Zebras look like horses, but they are not the same... Zebras, despite man's best efforts, cannot be tamed. The Wi-Fi we have engineered today remains fundamentally a collection of autonomous, uncoordinated things—zebras that simply cannot be harnessed.

Fi-Wi is architected from the ground up to be controllable, coordinated, and directed — the horse we need for in-building communications and sensing. As latency demands tighten and building densities increase, Fi-Wi isn't just a better future; it's the future we can build today.

1. Motivation and Problem Statement

"Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away." — Antoine de Saint-Exupéry

"Everything should be made as simple as possible, but not simpler." — Albert Einstein

With 23.3 billion Wi-Fi devices in use worldwide and 5.5 billion people depending on internet connectivity, and growing, Wi-Fi has become the primary way we access the internet. So much so many people think Wi-Fi is the internet. It's how a home healthcare worker video-calls to check on a patient, or a cancer patient connects to their support group. It’s how a parent works remotely while their child attends school online, and how lifelong learners access the information they need to grow. It’s how a grandmother monitors her heart condition through a telehealth app. It’s how a family member finds their next job, or how a neighbor orders a meal.

Running quietly in the background are autonomous systems we've come to depend on: security cameras that alert us to threats, medical monitors that track vital signs, smart home systems that manage climate and safety, IoT sensors that detect water leaks or carbon monoxide. These systems don't wait for us to notice problems—they operate continuously, silently, keeping people safe.

We've moved far beyond entertainment and convenience. Wi-Fi now carries the infrastructure of daily survival. When it breaks down under density or congestion, it's not just buffering that fails. It's jobs, healthcare access, human connection, and the life-safety systems we trust to work when we're not watching. The $4.9 trillion Wi-Fi contributes to the global economy isn't an abstract number. It's the cumulative value of billions of human activities and critical systems that simply stop working when the network fails.

Why Traditional Wi-Fi Cannot Support L4S

The infrastructure supporting all of this is failing at scale, and it must be addressed for all. The industry is moving toward L4S and ECN-based control to eliminate bufferbloat, but traditional Wi-Fi makes this impossible. Legacy congestion-control loops fail by design once a single flow saturates the bottleneck queue, and even modern ECN-based systems such as L4S cannot converge when Wi-Fi hides queue depth, induces collision storms, injects firmware-created delays that look like queues, and constantly shifts transmission (PHY) rates through its rate-control and aggregation machinery. Mesh networks and more APs catalyze intolerable user experiences by injecting more uncoordinated radios into an already chaotic RF environment. And because the AP industry understands these limits, it is no surprise that even major vendors publicly state that L4S cannot operate correctly over the products they sell.

Adding more Ethernet-attached APs makes it worse by creating more overlapping contention domains. Hidden queues in SoCs, rate-control firmware, and aggregation pipelines obscure the true bottleneck. In control-theory terms: the bottleneck queue cannot expose its state, the PHY rate is not stationary, and the closed loop cannot stabilize. This is why user experience fails in many apartments and homes, in hotels, MDUs, stadiums, and high-density buildings long before “capacity” is reached.

QoS cannot rescue this architecture. Because the bottleneck queue inside a Wi-Fi AP has no information about actual flow urgency or priority, no QoS mechanism can operate meaningfully. The only real solution is to avoid congestion altogether — which is exactly what L4S researchers have designed for and exactly what Fi-Wi supports.

Why Copper Infrastructure Has Reached Its Limits

While the protocol fails in the air, the physical infrastructure fails in the walls - the industry’s traditional answer of running copper Ethernet to APs — simply extends the lifetime of an architecture that has reached its limits. Copper requires periodic rip-and-replace cycles: Cat5 becomes Cat6, then Cat7, then Cat8. A home builder has no idea what communications wiring to install. The RJ45 connector and its plastic tab is fragile, outdated and end of life. And at 25G, 40G, or 100G, physics takes over: copper loses signal in dB per inch. Data centers have abandoned structured cabling (long-run copper) for core transport, restricting copper only to short-reach intra-rack DACs. Fi-Wi applies this same logic to the building: Fiber for the long haul (halls/walls), radio for the short hop.

How Fi-Wi Breaks Both Cycles

Fi-Wi breaks the cycle. Install fiber once — and never revisit behind walls or ceilings again. The glass is permanent; only the optics evolve. Fiber is already the universal medium for 100G/400G data centers, DWDM long-haul transport, and now PCIe throughout a building with Fi-Wi. Remote Radio Heads simply convert between fiber and 802.11, eliminating embedded routing, rate-control SoCs, switching silicon, and the security-patch treadmill they require. When Wi-Fi standards evolve, you replace the small radio module(s) — that's all.

Fi-Wi turns fiber combined with 802.11 into the permanent, predictable, control-theory-friendly transport that the L4S control loop requires, and treats 802.11 radio heads as the small, disposable, last-meters, connector-free interface where the in-building network behaves deterministically. And because fiber increases the long-term value of a building, the investment is not just technically durable — it is financially durable.

The Opportunity Is Here

There is no law of physics that says Wi-Fi cannot work at scale. The collapse we're seeing in apartments, hotels, and high-density buildings isn't inevitable. The researchers have shown engineers how to proceed. We know how to build stable control loops. We know how to coordinate radios. We know how to deploy permanent infrastructure.

The conditions for solving this are here, now. Engineering talent exists across our industry. The market has already validated the foundation: China's FTTR deployments have installed fiber to millions of rooms, proving that permanent infrastructure at this scale is not just feasible—it's already happening at volume. What's missing is capital directed at the right architecture. Investors are essential to this challenge. Their capital will enable the engineering to serve the market. And, once proven, market signals will sustain the development, directing human resources toward building what humanity needs for continued advancement.

Fi-Wi is Umber's answer, but the underlying challenge belongs to all of us. The 5.5 billion people depending on this infrastructure deserve better than a system designed for convenience that we've repurposed for survival. This is solvable engineering—the talent is ready, the manufacturing exists, and the market is waiting. It's time we came together and fixed this.

2. The Wi-Fi Crisis: Why Evolution Failed and Control Was Lost

The failure of modern Wi-Fi to support low-latency applications (L4S) is not a failure of bandwidth; it is a failure of control. With 23.5 billion Wi-Fi devices deployed globally, the protocol has hit an asymptotic limit where adding complexity yields diminishing returns.

As density rises, autonomous contention scales super-linearly—effectively operating as the inverse of Metcalfe's Law. The result is a rising noise floor and media access collisions that render unlicensed spectrum unusable for the deterministic performance required by next-generation applications.

2.1 The Evolutionary Trap: Why Incremental Improvements Failed

Evolutionary engineering is powerful; it gave us twenty-five years of Wi-Fi speed improvements. But every evolutionary curve eventually hits an asymptote—a point where adding more complexity yields diminishing returns. We have reached that point.

"The IEEE 802.11 working group behaves like a composer writing a symphony that effectively cannot be played. They continually add instruments—4096-QAM, Puncturing, MLO—without considering that the musician (the silicon) has only microseconds to react."

The decision matrix for a Wi-Fi chip has exploded combinatorially. We can trace this through the Modulation and Coding Scheme (MCS) Table:

• 1999 (802.11b): 4 choices (1, 2, 5.5, 11 Mbps). A simple if/else ladder.
• 2003 (802.11g): 12 choices with OFDM.
• 2009 (802.11n): 77 choices with MIMO and Channel Bonding.
• 2013 (802.11ac): 80+ choices with 256-QAM and wider channels.
• 2024 (802.11be / Wi-Fi 7): Thousands of permutations involving 4096-QAM, 320 MHz channels, 16 Spatial Streams, and Puncturing patterns.

The Physical Trap: When the firmware engineer fails to optimize the radio, can we simply redesign the chip? No, because of RTL (Register Transfer Level) Accretion. In software, engineers "refactor" unwieldy code. In hardware, refactoring is economically forbidden. A complex SoC takes 18–24 months to validate; removing "dead" logic risks breaking obscure corner cases. Consequently, vendors only add; they never subtract. 802.11be logic wraps around 802.11ax logic, which wraps around 802.11ac logic—twenty-five years of accumulated technical debt consuming area and leakage power.

The Market Signal: The ultimate proof that the standard has reached gridlock is the behavior of market leaders like Samsung and Apple. They no longer rush to support every new feature—they aggressively whitelist features and blacklist others because complexity drains battery and destabilizes connections. When the two largest consumers of wireless silicon effectively stop buying the complexity argument, the evolutionary roadmap is broken.

2.2 The Density Paradox: More Capacity, Less Performance

The fundamental instability of 802.11 stems from the Birthday Paradox applied to media access. In an autonomous system, as the number of contending stations (n) increases linearly, the probability of collision increases combinatorially:

Simulation data confirms that even with moderate client density, collision probability quickly exceeds 50%, forcing the network into a state of "Drift" where latency becomes unbounded. Under these conditions, the network is no longer constrained by PHY capacity, but by the probability of successful media access.

This is Metcalfe's Law in reverse: instead of each new node increasing the value of the network, each new node increases the chance of interference and reduces usable capacity.

2.3 The Three Technical Failure Modes

The collapse of the operator model is driven by three distinct architectural failures inherent to the 802.11 standard.

2.3.1 Protocol Tax: The Hidden Node Penalty

Standard Wi-Fi relies on Carrier Sense Multiple Access (CSMA), which assumes that all stations can hear each other. In real-world MDU (Multi-Dwelling Unit) environments, this assumption fails catastrophically.

Field measurements using ESP32-based sensors reveal that hidden node contention consumes 30-50% of available airtime in typical MDU deployments—airtime paid for in spectrum acquisition costs but lost to protocol overhead invisible to traditional monitoring. This represents a massive protocol tax where significant airtime is consumed by retries and backoff slots rather than payload delivery.

2.3.2 The MCS Matrix: Un-Engineerable Complexity

The most critical failure for a network operator is the loss of state control. Modern 802.11ax supports 12 MCS indices × 4 bandwidth options × 8 spatial stream configurations × 3 guard intervals = >1,000 valid PHY states. Autonomous rate selection must navigate this space at sub-millisecond timescales under non-stationary noise.

This creates a Non-Stationary System:

• Non-Stationary Chaos (Wi-Fi): The noise profile is time-variant and driven by unbounded interference (neighbors). This results in an unpredictable latency tail that cannot be engineered away.
• Stationary Stochasticity (Wired Networks): By contrast, cable and fiber networks operate on shielded media with known thermal noise profiles. The distribution is time-invariant, allowing for engineerable reliability (99.999%).

Because Wi-Fi is non-stationary, autonomous rate selection under contention has no bounded outcome. The IEEE 802.11 standard has allowed the MCS table to explode into hundreds of valid permutations—a chaotic state space that firmware must navigate in microseconds with incomplete information.

2.3.3 The Spatial Contention Cascade

As load increases, the spatial precision of the network degrades. Mathematical modeling shows that the condition number (κ)—a measure of how well-conditioned the MIMO channel matrix is—degrades from 6 dB (excellent spatial separation) to >12 dB (severe interference) under load. This collapse means that 4×4 MIMO effectively degrades to 2×2 or worse, turning additional spatial streams into self-interference rather than capacity.

This degradation collapses the theoretical gains of Mu-MIMO, transforming high-order spatial streams into interference rather than usable capacity. The "Efficiency Paradox" emerges: Wi-Fi evolution has focused on shrinking Payload Duration (faster PHY rates like 4096-QAM) while MAC Overhead (LBT, Backoff, Preamble) remains constant. To amortize the overhead, chips must build massive Aggregates (A-MPDUs). This destroys latency. We have engineered a Ferrari engine (the PHY) inside a garbage truck (the MAC).

2.3.4 Transport-Layer Assumption Violations

TCP congestion control and L4S both assume a service interval that is bounded and approximately stationary — a pipe with a knowable rate. CSMA/CA under saturation violates this assumption structurally. Service time per station is not a fixed function of MCS and payload size; it is a random variable whose tail is determined by contention depth, hidden-node geometry, and retry behavior. Under saturation, service intervals can spike from milliseconds to seconds. TCP interprets this as loss and backs off; L4S interprets it as a signal to pace back. Both responses reduce offered load, which partially relieves contention — but the relief is transient, the control loop is slow relative to the contention timescale, and the result is the sawtooth behavior described in §2.3.3.

No IP-layer congestion control can stabilize a MAC layer whose service-time distribution has an unbounded tail. The fix requires bounding the service-time distribution at the MAC layer itself, which requires a scheduler with enough state to assign slots rather than arbitrate contention. This is why IP-layer tools and architectural changes to the MAC are not substitutes for each other: they operate at different layers of the problem, on different timescales, with different information available to them.

2.3.5 The Hidden Controller Problem

Queuing theory describes the symptom. Control theory explains why the service center becomes unstable.

Every AP is running a hidden controller with a small local state vector. When loss occurs, that controller cannot reliably distinguish weak signal, hidden-node collision, OBSS interference, aggregation cost, or airtime overload. Those are different failure modes, but they collapse to the same local observation: retries, missed ACKs, and degraded service time. The controller must choose a response without knowing which failure mode it is in.

The wrong control action makes the system worse. Lowering MCS may be correct for weak signal, but it is harmful when the real problem is hidden-node collision or airtime saturation. Lower MCS increases airtime occupancy, longer airtime increases collision exposure, and more collisions drive more retries and more fallback. That is an unstable discrete-time control loop: a pole near or outside the unit circle in the z-plane.

There are also two conflicting control objectives. A capacity-optimized controller maximizes delivered bits per second over an averaging window. A latency-optimized controller minimizes service-time variance and tail delay under load. Those optima are not the same. Larger aggregates, aggressive MCS selection, and retry persistence can improve average throughput while worsening tail latency. The AP firmware may be doing exactly what it was designed to do, but it is optimizing the wrong cost function for user experience.

Fi-Wi changes the control problem by increasing observability and changing the objective function. The concentrator has a richer state vector: per-STA PHY state, per-RRH receive visibility, retry probability, SINR, service-time history, aggregation cost, and WAN/WLAN congestion state. That lets the controller distinguish the cause of service-time variance before choosing the control action. The objective function is explicitly latency and service-time variance, not peak throughput.

An autonomous AP cannot expand its state vector or change the building-wide cost function without a fronthaul. The observability constraint is architectural. This is true of Wi-Fi 5, 6, 7, 8, or any future generation that preserves the autonomous AP model: the controller is local, its state vector is small, and its objective function is capacity. Changing any of those three properties requires moving the MAC off the contested medium and behind a fronthaul.

2.4 The Operator's Dilemma

For network operators—whether cable MSOs, telcos, or fiber providers—this architectural chaos presents a fundamental business risk: You own the customer experience, but not the air interface.

• Wasted Infrastructure Investment: Investments in L4S congestion control, low-latency transport, and multi-gigabit access are negated if the final 10 meters are non-deterministic. The bottleneck shifts from the core to the customer premise, and operators lose visibility into the failure point.
• The Silicon Trap: Vendors market 4096-QAM (Wi-Fi 7), which requires ~40 dB SNR—physically impossible to sustain in dense environments. Operators pay for silicon logic that creates marketing value but zero operational value. The promised performance exists only in anechoic chambers, not in apartments with 30+ interfering networks.
• Operational Drag: The "un-engineerable gap" drives OPEX through support calls, truck rolls, and customer churn. When a subscriber's video call freezes, they call their ISP—even though the failure is in the Wi-Fi AP, not the WAN link.
• Competitive Vulnerability: Fiber-to-the-home providers market deterministic performance as a premium feature. Operators relying on legacy Wi-Fi CPE risk commoditization if they cannot match these latency and reliability guarantees.

2.5 Why Conventional Solutions Don't Scale

Traditional attempts to solve Wi-Fi density problems fail because they address symptoms rather than the underlying architectural failure:

• More APs: Adding APs increases the number of contending radios, exacerbating the Birthday Paradox. Collision probability grows combinatorially. More APs simply create more independent collision domains with the same per-domain collapse behavior. Field practitioners have long observed the corollary: in over-deployed buildings, removing APs raises throughput. Fi-Wi agrees with this observation for distributed MACs and inverts it inside a single scheduling domain — see Appendix R for the full treatment, including why centralized radio density is not a return to Meru/SCA.
• DFS (Dynamic Frequency Selection): Regulatory limitations and radar evacuation requirements make DFS unreliable for latency-sensitive applications. When a radar is detected, the AP must vacate the channel within seconds, causing multi-second outages.
• 6 GHz Spectrum: Additional spectrum doesn't solve density—it just shifts the collision problem to a new band. The fundamental CSMA/CA contention mechanism remains unchanged.
• Wi-Fi 7 / MLO (Multi-Link Operation): More MCS states = MORE chaos. Multi-Link Operation adds link diversity but doesn't eliminate autonomous contention. Each link still suffers from the same collision storms and hidden queue problems.
• Mesh Networks: Mesh adds relay hops, multiplying latency and cutting effective throughput by 50% per hop. Multi-hop mesh in dense environments becomes a latency disaster.
• Software-Only L4S + EDCA Redesign: The strongest version of the autonomous-AP-plus-software counter-argument is recent work from Nokia Bell Labs (Chang, Coomans, De Schepper, 2026), which combines a host-resident DualPI2 implementation with a new low-latency EDCA parameter set called AC_N15. On commodity Wi-Fi 6 hardware with software-only changes, the result reduces P99.99 latency under 4-STA contention from roughly 105 ms to roughly 35 ms. This is a real and useful result within its scope. It does not coordinate across APs, does not address spatial-channel conditioning under density, does not pre-position resources based on building-level temporal patterns, and does not mark L4S based on coordinated airtime allocation. It improves each autonomous AP locally. It does not produce the centralized BSS manager that 802.11 EDCA assumed would exist. See Appendix P for the full treatment of how Fi-Wi relates to this work.
• IP-Layer Queue Management (CAKE, fq_codel, LibreQoS): Tools that operate in the IP forwarding path improve queue dynamics above the 802.11 layer. They do not alter TXOP allocation, EDCA contention behavior, retransmission airtime fraction, AMPDU behavior, hidden-node behavior, or PHY error recovery inside the 802.11 system. Every shaped packet still enters a contention-governed MAC/PHY domain once it reaches the air interface. Application-visible improvement from IP-layer shaping does not establish that the underlying 802.11 contention problem has been solved — it establishes that reducing the offered load into a contended system helps, which is true and useful but is not the same thing. See Appendix M for the full treatment of where fq_codel and CAKE remain appropriate and where they do not reach.
• Inter-AP RF Coordination (BSS Coloring, Multi-AP): Coordination between neighboring networks addresses OBSS interference and is a genuine improvement. It is not a substitute for intra-network coordination. The 134 million occupied US housing units span a wide range: roughly 86 million single-family homes where inter-network coordination addresses most of the RF environment problem, and 48 million multifamily units where intra-building contention among many STAs sharing the same AP infrastructure dominates. No single coordination fraction applies across the install base. More importantly, coordination across administrative boundaries requires cooperation between independent operators and devices with no shared control plane — a deployment and policy problem that 802.11bn's Multi-AP coordination work has not yet solved at scale. Fi-Wi's coordination is intra-building and operates over a shared control plane the operator owns. It does not require neighboring networks to participate.

The Trillion-Dollar Context: The mobile industry spent $600 billion building 5G to get scheduled, deterministic performance outdoors. They understand that unlicensed spectrum + autonomous contention = chaos. The genius of 5G is its architecture; its Achilles heel is its cost. In recent auctions, 20 MHz of licensed mid-band spectrum sold for over $17 billion for U.S. rights alone.

Fi-Wi applies the cellular C-RAN architecture indoors—but on unlicensed spectrum that costs nothing. This is the arbitrage opportunity.

2.6 The Client Side: Autonomous STAs and the Limits of EDCA

The crisis is not confined to the infrastructure side. In legacy Wi-Fi, the Station (STA) is itself an autonomous agent that fights for upstream airtime using EDCA. It maintains its own local WMM queues and transmits whenever it wins a contention window, often oblivious to the fact that the AP’s receive buffer is already full or that other STAs in the same BSS are about to collide on the same backoff slot. The STA cannot coordinate with peers because there is no shared view of medium state to coordinate against.

The consequence is a push model: the STA pushes packets into the air as fast as the PHY allows. If the AP is congested, the STA does not know until packets drop or layer-2 retries spike, and latency accumulates in the STA’s own driver queue because the “pipe” is smaller than the “faucet.” This is the client-side dual of the infrastructure problems described in §2.1–§2.5. Solving it requires either (a) per-STA modifications that 25 billion already-deployed devices will never accept, or (b) an infrastructure-side architecture that influences uplink behavior without requiring universal STA cooperation. §3 introduces the architectural response; §3.3 and §3.10 detail the specific mechanisms by which the Wireless Fabric reaches even legacy STAs.

3. The Wireless Fabric: A Third Data-Plane Primitive

Networking has long had two dominant data-plane primitives: IP routing, which maps destination prefixes to next hops, and Ethernet bridging, which maps learned MAC addresses to ports and broadcast domains. Each defined an industry. Umber is building a third primitive for dense wireless systems: a Wireless Fabric that maps wireless aggregates to scheduled radio resources (RRH, time, frequency, and spatial eligibility) under centralized control. The Wireless Fabric is a logical primitive: a software data plane running on commodity host hardware, not switching silicon. Fi-Wi is Umber’s reference implementation of the Wireless Fabric.

The Wireless Fabric is a single data-plane primitive with three tightly-coupled aspects: forwarding (which scheduled radio resource each aggregate is bound to), scheduling/pacing (when and how it transmits), and marking (how the bottleneck signals back to transport). The aspects share state and execute in the same memory and timing domain. Marking is the cross-layer aspect: its inputs span 802.11 airtime and queue state, IP-layer flow context, and observed transport behavior, and its output writes ECN codepoints into the IP header so that L4S-aware senders react. The rest of this section unpacks the architecture: where the Wireless Fabric sits in the networking stack, how its aspects couple, what the execution environment looks like (DPDK and lcore mapping), and where it is expected to fail. Subsequent sections elaborate on the implementation: the system diagram (§4), key Fi-Wi mechanisms (§5), control architecture (§6), and onward.

Coverage is defined bottom-up by the room, not top-down by the building. The deployment minimum is one cell per room served by one RRH, with the in-room serving RSSI no weaker than −65 dBm. Multiple RRHs covering the same area are encouraged: they provide active redundancy (§9) and improve the spatial-channel eigenspaces that MU-MIMO depends on, keeping the condition number low under load (§3.14 demonstrates the collapse that autonomous APs cannot avoid). A whole-building Wireless Fabric is therefore not a handful of large APs but tens to hundreds of small, deliberately per-room RRHs coordinated by one concentrator.

The fabric’s outer boundary is set by whichever of these binds first: the scheduler’s 250 µs TXOP target (loop period for grant emission to each RRH), the PCIe-over-fiber link reach, the TSF coherence window, the RF coupling envelope within which coordination is useful, and the administrative boundary of a single operator’s policy domain. The intended near-term deployment target is one MDU, office building, or small campus, where those bounds are expected to be manageable. An autonomous AP defines coverage as its own ~30 m RF cell, with neighboring APs creating overlap that becomes interference. The Wireless Fabric inverts this: many small per-room cells with cross-coupling treated as a scheduling input rather than interference. §3.11 describes how some aspects can be distributed across head-end and building over DWDM, but the coordinated-scheduling domain itself stays building-bounded.

/* Linear regression over sliding window of TSF sample pairs */
static void map_refit(struct tsf_affine_map *m)
{
    double var_dm = 0.0, cov_dm_dr = 0.0;
    double mean_dm = sum_dm / n, mean_dr = sum_dr / n;

    for (i = 0; i < n; i++) {
        double x = dm[i] - mean_dm;
        double y = dr[i] - mean_dr;
        var_dm    += x * x;
        cov_dm_dr += x * y;
    }
    m->alpha = cov_dm_dr / var_dm;   /* drift rate */
    m->beta  = mean_dr - alpha * mean_dm;  /* offset */
}

/* Project master TSF to predicted radio TSF */
static uint64_t predict_radio(const struct tsf_affine_map *m, uint64_t master_tsf)
{
    int64_t dm   = (int64_t)(master_tsf - m->anchor_master);
    double dr_est = m->alpha * (double)dm + m->beta;
    return tsf_add_s64(m->anchor_radio, (int64_t)llround(dr_est));
}

/* Project master to all RRHs in one call */
void tsf_affine_pool_master_to_all(const struct tsf_affine_pool *p,
                                   uint64_t master_tsf,
                                   uint64_t out_radio[TSF_AFFINE_MAX_RADIOS]);

3.13 Splitting the Graph: Concentrator-vs-RRH Decision Decomposition

The preceding sub-sections describe what the Wireless Fabric is: a data-plane primitive with three coupled aspects, a single concentrator memory domain, a DPDK execution environment, and a defined set of failure modes. This sub-section offers the corresponding mental model: an intuition for why the decomposition into concentrator-side decisions and RRH-side execution is the right one. The lens is the MCS State Graph introduced in §2.3.2. The Wireless Fabric’s architectural breakthrough is decoupling that graph into its constituent parts:

• The Concentrator Owns the Edges (Navigation): The computational complexity of Wi-Fi lies in the transitions: deciding when to move from one state to another based on interference, queue depth, and CSI. By moving this logic to a workstation-class processor, the Concentrator can traverse the graph in software, evaluating the cost of every edge in real-time. These edge costs are computed from CSI vectors (channel quality), queue sojourn time (congestion), and spatial correlation matrices (interference), allowing the Concentrator to choose the lowest-cost path through the MCS state space for each packet.
• The Radio Head Owns the Nodes (Execution): The Radio Head is reduced architecturally toward a simple instrument that owns the physical states (the nodes). Current commodity Wi-Fi firmware retains parts of rate control, aggregation, and retry policy that the Wireless Fabric target moves to the Concentrator; the architecture reduces the RRH as far as commodity firmware permits and adds host-side override paths (§3.9) where firmware ownership blocks the move. In the target state the RRH does not decide which node to occupy; it acts as a stateless executor, converting the Concentrator's command (MCS_INDEX, N_SS, TXOP_DURATION) into the precise RF waveform without autonomous deliberation.
• The Graph Is Not Static: Unlike traditional rate adaptation (e.g., Minstrel), the Concentrator's graph is continuously recomputed based on global state. The edge costs reflect not just per-client SNR, but cross-client interference, queue depth across all flows, and spatial correlation between RRHs. This global visibility is impossible in autonomous APs, where each device sees only its own local channel state.

This architectural shift, from distributed chaos toward centralized control, mirrors the evolution from analog transmission systems (noise-prone, operator-invisible) to digital QAM (predictable, monitorable). Fi-Wi completes this transformation for the last 10 meters, moving the network from a model of probabilistic negotiation toward one of scheduled, observable execution.

Section 17 describes the Concentrator's scheduling algorithm that implements this graph traversal, while Appendix C details the RRH's scatter-gather DMA mechanism that executes the chosen state transitions at microsecond timescales.

3.14 Interactive Visualization: The MCS Collapse Under Load

The following interactive simulation demonstrates the architectural differences between Fi-Wi, autonomous APs, and mesh networks under varying load conditions. It visualizes the MCS State Graph discussed in §3.13, showing how autonomous systems fail to navigate this state space under density.

Each "room" represents a device with a 4 × 12 grid of MCS states (4 spatial streams × 12 MCS indices). The ghost node (dashed) shows the ideal state based on channel quality, while the active node shows the actual state selected by the rate control algorithm.

Technical Details: Understanding the Visualization

MCS Grid: Each 4×12 grid shows all possible MCS states. Top rows = Mu-MIMO (multi-user), bottom rows = standard 2×2 MIMO. Columns = MCS index (0-11, higher = faster but needs better SNR).

Eigenvalues (λ₁, λ₂): Strength of spatial modes in the MIMO channel. As density increases in autonomous mode, λ₂ collapses → spatial interference.

Condition Number (κ): Ratio λ₁/λ₂ in dB. Low (~6 dB) = good. High (>12 dB) = Mu-MIMO degraded to single-stream. This directly demonstrates the "Spatial Contention Cascade" from Section 2.3.3.

Collision Probability: Computed using Birthday Paradox formula: n(n-1)/2 collision pairs. When this exceeds 50%, the network enters "Drift" state with unbounded latency.

Why This Matters for Network Operators

This visualization proves the loss of control described in Section 2.4. In autonomous mode, operators cannot engineer performance because the system navigates a 1,000+ state MCS graph with no global coordination.

In Fi-Wi mode, the Concentrator's global state visibility allows it to:

• Choose optimal MCS/SS combinations based on current cross-room interference
• Coordinate spatial reuse to prevent Mu-MIMO collapse (maintain κ < 12 dB)
• Eliminate collisions through centralized scheduling (Birthday Paradox solved)
• Apply accurate L4S ECN marking with full queue visibility

The result: predictable, engineerable performance that scales with density instead of collapsing. The difference becomes visceral when you watch autonomous mode turn red under the same load that Fi-Wi handles in green.

4. System Picture

4.1 Classical Stack vs. Fi-Wi (The C-RAN Shift)

To understand Fi-Wi, we must first unlearn the definition of an "Access Point."

In a typical controller-managed enterprise Wi-Fi deployment, a centralized controller (e.g., Cisco WLC, Aruba Mobility Controller, Ubiquiti UniFi Controller) coordinates AP configuration: channel assignment, transmit power, client steering recommendations, and SSID management. However, each AP remains autonomous at the data plane:

• Each AP maintains its own queues and packet buffers
• Each AP makes its own MAC scheduling, aggregation, and TXOP decisions
• APs have no visibility into other APs' per-packet state, queue depths, or precise timing
• CSI (Channel State Information) remains local to each AP
• There is no shared timestamp reference or coordinated DMA across APs

These systems are loosely-coupled: the controller manages the control plane (configuration, policy) but the data plane — queuing, MAC scheduling, aggregation, and packet forwarding — remains distributed and autonomous across individual APs.

In Umber Fi-Wi (C-RAN for Wi-Fi), we split the AP and cellularize the RF domain, down to room-level. The concentrator sees all flows, all queues, and all RRHs. The RRHs handle 802.11 MAC/PHY but are tightly time-synchronized and behave as DMA-driven PHY/MAC endpoints rather than autonomous APs. A set of RRHs and their shared queues form a cellularized Wi-Fi domain within the building, often at “cell per room” granularity.

Fi-Wi centralizes both control plane AND data plane with shared state across all RRHs. The concentrator doesn't just configure RRHs; it directly manages their queues, schedules their TXOPs, and maintains unified timestamp-synchronized state across the entire cellularized RF domain.

4.2 Dual-Loop Control Model

Conceptually, Fi-Wi decouples the system into two nested feedback loops, separated by timescale:

The Outer Loop manages congestion and end-to-end latency (Internet speed). The Inner Loop manages MAC efficiency and radio timing (Airtime).

The Problem with Legacy Wi-Fi: Traditional APs couple these loops unpredictably, creating "sawtooth" latency patterns that confuse TCP.

The Fi-Wi Solution: By centralizing both loops in the Concentrator, Fi-Wi enforces a strict Time-Scale Separation. The Inner Loop runs so fast (3–5 kHz) that it appears as "constant service" to the slower Outer Loop (10–20 Hz), allowing L4S to stabilize perfectly.

(See Section 6: Control Architecture for the rigorous control-theoretic analysis and stability criteria.)

5. Key Fi-Wi Mechanisms

5.1 Time Synchronization

Fi-Wi operates across two distinct MAC-layer time domains simultaneously; a third clock domain — the RF sampling clock governing coherence across RRH front ends — is an RF signal-processing concern treated separately in Section 10. The first is the concentrator's internal master clock, maintained via PCIe-native timing (PCIe PTM) over the fronthaul — with PTP/802.1AS available as an optional external reference where the deployment requires absolute time (detailed in Section 10). The second is the 802.11 TSF (Timing Synchronization Function) domain that 802.11 clients use to coordinate with the MAC layer. In a traditional AP these two clocks are decoupled — the AP runs one TSF and one clock. In Fi-Wi, with 24 RRHs each presenting a TSF-aware BSS, managing the relationship between them is a foundational architectural responsibility of the concentrator.

5.1.1 The Fronthaul Clock: PCIe PTM

The concentrator synchronizes its master clock to all attached RRHs to within tens of nanoseconds (20–50 ns via PCIe PTM — see Section 10 for the full hardware chain and the comparison against Ethernet PTP). This master clock gives every packet:

• A meaningful ingress timestamp when it is enqueued in central memory
• A meaningful service timestamp when the RRH actually transmits it over the air
• A consistent time base for measuring queue sojourn time and service behavior per RRH or per RF group

This clock lives entirely inside the Fi-Wi domain. Clients never see it directly. It is the coordinate system in which shim header timestamps (Section 5.2), AQM marking decisions (Section 5.3), and the ML training corpus (Section 16) are all expressed. Because all packet timestamps, service events, and queue measurements are expressed in this single master time domain, Fi-Wi can compute precise per-packet sojourn times independent of the TSF domain, enabling stable ECN marking and L4S control across the system.

This synchronization rides the same PCIe-over-fiber links that carry data — no separate timing network is required. The full hardware chain (concentrator grandmaster oscillator, jitter attenuator, chipset Ext_Ref pin, GPS-disciplined vs. free-wheeling modes, and the PTM vs. Ethernet PTP comparison) is covered in Section 10: The Three Clock Domains.

5.1.2 The 802.11 TSF Domain

The 802.11 TSF is a 64-bit microsecond counter that every client associates with a BSS. Clients set their local TSF from beacons. They use it to wake from power save at the right moment, to interpret TBTT (Target Beacon Transmission Time), and to coordinate TXOP timing. The TSF is the only MAC-visible clock the 802.11 standard exposes at the MAC layer.

In a traditional single-AP deployment this is trivial: one AP, one TSF, one beacon stream. In Fi-Wi it is not. Consider a client in a room served by two RRHs in the same airtime domain. That client will receive beacons from both RRHs. If those beacons carry inconsistent TSF values, even small inconsistencies can lead to misaligned power-save wakeups, ambiguous TBTT interpretation, and in some implementations degraded performance or reassociation. The coherence of the TSF domain across all RRHs in a BSS is not optional; it is a hard correctness requirement.

Fi-Wi satisfies this requirement by construction: the concentrator generates all beacon frames. No RRH constructs its own beacon. The concentrator writes the TSF value into every beacon before dispatching it to the appropriate RRH for transmission. Because all TSF values originate from the same source and are derived from the same master clock, they are consistent by design rather than by coordination protocol. Within a given BSS, TSF values are identical across all participating RRHs; multiple TSF domains arise only when multiple BSS instances are present.

5.1.3 The Concentrator as Time Origin

The concentrator maintains 25 simultaneous time references: its own concentrator master clock (the concentrator is the grandmaster for the Fi-Wi domain; the UAX-8 prototype uses the MS-01 motherboard oscillator, sufficient for TSF affine correction; a production concentrator may use a higher-stability oscillator) and one 802.11 TSF per RRH. Each TSF has its own epoch (established at BSS creation) and its own drift correction term, derived from periodic synchronization updates over the fronthaul (PCIe PTM), which bound long-term drift. The per-RRH terms exist to compensate for each radio's local oscillator: they are a correction mechanism, not a divergence, ensuring that the TSF value transmitted over the air is identical across all RRHs participating in a BSS (Section 5.1.2). The concentrator knows the exact affine mapping between the master clock and every client-visible TSF domain at all times:

TSF_i(t) = (t_master - epoch_i) + drift_correction_i(t)

Any event — a packet enqueue, an ECN mark, a TXOP start, a beacon transmission — can be expressed in any of the 25 frames without loss of precision. This is the time-domain analog of a coordinate transformation: the concentrator is the origin from which all other reference frames are derived, and any event timestamp can be mapped between frames via a known, invertible affine transform, updated continuously via the fronthaul synchronization loop.

This unified time model also enables the concentrator to schedule transmissions across RRHs against a single global timeline, rather than relying on independent per-RRH contention processes. TSF continuity across RRH handoffs is a direct consequence of centralized beacon generation, and it is what makes Fi-Wi's active redundancy claims in Section 9 operationally credible: per-packet steering between RRHs is transparent to clients because the client's MAC-layer time reference never changes. This unified time model enables not only precise measurement, but coordinated control of transmission behavior across RRHs, as described in Section 5.1.4.

The implementation of the affine clock model — including the least-squares regression over a sliding sample window, the alpha/beta parameters, and the tsf_affine_pool_master_to_all() projection call — is described in detail in §3.4.5 (TSF Synchronization: Affine Clock Correction).

5.1.4 Time-Driven EDCA Orchestration

The unified time model described above is not only a measurement framework; it is the foundation for Fi-Wi's centralized MAC scheduling. In conventional 802.11 deployments, EDCA (Enhanced Distributed Channel Access) operates as a stochastic contention mechanism: each AP independently selects random backoff values within its CWmin/CWmax range, and medium access emerges probabilistically.

In Fi-Wi, EDCA is not treated as a distributed random process. It is treated as a centrally orchestrated actuation layer, driven by the concentrator's master time reference.

Because the concentrator maintains:

• A global view of all per-RRH transmit queues
• Precise enqueue and service timestamps in the master clock domain
• A coherent TSF domain across all RRHs participating in a BSS

it can shape medium access behavior across RRHs by dynamically controlling EDCA parameters on a per-radio basis. The key parameters are:

• CWmin / CWmax: control the distribution of backoff intervals
• AIFS: offsets contention start times between RRHs
• TXOP: defines the service quantum once access is obtained

By assigning narrowly bounded contention windows and staggered AIFS values across RRHs, the concentrator can bias contention outcomes such that one RRH is overwhelmingly likely to win access at a given moment. Rotating these parameters over time creates a soft time-division multiplexing (TDM) effect using standard EDCA semantics.

This transformation is only possible because all RRHs share a common time reference. The concentrator can schedule EDCA parameter updates relative to the master clock and ensure that all RRHs apply them in a coordinated manner. Without this shared time base, independent EDCA processes would quickly decorrelate and revert to stochastic contention.

Conceptually, the concentrator executes a scheduling loop:

for each scheduling interval:
  observe queue state across RRHs        // centralized visibility
  select next RRH (or RF group) to serve // queue-aware decision
  assign EDCA parameters (CWmin, CWmax, AIFS, TXOP)
  enforce timing relative to master clock // coordinated application

The result is not strict TDMA — 802.11 contention semantics are preserved and the system remains compliant with standard client behavior — but the distribution of outcomes is shaped by the concentrator. Over short time horizons, access becomes highly predictable and service intervals can be bounded. This has two critical consequences:

• Deterministic service timing: Packets experience bounded and observable delay between enqueue and transmission, enabling stable AQM and ECN marking.
• Queue-aware scheduling: RRHs with pending traffic can be prioritized immediately, while idle RRHs do not contend unnecessarily for airtime.

Because TSF values are consistent across RRHs, these scheduling decisions are MAC-transparent to clients. From the client's perspective, the network behaves as a single, coherent AP with stable timing characteristics, even as transmissions are steered across multiple physical radios.

This time-driven EDCA orchestration is the mechanism by which Fi-Wi converts the inherently stochastic 802.11 MAC into a predictable, centrally scheduled system — completing the chain from time synchronization through queue observability to stable L4S control.

5.2 Fi-Wi Shim Header

Between 802.3/IP and the fronthaul link we add a small internal metadata header. Conceptual form:

struct FiWiMeta {
  uint64_t seq;          // fronthaul sequence number
  uint64_t t_ingress_us; // time packet enqueued into group queue (central DRAM)
  uint32_t txop_id;      // TXOP this MSDU is in
  uint8_t  mpdu_idx;     // index within aggregate
  uint8_t  mpdu_cnt;     // total MSDUs in this TXOP
  uint8_t  ecn_flags;    // CE applied? which queue? reason bits
  uint32_t qlen_pkts;    // queue depth snapshot at TXOP start
};

This header is visible only inside the Fi-Wi domain. It lets us:

• Record timestamps and queue snapshots at enqueue and dequeue
• Mark ECN based on central queue state at the moment of service
• Reconstruct at the concentrator exactly when a packet was queued and actually transmitted
• Compute a real feedback delay Td = now - t_mark per flow or per RF group

5.3 AQM / L4S Marking Placement

We choose the group queues in the concentrator—each corresponding to a cellularized airtime domain shared by one RRH or by multiple interfering RRHs—as the only places where deep queues are allowed and where we apply ECN:

• Downlink: group queues feeding downlink RRHs
• Uplink: group queues collecting uplink traffic from RRHs

Other queues (within RRH hardware, on the fiber/fronthaul link) are kept shallow via pacing and controlled descriptor posting. The group queues become the single bottlenecks in each cellularized airtime domain, which is exactly what L4S wants: a small number of stable, well-behaved bottlenecks with known behavior. The control policy is explicitly tuned to keep both average and tail queueing delay low.

In a variable-rate wireless medium, queue sojourn time alone is an unreliable congestion signal — service time varies per client based on PHY rates, contention, and retries. Fi-Wi replaces backward-looking buffer metrics with a forward model of wireless service time based on airtime debt. The full airtime debt model is covered in §18 (Airtime-Assisted ECN).

5.4 Centralized Packet Memory and DMA

In Fi-Wi, packet memory is centralized in the concentrator:

• Each RF airtime domain has one or more group queues implemented in concentrator DRAM.
• RRHs act as DMA engines that push/pull frames to/from those central queues.
• On-chip RRH buffers are deliberately small and just-in-time filled, to avoid hidden queues.

This design:

• Makes all meaningful queues fully visible to the control plane
• Lets AQM / ECN marking run exactly where the integrator (queue) lives
• Ensures there are no uncontrolled “mystery buffers” in the RF domain

5.5 RRH Edge Control via Beacon Power Shaping

Because the Fi-Wi concentrator maintains shared state for the entire RF domain, it can directly control the RF footprint of each RRH by adjusting per-RRH beacon transmit power. This alters:

• Coverage footprint – how far each RRH is visible as a candidate AP
• Cell edge location – which STA sees which RRH as the dominant beacon source
• Interference domains – how many RRHs mutually interfere for a given channel
• Candidate RRHs for grouping – beacon shaping influences which RRHs should share a queue

Beacon power is one of the most effective tools for dynamic RF cell shaping because it affects STA association and roaming decisions without modifying data-plane PHY rates. By lowering beacon power at certain RRHs and raising it at others, the concentrator can:

• Steer a STA toward a preferred RRH or RF group
• Increase isolation between RRHs to allow more independent airtime domains
• Reduce unnecessary overlap that causes hidden-node problems or collapses spatial reuse
• Stabilize spatial-stream structure by improving per-RRH dominance or diversity in the CSI

Traditional controller+AP systems attempt similar behavior but lack true shared state because each AP maintains its own queueing and PHY decisions. In Fi-Wi, beacon shaping is coordinated with:

• Current queue group assignments
• CSI-driven spatial-stream and eigenvector analysis
• Predicted load and collapse risk

This makes beacon power a first-class control variable in defining and stabilizing the boundaries of each cellularized RF domain.

5.6 Fronthaul Requirements and Feasibility

The Fi-Wi architecture requires deterministic, low-latency fronthaul links between the concentrator and RRHs. Because RRHs function as DMA engines accessing centralized packet memory, Umber's implementation uses PCIe (PCI Express) over fiber rather than Ethernet. PCIe avoids the network stack entirely: no TCP/IP processing, no Ethernet framing, no switch store-and-forward, and direct memory-to-memory transfers with credit-based flow control and hardware-enforced QoS. For the full bandwidth analysis, latency budget, queue and metadata mapping, and hot-swap behavior, see Section 13: PCIe Fronthaul.

6. Control Architecture: The Dual-Integrator System

A rigorous control-theoretic analysis of Wi-Fi reveals a fundamental challenge: there are not one, but two distinct integrators in the transmit path. In traditional autonomous APs, these integrators are coupled in undefined ways, leading to instability (bufferbloat) and poor interaction with TCP congestion control. Fi-Wi explicitly separates these integrators, applies distinct control laws to each, and enforces a strict Time-Scale Separation to guarantee system stability.

6.1 The Two Integrators

To achieve stability, we must model and control two distinct accumulation processes:

1. The Outer Integrator (Group Queue): Located in the Concentrator. This accumulates packets based on the mismatch between arriving traffic (internet speed) and the wireless link capacity. It operates on the RTT timescale (milliseconds).
2. The Inner Integrator (Aggregation Buffer): Located logically between the Concentrator and RRH. This accumulates packets to build 802.11 A-MPDU aggregates for PHY efficiency. It operates on the TXOP timescale (hundreds of microseconds).

6.2 The Outer Loop: L4S and Group Queue Dynamics

The primary bottleneck managed by the AQM (Active Queue Management) is the Group Queue. This loop drives the end-to-end congestion control (L4S/TCP).

6.2.1 Queue Dynamics

The queue depth Q(t) evolves based on the mismatch between the arrival rate λ(t) and the effective service rate μ(t):

dQ/dt = λ(t - τ_fwd) - μ(t)

6.2.2 The PI² Control Law

Fi-Wi uses a PI² controller to calculate a marking probability \( p(t) \), targeting a shallow queue reference \( Q_{ref} \) (typically 200 µs). This provides a coherent signal to L4S senders:

p(t) = K_alpha * (Q(t) - Q_ref) + K_beta * ∫ (Q(t) - Q_ref) dt

6.3 The Inner Loop: MAC Aggregation and TXOPs

The Inner Loop manages the trade-off between PHY efficiency (large aggregates) and latency (small aggregates). In traditional APs, this integrator is effectively unbounded to maximize benchmark scores, creating a "sawtooth" latency pattern that confuses TCP.

Fi-Wi bounds this integrator via two mechanisms:

• Bounded TXOP Duration: Limiting TXOPs to ≈250 µs forces the aggregation integrator to discharge frequently (3–5 kHz).
• Bounded Aggregate Size: Limiting \( A_{max} \) ensures the "sawtooth" amplitude never exceeds the L4S latency budget.

6.4 System Integration: Time-Scale Separation

For the nested loops to remain stable, the Inner Loop must look like "constant service" to the Outer Loop. This requires the Inner Loop bandwidth (ω_mac) to be significantly higher than the Outer Loop bandwidth (ω_tcp):

ω_mac >> ω_tcp   (typically > 20:1 ratio)

6.4.1 Frequency Domain Constraint

By forcing the MAC to operate at a frequency of 3–5 kHz (via 250 µs TXOPs), the aggregation noise is pushed high enough that it is naturally filtered out by the TCP loop (which operates at 10–20 Hz).

6.4.2 A-MPDU Aggregation Coherence and ECN Marking Precision

The 250 µs TXOP constraint serves a dual purpose: it maintains time-scale separation and ensures L4S receives coherent ECN feedback. Traditional Wi-Fi's massive A-MPDU aggregation creates a fundamental mismatch between Layer 2 efficiency and Layer 3 control precision.

The Aggregation-Feedback Mismatch

In wide-channel deployments (160 MHz), APs build large A-MPDU aggregates containing dozens of IP packets to amortize MAC overhead. This creates three control-loop pathologies:

• Binary ECN Feedback: Instead of smooth, per-packet congestion signals, the sender receives bursts of marks (when the aggregate transmits) or extended silence (during assembly), violating the PI² controller's assumption of sub-RTT feedback granularity.
• Buffering-Induced Latency: Building efficient aggregates requires multi-millisecond buffering. This artificial delay appears as congestion to L4S senders, causing premature rate reduction even when the medium is idle.
• Sojourn Time Measurement Loss: Once packets enter hardware aggregation pipelines, per-packet queueing delay becomes invisible, preventing accurate ECN marking per RFC 9332.

Fi-Wi's Coherence Strategy

Fi-Wi resolves this through coordinated design:

1. 40 MHz Channel Width: Narrower channels require smaller aggregates, naturally increasing MAC service frequency. More frequent transmissions with smaller payloads ensure sojourn time measurement occurs at packet granularity.
2. Concentrator-Level ECN Marking: The Concentrator performs sojourn time measurement and ECN marking before handing packets to RRHs for PHY transmission, preserving microsecond-level queueing visibility.
3. Bounded TXOP Duration: The 250 µs maximum ensures MAC service frequency remains >10× higher than L4S control frequency (~1 RTT), enabling senders to interpret ECN marks as smooth probability signals rather than discrete bursts.

This approach maintains the benefits of A-MPDU efficiency while preserving the feedback coherence L4S requires. The result: DualQ can sustain its ~1ms target drain time without artificial inflation from aggregate assembly delays. For detailed analysis, see Appendix I.7.

6.4.3 Design Parameters for Stability

Fi-Wi uses these parameters to ensure the system remains critically damped:

7. Airtime Domains and Dynamic Queue Grouping

In Fi-Wi, the core rule is: there is one deep queue per independent airtime resource. The physical queue lives in concentrator memory, but it represents the airtime of one RRH or a dynamic group of RRHs whose RF signals are coupled strongly enough to behave like a single cell.

If two RRHs can interfere, they cannot transmit simultaneously and therefore must share a single logical queue. If RRHs are RF-isolated, each receives its own queue. This preserves the “one bottleneck per control loop” structure required by L4S.

7.1 Why airtime determines queue structure

Service at each queue corresponds to over-the-air transmission. Any RRHs that share RF space must share a service process and therefore share a queue. RRHs that do not interfere have independent airtime and get independent queues.

7.2 Forming airtime groups dynamically

Crucially, these RF groups and their queues are not static. The concentrator forms and maintains airtime domains dynamically using:

• Static provisioning from building layout / installer input as an initial guess
• Passive RF discovery (RSSI correlation, channel utilization, beacon strength)
• Runtime adaptation (retry bursts, rate shifts, noise-floor correlation, airtime logs, CSI)

Beyond simple interference, Fi-Wi’s groupings also consider the spatial structure of the channels:

• CSI reveals the dominant MIMO eigenvectors and eigenvalues between a set of RRHs and a set of STAs.
• Groups can be chosen so that RRHs in the same domain not only share airtime, but can jointly form useful spatial streams and nulls when coordinated.
• RRHs whose channels are nearly parallel for a given STA might be grouped differently from RRHs whose channels provide orthogonal eigenmodes.

Over time, the Fi-Wi system continuously adjusts:

• Which RRHs belong to the same airtime domain
• Which RRHs can be treated as isolated cells
• How much capacity is budgeted per domain, including how many effective spatial streams it can carry

Groups may merge if interference appears or split if RRHs become effectively isolated (e.g., after a channel change or power adjustment, including beacon power shaping). The AQM and ECN marking logic always runs at the current group queue, so L4S always sees a single, well-defined bottleneck per cellularized domain.

Because all RRHs expose real-time CSI, queue metrics, retry statistics, airtime usage, and beacon reports into the concentrator’s shared state, Fi-Wi can form RF groups that are tuned not just for coverage but for:

• Interference minimization
• Maximum usable spatial streams (eigenvector structure)
• Tail-latency behavior under L4S marking

7.3 Room-Level RRH Density (FTTR-Class Deployment)

Fi-Wi is not designed around a small number of big AP cells per floor. The architecture assumes something much closer to Fiber-to-the-Room (FTTR): one cell per room, with fiber or equivalent deterministic fronthaul feeding small RRHs in each room.

In higher-end deployments, each room can contain multiple RRHs (e.g., 2–4 per room) to support:

• Location services, using multi-angle phase and RSSI measurements across RRHs
• Multi-static sensing for occupancy, motion, and other sensing applications
• Increased spatial-stream or eigenmode diversity, enabling higher-rank MIMO than a single AP could offer
• Robust redundancy — automatic handovers or multipath transmission without STA involvement

This density dramatically improves RF control. With RRHs separated by just a few meters, the concentrator sees:

• Sharper CSI contrasts between RRHs
• More stable eigenvectors for multi-RRH coordination
• Finer-grained airtime-group detection
• Better prediction of collapse risk or interference coupling at the room level
• Structural reduction in hidden node geometry: with 3–7 STAs per RRH in a confined space, STAs associated to the same RRH are geographically close enough to hear each other. Hidden node requires two STAs that cannot hear each other but share the same AP; high RRH density makes that condition rare by construction, without requiring any STA cooperation or RTS/CTS overhead.
• Per-RRH EDCA tuning: because each RRH is its own BSS with its own beacon, the concentrator can set EDCA parameters independently per RRH based on the specific STA population it serves. This is not per-STA EDCA, but with small per-RRH populations it is close enough to matter.

Traditional AP-based architectures cannot achieve this cleanly because they lack shared state and maintain separate, isolated queues and PHY/MAC processes in each AP. Even with a central controller, they are limited to heuristic steering and static power/channel tweaks.

Fi-Wi, by contrast:

• Places the queues and packet memory in the concentrator, not in room APs
• Makes every RRH part of a shared state space driven by centralized timing, CSI, queueing, and scheduling
• Uses dynamic RF grouping to decide which RRHs are in the same airtime domain and which are isolated cells
• Treats multi-RRH-per-room deployments as an opportunity for richer spatial streams and sensing, not as a management headache

A cell-per-room architecture makes Fi-Wi fundamentally different from controller-based Wi-Fi: it behaves more like cellular small cells with centralized coordination than like a set of autonomous APs.

To the WLAN practitioner, this density prescription reads as the exact opposite of hard-won field experience: in distributed-MAC deployments, over-density is the disease and removing APs is the cure. Appendix R takes that objection head-on — including why this is not a return to Meru/SCA-style single-channel coordination — and states the density claim in falsifiable, measurable form.

8. Queue Architecture for Fi-Wi

Fi-Wi centralizes packet memory, queueing, AQM, and TXOP scheduling inside the concentrator. Because the concentrator is the true bottleneck for all wireless transmissions, Fi-Wi can use a clean, minimal queue structure that behaves predictably under load and exposes stable delay semantics to L4S congestion controllers. This stands in contrast to traditional APs, where dozens of hidden queues (per-station, per-TID, firmware rings, retry/BA windows, PS-poll buffers, rate-control queues) produce variable and unobservable queueing delay.

This section describes Fi-Wi’s queue architecture, why WMM priority becomes largely unnecessary, and how centralized TXOP scheduling eliminates the stochastic contention that drives Wi-Fi collapse in legacy systems. The goal is simple: a minimal number of queues, explicit queue semantics, and predictable latency for all traffic classes.

8.1 Why queue architecture matters

Because all packets live in the concentrator’s memory until the moment they are transmitted over the air, Fi-Wi can explicitly control:

• exact queue depths per airtime domain,
• sojourn-time measurement for L4S AQM,
• TXOP length and pacing (typically 200–250 µs),
• which RRH transmits each TXOP, based on grouping and CSI.

This allows Fi-Wi to do what distributed APs cannot: construct a consistent, visible bottleneck queue that L4S congestion controllers can lock onto with stable behavior.

8.2 The theoretical case: L4S makes most priority obsolete

If queue delay is capped around 500 µs, legacy WMM categories provide little additional value. For example, consider a voice stream:

If L4S keeps queueing delay under ~500 µs, then all traffic — including voice — stays far inside its latency budget. WMM’s role in combatting bufferbloat disappears when bufferbloat itself is removed.

8.3 Practical complications

Three real-world issues motivate a cautious design:

• UDP does not respond to ECN

Voice and video often use UDP. They:

• do not slow down when marked with ECN,
• can starve TCP/QUIC if left unchecked,
• may need guardrails even when queue delay is low.

Fi-Wi can mitigate this using per-flow fair queuing inside the L4S queue, keeping UDP in check without needing a separate WMM hierarchy.

• Airtime vs. queue time

WMM historically manipulates AIFS, CW, and TXOP to reduce contention delay. Fi-Wi eliminates contention entirely using centralized TXOP scheduling, so WMM’s airtime hacks lose relevance.

• Failure modes and defense-in-depth

Even L4S can fail under:

• broken ECN in sender stacks,
• paths that zero or strip ECN bits,
• AQM or scheduler bugs during early rollout,
• extreme bursts that exceed control-loop bandwidth.

Hence, Fi-Wi benefits from a small amount of priority separation, at least in early deployments.

8.4 Minimal 3-queue structure

The theoretically sufficient minimal queue architecture for Fi-Wi is three queues:

• Management/Control Queue (Q_mgmt)
  Strict priority; beacons, probe responses, authentication, association, RTS/CTS, BA frames. Small, bounded, and never ECN-marked.
• L4S Queue (Q_L4S)
  ECT(1) traffic with dual-queue AQM; may use per-flow fair queuing internally.
• Classic Queue (Q_classic)
  ECT(0) and legacy traffic; part of the coupled AQM but with classic behavior.

In this design, WMM is unnecessary at the wireless bottleneck. All data traffic benefits from the same controlled queue delay, and fairness is enforced by per-flow scheduling rather than EDCA.

8.5 Pragmatic 5-queue structure

A more conservative deployment uses five queues per airtime domain:

1. Q_mgmt — Management & control (strict priority)
2. Q_L4S-hi — High-priority L4S (voice, control)
3. Q_classic-hi — High-priority classic (legacy VoIP)
4. Q_L4S-be — L4S best-effort (bulk QUIC/TCP)
5. Q_classic-be — Classic best-effort (legacy devices)

8.6 Numerical examples

Consider 10 simultaneous HD video calls (~20 Mbps total) plus a saturating background TCP flow:

Legacy WMM:

• Voice/video in AC_VI or AC_VO.
• Bulk traffic in AC_BE.
• Queues: video ~10 ms, best-effort ~100 ms under load.

Fi-Wi with L4S + fair queuing:

• 20 Mbps / 1 Gbps ≈ 2% of airtime.
• Queue held near 500 µs.
• Voice/video see ≈10–20 µs queue contribution after fair-share weighting.

This is roughly 1000× lower queueing latency than legacy WMM systems, and it applies to all traffic, not only traffic in a “priority” AC.

8.7 Deployment strategy

Fi-Wi can phase its queue structure over time:

• Phase 1 (Conservative): 5-queue system for operator comfort and defense-in-depth.
• Phase 2 (Aggressive): collapse to the minimal 3-queue system once production evidence confirms L4S stability across real workloads.

Metrics to monitor include:

• Voice MOS under background load,
• p99/p999 latency for all flows,
• video stall rates,
• behavior under intentional overload tests.

8.8 WMM support in Fi-Wi

WMM exists to correct three historical problems in distributed Wi-Fi:

• Bufferbloat (queues growing into tens or hundreds of milliseconds),
• UDP starving TCP due to non-responsive senders, and
• stochastic EDCA contention that produces unfair and unpredictable airtime.

Fi-Wi removes the root causes of these behaviors:

• Queueing lives exclusively in the concentrator, tightly bounded (200–500 µs).
• Per-flow fair queuing inside the L4S queue prevents UDP starvation.
• Centralized TXOP scheduling eliminates EDCA contention entirely.

Because of this, full WMM support at the air bottleneck is not necessary. However, Fi-Wi does support WMM semantics for:

• policy compatibility with existing enterprise QoS frameworks,
• client expectations (e.g., phones marking DSCP=EF),
• defense-in-depth during overload or L4S misbehavior.

Fi-Wi handles WMM as an admission-time mapping:

• WMM voice/video → high-priority L4S or Classic queues, depending on ECN capability.
• WMM best-effort → Q_L4S-be or Q_classic-be.
• WMM background → Classic best-effort.

This preserves compatibility while avoiding the complexity and unpredictability of EDCA-based priority systems. Over time, Fi-Wi deployments can rely on pure L4S semantics and collapse WMM to a compatibility shim, not a required scheduling mechanism.

8.9 Summary

Fi-Wi’s centralized queue architecture enables:

• a minimal 3-queue design (management, L4S, classic),
• a pragmatic 5-queue variant for early deployment,
• predictable delay semantics compatible with L4S,
• elimination of WMM’s complexity and EDCA’s stochastic behavior,
• clean integration with centralized TXOP scheduling.

Traditional Wi-Fi uses WMM to work around bufferbloat and contention. Fi-Wi removes those problems entirely through tight queue control, shared state, and central scheduling. Priority becomes a policy choice — not a crutch for an unstable MAC.

In Fi-Wi, the Carve-Out ensures the voice packet (L4S) bypasses the accumulated Classic bulk data completely. The file download continues to saturate the link, but the latency of the L4S flow is decoupled from the load of the Classic flow.

• Classic View: "I am getting 950 Mbps, and my ping is 20ms."
• L4S View: "I am getting 50 Mbps, and my ping is 1ms."
• Result: Both flows coexist on the same RF channel without the bulk flow destroying the real-time flow.

9. RRH-Level Active Redundancy, Dynamic Point Selection, and Intelligent Frequency Reuse

Fi-Wi’s centralized shared state across RRHs makes it natural to treat multiple radios as an active redundant set for the same STA or room. This is analogous in spirit to 802.11be’s Multi-Link Operation (MLO), where a single multi-link device (MLD) can use multiple links for reliability and capacity. In Fi-Wi, the concentrator is the coordination point leveraging shared state, and the RRHs are the distributed radios providing multiple RF paths.

9.1 Uplink: Duplicate Reception & Diversity

In many deployments, a client STA will be audible at more than one RRH (overlapping coverage). On the uplink, Fi-Wi exploits this spatial diversity to improve reliability without requiring changes to the client.

1. Multi-Point Reception: Multiple RRHs may receive the same MPDU from a transmitting STA, potentially at different SNR/MCS levels.
2. Forwarding: Each RRH decodes the frame locally. If the Frame Check Sequence (FCS) passes, the RRH timestamps the frame (using the shared global timebase), attaches metadata (RSSI, SNR, Channel State Information), and forwards it to the Concentrator via the PCIe-over-Fiber link.
3. Post-Detection Selection: Effectively, the Concentrator acts as a Post-Detection Selection Diversity combiner:
   • Identification: The Concentrator buffers incoming frames for a defined window (e.g., 100 µs) and identifies duplicates by matching the Source MAC and 802.11 Sequence Control fields (Sequence Number + Fragment Number).
   • Selection: If multiple RRHs capture the same frame successfully, the Concentrator compares their metadata and forwards only the single copy with the highest SNR, discarding the others.

This approach leverages the spatial diversity of distributed RRHs to mitigate shadowing and multipath fading. Because the selection logic operates on valid MAC frames (after FCS verification) rather than raw I/Q samples, this architecture maintains compatibility with standard COTS Wi-Fi silicon at the Radio Head. Appendix K discusses the information-theoretic ceiling above this baseline — joint multi-user decoding across synchronized RRH observations — and the architectural prerequisites Fi-Wi already satisfies for reaching it.

9.2 Downlink: per-packet steering

On the downlink, the concentrator can treat multiple RRHs as candidate transmitters for a given STA or room:

• Each downlink packet is enqueued once in the group queue for that airtime domain.
• The scheduler assigns the corresponding TXOP to one RRH (or occasionally duplicates across two RRHs) based on:
  • Recent CSI and SNR from each RRH to that STA
  • Per-RRH load and airtime usage
  • Policy (reliability vs. capacity vs. isolation)

This gives Fi-Wi:

• Per-packet (or per-TXOP) steering between RRHs, with the group queue as the single source of truth for the flow.
• Fast failover: if one RRH degrades (noise, hardware issue), the scheduler can switch the STA’s downlink traffic to another RRH in the same cellularized domain without changing IP or transport state.

9.2.1 Listen-Before-Talk (LBT) and RRH Eligibility for Downlink Scheduling

In a multi-RRH Fi-Wi deployment, each radio head operates on the same BSSID and channel but sits in a different physical location with its own RF conditions. While Fi-Wi centralizes all queueing and scheduling decisions, every RRH must still obey the fundamental 802.11 rule: listen-before-talk (LBT).

This is where Fi-Wi diverges sharply from classical multi-AP systems. In UniFi, Ruckus, Aruba, and all controller-based Wi-Fi architectures, each AP queue is blind to the RF medium state until it attempts to transmit. The AP commits a packet to the hardware queue, and if the medium is busy, the packet waits (Head-of-Line blocking) while the AP performs backoff.

Fi-Wi inverts this. RRHs continuously report their LBT Eligibility Status (Clear/Busy) to the Concentrator via the high-speed telemetry path. RRHs report LBT eligibility status via PCIe telemetry with update intervals of 100–500 µs, well-matched to inter-TXOP scheduling decisions. While the Concentrator cannot react within a single 9µs backoff slot, it operates on the Inter-TXOP timescale (200–500 µs¹).

Before posting a new DMA descriptor to an RRH, the Scheduler checks this eligibility:

• Eligible: The RRH reports the medium is statistically idle. The Concentrator posts the descriptor, "priming" that RRH for immediate transmission.
• Ineligible: The RRH reports heavy interference or high Energy Detect (ED). The Scheduler skips this RRH and routes the descriptor to an alternative RRH in the same airtime domain that reports clear air.

This prevents Head-of-Line Blocking where a packet sits in a hardware queue on a jammed radio. When multiple RRHs report clear airtime, Fi-Wi selects among them based on link quality (CSI) and predicted airtime efficiency. Conversely, if all RRHs report medium-busy, no RRH is primed; the scheduler pauses the flow to prevent backpressure from accumulating in the RRH hardware, keeping the queue depth visible in the Concentrator where L4S can measure it.

The result is a form of Centralized Selection based on LBT Eligibility. Multi-AP systems coordinate configuration (channels, power), but they cannot coordinate transmit starts because they lack the real-time feedback loop to steer packets away from busy radios before they are queued.

¹ Representative scheduling interval for mixed traffic workloads; actual TXOP durations range from tens of microseconds (small frames) to several milliseconds (large aggregates). ↩

9.3 Analogy to Wi-Fi 7 MLO

802.11be MLO allows a multi-link device (AP/STA) to use multiple links (e.g., 2.4G, 5G, 6G bands or channels) under a single MAC entity. Features include:

• Link aggregation: parallel links for higher throughput.
• Link redundancy: multi-link duplication for reliability.
• Per-packet steering: send each frame on the “best” link.

Fi-Wi provides a similar effect at the building scale, but with important differences:

• The coordination point is the Fi-Wi concentrator using shared state, not co-located with any single AP.
• The “links” are RRHs connected by deterministic fiber (or similar fronthaul) to the concentrator.
• The same centralized queues and L4S/AQM logic govern all RRHs in a cellularized domain.

Because the RRHs are spatially distributed around rooms and hallways, Fi-Wi gains advantages that co-located antennas cannot provide:

• More selection options: The Concentrator can choose from multiple spatially separated RRHs, each seeing different channel conditions to any given client
• Macro-diversity: Uplink transmissions received by multiple RRHs provide selection diversity that improves reliability at cell edges
• Spatial reuse opportunities: RRHs in different rooms can reuse the same channel when spatial isolation is sufficient (>25 dB attenuation)
• Better coverage uniformity: Multiple distributed RRHs eliminate dead zones and coverage gaps that single APs create

These advantages come from intelligent packet routing and dynamic RRH selection, not from RF phase coordination or simultaneous beamforming across RRHs.

9.3.1 Fi-Wi vs Wi-Fi 7 MLO: Compliance and Control

Fi-Wi strictly adheres to local regulatory compliance. The Concentrator manages the queue and the schedule, but the RRH manages the compliance.

When the Scheduler assigns a TXOP to an RRH, it posts a descriptor. The RRH hardware then performs standard 802.11 EDCA:

1. It senses the medium.
2. It draws a random backoff counter.
3. It counts down only when the medium is idle.
4. It transmits when the counter reaches zero.

The Architectural Difference:

In MLO or Mesh: If an AP commits a packet to a radio and that radio hits congestion, the packet is trapped in the local buffer. The backoff might take 50ms. During this time, the AP's other radios (or other APs in the mesh) might be idle, but they cannot help because the packet is already "owned" by the busy MAC.

In Fi-Wi: The packet remains in the Concentrator's central memory until the last possible moment (see Appendix F). If the Concentrator sees an RRH entering deep backoff (via real-time telemetry) or reporting "Busy," it stops posting new descriptors to that RRH and steers subsequent traffic to a free RRH. The backoff engine remains local (compliance), but the queue feeding it is steered globally (performance).

This allows Fi-Wi to scale airtime domains across an entire building while preventing the multi-node contention collapse that plagues traditional Wi-Fi networks.

9.4 Preserving the "single bottleneck" L4S view

To keep L4S happy, Fi-Wi needs to preserve a single bottleneck queue per flow even while using multiple RRHs:

• If two RRHs are in the same RF airtime group, they share the same group queue and marking logic. Uplink redundancy and downlink dynamic point selection do not create multiple bottlenecks — they just add spatial diversity and selection options within the same queue.
• If two RRHs are in different airtime domains (e.g., different channels or sufficiently isolated), the concentrator typically designates one domain as the primary bottleneck for that flow and uses the other only for failover or low-rate redundancy, to avoid creating competing bottlenecks for the same traffic.

In other words:

• Redundancy and spatial diversity are RF-level; queuing discipline remains centralized and unique per flow/domain.
• L4S still sees a single marking point (the group queue) for each path its packets traverse.

9.5 Dynamic Point Selection and Intelligent Frequency Reuse

Traditional Wi-Fi deployments suffer from two fundamental problems in high-density environments: (1) clients are statically associated to a single AP based on initial connection, leading to suboptimal performance as they move, and (2) autonomous APs compete for airtime through CSMA/CA contention, creating interference. Fi-Wi inverts this paradigm through Dynamic Point Selection—continuously choosing the optimal RRH per packet—and Intelligent Frequency Reuse—leveraging spatial isolation to maximize capacity.

9.5.1 Dynamic Point Selection: The Core Capability

Unlike traditional Wi-Fi where clients are physically and logically tied to a single Access Point (AP), Fi-Wi treats the entire building as a single Virtual Cell. The Concentrator maintains real-time Channel State Information (CSI) from all RRHs and dynamically selects the optimal transmission point for each individual packet.

10.1.1 The Roaming Paradigm Shift: Negotiation vs. Execution

To understand the magnitude of this shift, we must compare the standard "Fast BSS Transition" (802.11r) with the Fi-Wi approach. In standard Wi-Fi, mobility is a negotiation. In Fi-Wi, it is an execution.

While 802.11r is sufficient for buffered video (Netflix), it typically breaks real-time applications like Voice over Wi-Fi (VoWiFi) and VR/XR, where a 50ms gap causes audio dropouts or visual artifacts. Fi-Wi's sub-millisecond switching ensures true continuity.

10.1.2 How It Works

• Continuous CSI Ingestion: The Concentrator constantly reads memory-mapped uplink metadata (RSSI, SNR) from all RRHs. This is not a periodic report, but a real-time stream of the physical environment.
• Per-Packet Decision: For each downlink packet, the Scheduler examines current conditions and selects the RRH with the best channel to that specific client.
• Instant Routing: The packet is DMA-transferred via fiber to the selected RRH, which performs standard CSMA/CA and transmits.
• Seamless Adaptation: As the client moves or RF conditions change (e.g., a door closes), the Concentrator automatically switches to a different RRH.

10.1.3 Example Scenario

Consider "Alice" on a VR headset walking down a hallway:

1. Alice starts a session in Room 304 (near RRH-A: RSSI -40 dBm). The Concentrator routes packets via RRH-A.
2. Alice walks toward the doorway. RRH-A degrades (-55 dBm) while the hallway unit, RRH-B, improves (-45 dBm).
3. The Concentrator detects this crossing point in the CSI data.
4. For the very next packet, the pointer switches to RRH-B.
5. Result: Alice's VR stream continues without a single dropped frame or latency spike. She is unaware that the transmission point changed.

9.5.3 Intelligent Frequency Reuse

In traditional Wi-Fi, neighboring APs on the same channel create co-channel interference. The standard solution is to assign different channels (e.g., AP-A uses Channel 36, AP-B uses Channel 48), but this wastes spectrum. Fi-Wi enables intelligent frequency reuse—using the same channel across multiple RRHs when spatial conditions allow.

When Frequency Reuse Works

Frequency reuse is viable when clients are in spatially separated locations with significant isolation (typically >25-30 dB attenuation due to walls, floors, or distance).

Example: Adjacent Rooms

• Client A in Room 304 (served by RRH-A, RSSI: -42 dBm)
• Client B in Room 305 (served by RRH-B, RSSI: -45 dBm)
• Cross-coupling: RRH-A → Client B = -75 dBm, RRH-B → Client A = -78 dBm
• Isolation: ~30-33 dB (wall attenuation)

The Fi-Wi Decision:

1. Concentrator detects >30 dB spatial isolation via CSI measurements
2. Configures both RRH-A and RRH-B to operate on Channel 36
3. Each RRH performs independent CSMA/CA in its local environment
4. Cross-interference is minimal due to spatial isolation
5. Result: Effective channel capacity is doubled without requiring additional spectrum

Dynamic Adaptation

The key advantage over static channel planning is real-time adaptation:

• Movement Detection: If Client A moves toward doorway, cross-coupling increases (e.g., -75 dBm → -55 dBm)
• Automatic Response: Concentrator detects degraded isolation via updated CSI
• Mitigation Options:
  • Switch one RRH to different channel (Channel 36 → Channel 48)
  • Use time-division: Alternate which RRH transmits in each TXOP window
  • Apply dynamic point selection: Route all traffic through whichever RRH provides best path to both clients

Why Autonomous APs Cannot Do This

9.5.4 Transparent Integration with L4S

The complexity of dynamic point selection and frequency reuse is hidden from the L4S congestion control loop. Traffic still lives in per-airtime-domain group queues. When the Concentrator enables frequency reuse or optimizes RRH selection, it simply affects the effective service rate μ(t) of the queue.

The PI² controller in the outer loop (see Section 6) sees the queue draining faster and naturally reduces ECN marking. This allows L4S senders (TCP Prague) to ramp up their congestion windows to fill the expanded capacity. The system automatically discovers and exploits available spatial capacity without requiring changes to congestion control algorithms or application awareness.

9.5.6 What Dynamic Point Selection Does NOT Enable

To maintain technical accuracy, it is important to clarify what Fi-Wi's dynamic point selection does not provide:

• Not Distributed MIMO: Fi-Wi does not perform simultaneous coherent transmission from multiple RRHs. Each transmission originates from a single selected RRH.
• Not Beamforming: While the Concentrator selects the optimal RRH, it does not control RF phase or perform spatial multiplexing across RRHs. Each RRH operates independently.
• Not Coordinated Simultaneous Transmission: When using frequency reuse, each RRH performs independent CSMA/CA. There is no explicit scheduling or coordination of transmission timing—the coordination comes from routing decisions and channel assignments.

These capabilities would require either:

• Custom ASIC development with externally-controllable RF synthesizers (for phase-coherent MIMO)
• Or extensive firmware modification to enable TSF-scheduled transmission (regulatory challenges in unlicensed spectrum)

Fi-Wi's architecture deliberately focuses on capabilities achievable with COTS Wi-Fi chips, providing 2-3x capacity improvement through intelligent management rather than pursuing 4-6x gains that would require custom silicon development.

9.5.7 Performance Expectations

Based on the capabilities described above, Fi-Wi provides the following performance improvements over traditional autonomous AP deployments:

• 2-3x Capacity Increase: Through intelligent frequency reuse and reduced CSMA/CA contention
• 50-80% Latency Reduction: Via dynamic point selection eliminating handoff delays
• 90%+ Cell Edge Reliability: Through selection diversity on uplink
• Zero Handoff Delay: Seamless mobility without 802.11 reassociation

These gains are achieved through centralized intelligence and microsecond-latency fronthaul, not through RF phase control or coordinated transmission. The architecture remains fully compliant with unlicensed spectrum regulations and works with commodity Wi-Fi chipsets.

9.5.8 Summary

Fi-Wi transforms the problem of wireless density by treating it as a routing and scheduling problem rather than an RF coordination problem. By centralizing packet memory and MAC scheduling, Fi-Wi converts adjacent radios from interferers into dynamically selected access points, allowing the network to scale capacity through intelligent management rather than collapsing under interference.

The key insight is that most Wi-Fi performance problems stem from poor decisions (wrong AP, wrong channel, wrong timing) rather than fundamental RF limitations. Fi-Wi solves this by providing the Concentrator with complete visibility and control, enabling microsecond-granularity optimization that autonomous APs cannot match.

10. The Three Clock Domains

10.1 Overview

Fi-Wi operates across three distinct clock domains simultaneously. Each domain has different coherence requirements, different distribution mechanisms, and different consequences for the receive array capability tiers the system can support. Confusing them leads to both overspecifying the easy case and underspecifying the hard one.

10.2 The Master Clock

Clock synchronization has two distinct jobs. The first is a trustworthy source of time. The second is delivering that time to where it is needed without quietly discarding the precision along the way. Both matter. A trustworthy source feeding sloppy delivery wastes the source. Precise delivery of an untrustworthy source is worse.

Fi-Wi's answers to both jobs are straightforward for the building-internal use case.

The source. The concentrator IS the grandmaster for the Fi-Wi domain. It maintains a free-wheeling timebase from its motherboard oscillator (or an optional higher-stability oscillator in a hardened deployment), optionally disciplined to GPS or an external PTP grandmaster if absolute wall-clock time is required by the deployment. In the baseline UAX-8 deployment, GPS disciplining is not required — Fi-Wi does not need UTC traceability. It needs internal coherence across RRHs for scheduling, TSF alignment, and AQM timestamps. The concentrator's clock only needs to be stable enough that the affine correction model stays valid between sample windows. A commodity motherboard oscillator is adequate for this. There is no external grandmaster in the baseline deployment.

The delivery. PCIe PTM (Precision Time Measurement) delivers 20–50 ns timing to RRHs over the same fiber that carries data — no separate timing network, no switch jitter, no PTP boundary clock accumulation error. PCIe PTM uses hardware timestamping at the PCIe transaction layer, not software-assisted PTP. For intra-building distances this puts it in the same class as White Rabbit in terms of delivery fidelity, without requiring a separate fiber timing plane or TSN-capable switching infrastructure. The precision is not discarded in transit.

This coordinate system — concentrator as source, PCIe PTM as delivery — is the frame in which all scheduling decisions, packet timestamps, sojourn time measurements, AQM marking decisions, and ML training labels are expressed. Its coherence requirement is microsecond-class: tight enough for scheduling loop closure and ECN marking stability, not tight enough to drive RF sampling (see §10.4 for the sampling clock domain).

10.2.1 Clock Hardware Chain: PCIe PTM to RRH

The "cellularization" of Wi-Fi relies on a unified timebase. In the Fi-Wi architecture, time is not merely used for logging; it is a control variable. To achieve coordinated scheduling, accurate queue measurements, and seamless mobility, every RRH must share a precise understanding of "now" down to the microsecond level.

To achieve this, Fi-Wi establishes a strict Hierarchical Clock Tree over the PCIe fronthaul, leveraging the native determinism of the bus rather than the best-effort nature of packet switching.

10.2.2 The Concentrator as Grandmaster (GM)

The Fi-Wi Concentrator acts as the timing grandmaster for the entire building, distributing its timebase over the PCIe fronthaul via PCIe PTM rather than network PTP (PTP/802.1AS remains available as an optional external reference — see §5.1.1). It houses the primary reference oscillator — the commodity motherboard oscillator in the baseline deployment (see §10.2), or optionally a high-stability oscillator (e.g. OCXO) in hardened deployments.

• Downstream Distribution: The Concentrator distributes the master clock reference to all connected RRHs via the fronthaul links.
• PCIe Precision Time Measurement (PTM): Over the PCIe-over-fiber links, Fi-Wi leverages PCIe PTM (a hardware-native feature since PCIe Gen 3) to continuously calculate and compensate for propagation delay variations in the fiber.
• RRH Frequency Lock: Each RRH maintains a local clock disciplined to the Concentrator's master clock — PCIe PTM for time synchronization, plus a distributed frequency reference (see §10.2.6 and §10.5) — ensuring frequency stability and preventing long-term drift.

10.2.3 What Clock Synchronization Actually Enables

A defining advantage of the Fi-Wi architecture is the use of "Hard Synchronization" via PCIe, rather than "Soft Synchronization" via Ethernet. While Ethernet-based APs rely on IEEE 1588 PTP, they are subject to switch jitter and software stack latency. PCIe over fiber eliminates these variables.

Important Note: While frequency-locked clocks provide excellent timing consistency, they do not enable RF phase control or coordinated simultaneous transmission. COTS Wi-Fi chips have independent RF synthesizers with arbitrary phase offsets that cannot be controlled externally. The value of clock synchronization lies in accurate timestamping for L4S queue management and consistent TSF counters for seamless client mobility, not in RF phase alignment.

10.2.4 Operating Modes: GPS-Disciplined vs. Free-Wheeling

The Concentrator's clock behavior depends on the deployment environment and regulatory requirements. There are two distinct modes of operation:

Mode A: GPS-Disciplined (Absolute Synchronization)

In this mode, the Concentrator is connected to an external GNSS (GPS/Galileo) receiver. The internal oscillator is disciplined to align with UTC (Coordinated Universal Time). This connects the internal timing of the Fi-Wi system to external absolute time.

Mode B: Free-Wheeling (Relative Synchronization)

In deep indoor environments (basements, bunkers) where GPS is unavailable, or cost-sensitive deployments where 6 GHz AFC is not required, the Concentrator operates in Free-Wheeling mode.

10.2.5 When Absolute Time Becomes Mandatory

While Free-Wheeling mode is sufficient for core system operation, GPS-Disciplined (Absolute) mode becomes mandatory when the Fi-Wi system interacts with external systems that require UTC timestamps:

1. 6 GHz AFC (Automated Frequency Coordination): To operate at Standard Power in the 6 GHz band (essential for outdoor or large-venue coverage), the FCC requires the system to check a central database for incumbent microwave links. The database operates on UTC. The Concentrator must sign its request with a precise, absolute timestamp and geolocation. A drifting clock will cause the AFC request to be rejected, forcing the system into Low Power Indoor (LPI) mode.
2. Inter-Concentrator Handoffs (Multi-Building Roaming): In a campus environment with two distinct Concentrators (e.g., Building A and Building B), a client roaming between them may experience time jumps. If Concentrator A and B are free-wheeling independently, their timestamps may differ by seconds. This jump can break high-level security protocols (like Kerberos or 802.1X re-authentication) that reject "replay attacks" based on timestamp windows.
3. Correlated Debugging: If a user reports a connectivity drop at 10:04 AM, but the Concentrator has drifted by 45 seconds, the system logs will be stamped 10:04:45. Correlating Fi-Wi logs with client-side logs (which are usually synced to NTP/Cellular time) becomes operationally difficult, complicating root-cause analysis.

10.2.6 RRH Clock Distribution Hardware

Standard enterprise APs utilize free-running crystal oscillators with ~20 ppm frequency error. This causes TSF counters to drift relative to each other, making seamless mobility difficult. To achieve the timing consistency required for Fi-Wi's coordinated operation, the RRH hardware architecture must be fundamentally different.

The Fi-Wi Solution: The RRH hardware uses Mobile-Class Wi-Fi Silicon (which natively supports external clock inputs) driven by a Fronthaul-Recovered Precision Clock.

The clock distribution chain operates as follows:

1. Concentrator (Grandmaster): Distributes the master timebase via PTM packets over the PCIe-over-fiber link.
2. RRH FPGA / Retimer: Recovers the implicit clock from the PCIe bitstream or explicit PTM timestamps.
3. Network Synchronizer (Jitter Attenuator):
   • Component: a dedicated network-synchronizer / jitter-cleaner IC.
   • Function: Feeds the "dirty" recovered clock digitally into this dedicated IC.
   • Cleaning: The IC uses an internal, narrow-bandwidth DSP servo loop to filter out PCIe transport jitter, synthesizing a pristine 40 MHz reference.
4. Wi-Fi SoC (Client SKU): The cleaned signal is fed directly into the chip's Ext_Ref / XO_IN pin. The chip's internal PLLs lock to this external frequency reference, ensuring consistent TSF counter operation across all RRHs.

Note — time synchronization vs. frequency distribution: PCIe PTM synchronizes time (the timestamps used for scheduling, sojourn measurement, and TSF affine correction). The frequency reference feeding the jitter attenuator and the chipset's Ext_Ref pin is a separate concern with its own delivery path. Recovering that frequency reference from the PCIe bitstream itself, as depicted by the clock-recovery stage in Diagram 10.2-2, corresponds to Option 1 in §10.5 and is a Phase 2 candidate — no discrete PCIe CDR part exists today. Phase 1 prototyping instead uses explicit clock distribution over a dedicated fiber (Option 2, §10.5), which feeds the same jitter-attenuator → Ext_Ref chain shown above.

10.2.7 Required Chipset Capabilities

The RRH requires three chipset capabilities; any Wi-Fi part that provides them is a candidate, regardless of SKU class. The first — an external clock input — is the one that matters for distributed operation, because it is what lets every RRH frequency-lock to the concentrator timebase and keep TSF counters consistent across radios. Client/mobile-class Wi-Fi silicon commonly exposes all three, which is why early Fi-Wi RRHs use such parts, but the requirement is the capability, not the SKU label.

• External Clock Input (required): The chipset must accept an external frequency reference rather than relying solely on a passive on-chip crystal. Fi-Wi drives each RRH's reference from the concentrator timebase so that all radios are frequency-locked; this is what makes TSF counters consistent across RRHs. Any chip that exposes an external reference-clock input (often labeled Ext_Ref, XO_IN, or equivalent) qualifies — the dependency is on this clocking path, not on the chip being a phone or laptop part. Chips that expect only a passive crystal resonator cannot be frequency-locked this way and are unsuitable.
• PCIe Endpoint: The chipset must present as a PCIe endpoint so it can be reached over the PCIe-over-fiber fronthaul and integrated into the Fi-Wi DMA path. Many client-class radios are already native PCIe endpoints.
• Power and Thermal Envelope: The radio must fit a fanless, cell-per-room RRH enclosure — on the order of a few watts. Parts designed for power-constrained devices meet this envelope comfortably.

10.2.8 What Frequency-Locked Clocks Enable and What Requires More

Frequency-locked clocks via PCIe PTM enable consistent TSF counter operation across all RRHs, stable L4S timestamps, seamless mobility, and accurate queue measurements. They are sufficient for Tier 1 (selection diversity) and Tier 2 (uplink MRC) of the receive array capability tiers described in §10.4.2.

Tier 3 — distributed transmit beamforming where multiple RRHs transmit the same waveform for constructive combining at a legacy STA — requires sampling clock coherence at RF wavelength scale (approximately 50 ps at 5 GHz). This is beyond what the frequency-locked clock chain described in this section provides. It is a Phase 2/3 capability requiring CDR integration or custom ASIC work. See §10.4.2 for the full three-tier framework and §10.5 for clock distribution options.

10.3 The TSF Domain

The 802.11 TSF (Timing Synchronization Function) is the only clock the 802.11 MAC standard exposes to clients. In Fi-Wi, the concentrator generates all beacon frames centrally, deriving TSF values from the master clock via a known affine transform per RRH. Within a given BSS, TSF values are identical across all participating RRHs by construction. The TSF domain is a MAC-layer correctness requirement, orthogonal to the sampling clock domain. See §5.1.2 and §5.1.3 for the full treatment.

TSF coherence does not imply RF sampling coherence. RF sampling coherence does not require TSF alignment. They are orthogonal problems that share a common reference: the master clock.

10.4 The Sampling Clock Domain

The sampling clock drives the ADC and DAC in each RRH — the clock that determines when the RF front end samples the incoming signal and when it emits the outgoing waveform. This is the clock domain that governs RF coherence across RRHs. Its requirements depend on what the concentrator is doing with the multi-RRH receive paths, and on where the functional split between RRH and concentrator is placed.

10.4.1 O-RAN Functional Split and Clock Implications

The O-RAN architecture defines functional split points that determine how much PHY processing stays at the radio versus moves to the central unit. The split point determines what the fronthaul carries and what sampling clock coherence is required.

At Split 8 (CPRI), raw IQ samples cross the fronthaul. All PHY processing — FFT, channel estimation, equalization, decoding — happens at the concentrator. The RRH is a pure RF and analog front end. Because the concentrator is combining raw IQ from multiple RRHs, the sampling clocks must be coherent at RF wavelength scale. At 5 GHz a quarter wavelength is approximately 1.5 cm, which corresponds to approximately 50 ps. This is the hardest clock coherence requirement.

At Split 7.2x (eCPRI), lower PHY processing — FFT, OFDM demodulation, some channel estimation — happens at the RRH. Frequency-domain symbols cross the fronthaul. The concentrator performs MRC combining and decoding from frequency-domain inputs. The sampling clocks must be frequency-locked and phase-stable across a PPDU, but the concentrator can estimate per-RRH channel phase from the Wi-Fi preamble on every packet. The requirement is phase stability during the packet, not fixed 50 ps absolute alignment across the building. For a 1 ms PPDU at 5 GHz with a 10 ppb inter-RRH frequency offset, the accumulated phase drift is approximately 18 degrees — within the estimation range of the preamble equalizer.

At Split 6 and above, more PHY processing stays at the RRH. Coherent combining at the concentrator is not possible at this split. Sampling clocks can be more independent.

Fi-Wi targets Split 7.2x for Phase 1. The RRH performs OFDM demodulation and passes frequency-domain IQ to the concentrator for MRC combining and decoding.

10.4.2 Three Receive Array Capability Tiers

The sampling clock coherence requirement depends on which receive array tier the system is operating in.

Tier 1 — Selection Diversity. The concentrator receives from multiple RRHs and selects the best decode. No sampling clock coherence required. Each RRH independently demodulates and decodes. Lowers PER because at least one RRH is likely to hear the STA cleanly. Achievable with independent per-RRH clocks. Available in Phase 1.

Tier 2 — Uplink Maximal Ratio Combining (MRC). The concentrator combines frequency-domain IQ or soft symbols from multiple RRHs. Requires frequency-locked sampling clocks and phase stability across the PPDU. The concentrator estimates per-RRH channel phase from the Wi-Fi preamble on every packet using the L-STF, L-LTF, and HT/VHT/HE preamble fields. Higher SNR from combining drives higher MCS, shorter TXOPs, lower PER, and reduced airtime per unit of data delivered — directly reducing contention pressure on legacy devices. Available in Phase 1 with appropriate clock distribution.

Tier 3 — Distributed Transmit Beamforming. Multiple RRHs transmit the same waveform to a single legacy STA with the goal of constructive combining at the STA. Requires sampling clock coherence at RF wavelength scale: approximately 50 ps at 5 GHz. This is the Split 8 class problem. Phase 2 or Phase 3 capability.

10.5 Clock Distribution Options

Three approaches to distributing a coherent sampling clock reference from the concentrator to the RRHs.

Option 1 — Clock and Data Recovery (CDR) from the PCIe Stream. CDR extracts a synchronized clock from the PCIe data stream at the RRH without a separate clock wire. Whether the recovered clock can drive the ADC/DAC sampling input directly, or whether a jitter cleaner is required, depends on CDR output jitter and chipset clock input tolerance. No discrete PCIe CDR chip exists at present; requires FPGA or custom ASIC integration. Phase 2 candidate.

Option 2 — Explicit Clock Distribution over Fiber. A master oscillator at the concentrator drives a fan-out clock buffer distributing the reference to each RRH over a dedicated fiber alongside the PCIe-over-fiber fronthaul. Propagation delay is deterministic and measurable; the concentrator applies per-RRH timing corrections for fiber length differences. Sufficient for Tier 2 uplink MRC. Recommended for Phase 1 prototyping. See §10.2.6 and §10.2.7 for the full hardware chain including the digital jitter attenuator and mobile chipset Ext_Ref pin.

Option 3 — Over-the-Air CW Reference. One RRH designated master transmits a continuous wave reference; all others phase-lock to it. No additional wiring required. Introduces dependency on RF path conditions. Better suited to outdoor distributed antenna scenarios than indoor building deployments.

11. Fi-Wi value vs. Traditional Distributed APs

Modern enterprise Wi-Fi deployments use centralized controllers (Cisco WLC, Aruba Mobility Controller, Ubiquiti UniFi, Ruckus SmartZone, etc.) to manage multiple APs. These controllers coordinate the control plane: channel assignment, transmit power, client association hints, roaming policies, and security. However, these remain loosely-coupled systems where the data plane — queueing, MAC scheduling, aggregation, and packet memory — remains distributed inside individual APs.

A traditional AP is not just “running EDCA.” It is running EDCA after juggling dozens or hundreds of logical MAC queues and state machines:

• Per-station queues (one or more per associated STA)
• Per-TID / per-WMM-class queues (4–8 AC/TIDs per station)
• Firmware queues for rate control, reordering, and AMPDU assembly
• Internal MAC hardware ring buffers for TX/RX DMA
• Rate-control queues tracking MCS and retry history
• PS-poll / U-APSD buffered-traffic queues for sleeping stations
• BAR/BA state (BlockAck windows of outstanding MPDUs)

With N stations, an AP can easily have on the order of N × (4–8) logical queues behind a single RF channel. Every AP in the same RF domain runs this large, isolated, queue-filled state machine independently. No AP has a global view; controllers see only coarse statistics.

The result:

• The bottleneck is smeared across many hidden queues, so L4S cannot see or control it.
• APs make inconsistent decisions, so collapse emerges at different times across APs.
• AP-side buffering encourages long TXOPs when an AP finally wins the medium, driving multi-millisecond bursts.

Fi-Wi is fundamentally different: it centralizes both control plane and data plane with shared state across all RRHs. The concentrator does not just configure RRHs; it directly manages their queues, schedules their TXOPs, maintains unified CSI and airtime state, and applies coordinated ECN marking for each airtime domain. This architectural difference — not just improved control-plane coordination — is what enables Fi-Wi’s latency, L4S, and spatial multiplexing advantages.

The following subsections detail specific benefits of Fi-Wi’s cellularized, tightly-coupled architecture compared to controller-managed, loosely-coupled AP systems.

11.1 Deterministic low latency

Traditional APs:

Each AP builds its own local queues. Under load, large aggregates, retries, and hidden buffering produce multi-millisecond queueing and service delays. Tail latency is largely uncontrolled, and varies across APs sharing the same channel.

Fi-Wi (cellularized Wi-Fi, cell-per-room):

• There is exactly one deep queue per airtime domain (RF group of interfering RRHs), not per AP.
• That queue is fully visible and centrally controlled in the concentrator.
• Sojourn-time or PI2-style control targets sub-millisecond delay (e.g., 200–400 µs) including tails.
• Concentrator pacing prevents upstream buffers from filling while respecting L4S behavior.

11.2 Stable L4S behavior

Traditional APs:

L4S flows traverse multiple hidden queues: wired bottlenecks, AP-local queues, firmware queues, and EDCA contention. ECN marking (if it exists at all) is inconsistent and not tied to a single bottleneck. Collapse produces noisy, bursty marking or loss, and the L4S control loop becomes oscillatory or falls back toward classic congestion behavior, especially in the tails that matter to users.

Fi-Wi:

• There is exactly one main marking point per airtime domain — the centralized group queue.
• ECN marks are proportional to queue delay (sojourn time / PI2) in that queue.
• The control loop sees a small number of well-behaved bottlenecks, which is precisely the design assumption behind L4S.

11.3 Aggregation without losing visibility

Traditional APs:

Aggregation improves PHY efficiency but hides individual packet timing from the congestion controller. The controller does not know which MSDUs were grouped into a TXOP, what the queue state was when the TXOP started, or how long each device has been waiting.

Fi-Wi:

• Aggregation is retained for efficiency, but the concentrator knows which MSDUs shared a TXOP.
• The queue state at TXOP start is known for each cellularized domain.
• ECN marks can be assigned within aggregates in a controlled way, preserving fine-grained congestion information.

This combination yields high PHY efficiency and transport-layer visibility into congestion, instead of having to choose one or the other.

11.4 Building-scale coordination

Controller-managed loosely-coupled APs:

The controller can adjust channels, power, and send steering hints (e.g., 802.11v), but it cannot see or control:

• Per-packet queue states across APs — each AP's queue depth, sojourn times, and retries remain opaque to the controller.
• MAC aggregation and TXOP decisions — each AP independently decides when to send aggregates and how long TXOPs last.
• Real-time CSI or spatial stream structure — CSI stays local to each AP's antenna array; the controller sees only coarse stats ("MCS 9, 80 MHz").
• Cross-AP packet routing — the controller cannot dynamically route individual packets to different APs based on instantaneous RF conditions.
• Coordinated frequency reuse — APs cannot adapt channel usage in real-time based on spatial isolation measurements.

As a result, these systems rely on heuristic, reactive policies: channel reassignment after interference is observed, power adjustments based on neighbor reports, and client steering using RSSI or airtime snapshots. These help, but they operate on coarse time scales (seconds to minutes) and cannot fix the fundamental data-plane issues of distributed queues, MAC contention, and tail latency under load.

Fi-Wi cellularized architecture:

The concentrator maintains true shared state across all RRHs in the building:

• Centralized queues per airtime domain — every packet's enqueue time, queue depth, and marking state is visible.
• Unified timestamp reference — µs-accurate synchronization enables precise sojourn-time measurement and coordinated scheduling.
• Aggregated CSI from all RRHs — the concentrator sees channel quality from every RRH to every active STA, enabling dynamic point selection.
• Coordinated MAC scheduling — TXOP timing, aggregation policies, and RRH selection are decided centrally with full visibility.
• Dynamic RF grouping — RRH-to-queue assignments adapt in real time based on measured interference, CSI, and load conditions.
• Dynamic point selection — each packet is routed to the optimal RRH based on current channel conditions, providing seamless mobility and selection diversity.

Because RRHs are distributed in space (often 2–4 per room in high-density deployments), Fi-Wi can leverage spatial separation for intelligent frequency reuse. The concentrator sees CSI from all RRHs and can make microsecond-granularity decisions about which RRH should transmit each packet — all while preserving the "single bottleneck queue per airtime domain" discipline required for stable L4S behavior.

11.5 Control Plane vs. Data Plane

The table below summarizes the architectural differences between controller-managed, loosely-coupled APs and Fi-Wi's cellularized, tightly-coupled architecture:

11.6 Operational and lifecycle advantages

Controller-managed loosely-coupled APs:

• Complex distributed software updates: each AP runs its own firmware image. Updating a building with 50+ APs requires orchestrated rollouts, version compatibility checks, and managing multi-version fleets.
• Per-AP troubleshooting: queue states, retries, and airtime stats are scattered across devices with limited visibility.
• Hardware scaling complexity: each new AP adds another autonomous system with its own memory, CPU, queue management, and potential firmware bugs.
• Failure diagnosis: hard to distinguish AP hardware failure from configuration or RF environment issues; local state is often incomplete.
• Massive Security Surface: Each AP runs a full Linux OS with millions of lines of code, open ports, and local services. A vulnerability in the kernel or a driver compromises the edge device.

Fi-Wi cellularized architecture:

• Simplified RRH hardware: RRHs are thin RF endpoints: RF front end, DMA engine, PTP sync, and a minimal control shim. No complex queue management, routing, or control logic.
• Centralized software complexity: the concentrator contains control logic, shared-state machinery, AQM, scheduling, and learning models. Software updates and new features happen at the concentrator, not across dozens of RRHs.
• Unified diagnostics and visibility: all queues, timestamps, CSI, and per-packet metadata flow through the concentrator:
  • View real-time queue depths and sojourn times per airtime domain.
  • See per-STA, per-RRH, or per-flow latency histograms (median, p95, p99).
  • Trace individual packets from ingress to air transmission with µs-accurate timestamps.
  • Identify which RRH, spatial stream, or MCS is causing retries or collapse.
• Scaling simplicity: adding RRHs to increase density (e.g., 2–4 per room) does not add new control complexity. The concentrator’s logic remains the same; only RF groupings change.
• Predictable failure modes:
  • RRH failure: the concentrator detects loss of DMA/PTP, removes that RRH from its airtime group, and redistributes its STAs to other RRHs in the same domain; queues and flow state remain intact.
  • Concentrator failure: in HA deployments, a standby concentrator with synchronized state can take over; RRHs remain thin endpoints.
• Installation and provisioning: RRHs are auto-configured by the concentrator (channel, power, beacon content, group membership, and policies). Installers mount RRHs, connect fiber, and the concentrator handles the rest—no per-RRH SSH or web UI.
• Cost structure advantages: RRH hardware is simpler and cheaper than full APs; software development and testing concentrate on one platform; support and diagnostics are centralized.
• Zero Security Surface at the Edge: RRHs run **zero lines of Linux**. They have no shell, no open network ports, and no OS to exploit. They are fixed-function hardware micro-bridges.

13. PCIe Fronthaul (Gen3 x1 over Fiber)

13.1 Why PCIe as the RRH interface

A central hardware design choice is to make the RRH look like a PCIe endpoint to the Fi-Wi concentrator. This leverages the fact that:

• All modern Wi-Fi chips speak PCIe already, especially mobile / laptop-class devices.
• The Wi-Fi silicon is validated with PCIe as its primary host interface.

Benefits of this choice:

• No need to invent a new device-side PHY or host interface.
• Reuse of commodity controller logic, driver models, and test infrastructure.
• Natural mapping to our DMA-based RRH model:
  • RRH presents DMA queues for Tx/Rx over PCIe.
  • Concentrator holds packet memory and group queues in DRAM.

We start with PCIe Gen3, one lane (x1), carried over fiber via a retimer + optical interface. Higher generations or widths (Gen4, x2/x4) are possible later but not required for the initial Fi-Wi performance targets.

13.2 Gen3 x1 throughput

PCIe Gen3 provides:

• Raw signaling rate: 8.0 GT/s per lane.
• 128b/130b encoding → effective line rate ≈ 7.88 Gb/s.

After protocol overhead (TLP headers, DLLPs, flow control), the sustained payload throughput for Gen3 x1 is in the rough range of 6–7 Gb/s for large transfers. This is more than sufficient for:

• Feeding a single RRH with:
  • Multi-Gbps Wi-Fi PHY peaks and realistic MAC efficiency.
  • CSI export and control traffic.
• Supporting the expected 2.4 GHz + 5/6 GHz bands per RRH with headroom.

If a future RRH design must exceed this, the same architecture scales to:

• Gen4 x1 (doubles per-lane throughput), or
• Multi-lane links (x2, x4) for very high-end RRHs.

For our initial Fi-Wi deployment assumptions, Gen3 x1 over fiber is a sensible and sufficient starting point.

13.3 Latency characteristics and budget

PCIe Gen3 latency has several components:

• Transaction layer and data link layer processing.
• PHY serialization / deserialization.
• Retimer pipeline delay.
• Fiber propagation.

Order-of-magnitude:

• PCIe controller + retimer pipeline: microsecond-scale latency per transaction (a handful of µs).
• Fiber propagation: ~5 µs/km, so tens to a few hundred meters of in-building fiber adds only tens to hundreds of nanoseconds.

Compared to:

• L4S RTTs (tens of ms, WAN) or a few ms (LAN).
• MAC TXOP durations (hundreds of µs).
• Queue sojourn-time targets (200–400 µs at group queues).

the PCIe-over-fiber latency is effectively negligible. It comfortably fits within the microsecond-level time base used for:

• t_ingress_us timestamps in FiWiMeta.
• RRH transmit time stamps.
• Queue-delay computation and ECN marking.

13.4 Mapping queues and metadata

The PCIe model fits naturally with the Fi-Wi queueing and metadata scheme. Each RRH behaves like a PCIe endpoint with:

• Downlink DMA rings: concentrator → RRH descriptors, pointing to packets in central DRAM.
• Uplink DMA rings: RRH → concentrator descriptors, carrying received frames and CSI/stats.

The FiWiMeta header lives in host memory adjacent to packet payloads and is referenced by these descriptors.

Downlink flow:

1. Concentrator enqueues IP/Ethernet packets into a group queue in DRAM, allocates or updates FiWiMeta (including t_ingress_us and queue snapshot).
2. Scheduler posts PCIe descriptors to the RRH for the next TXOP, selecting which MSDUs and which RF group/airtime domain.
3. RRH DMA-fetches the MSDUs via Gen3 x1, builds an aggregate (A-MPDU), transmits over the air, and reports:
   • Success / retry / PER data
   • CSI and MCS used
   • TX time / TXOP info

Uplink flow:

1. RRH receives 802.11 frames from STAs, decodes them, and attaches CSI and MAC status.
2. RRH DMA-writes the frames + metadata into concentrator DRAM via PCIe.
3. Concentrator:
   • Inserts packets into the appropriate group queues.
   • Updates per-flow and per-domain control state.
   • Feeds CSI and MAC outcomes into the learning model (Appendix B).

In both directions, the PCIe fronthaul:

• Provides more than enough throughput for RRH traffic and control.
• Provides low and stable latency consistent with microsecond-level timing.
• Aligns perfectly with the central group-queue and FiWiMeta assumptions of the control-plane design.

13.5 PCIe Hot Swap

A critical operational requirement for Fi-Wi is the ability to service, replace, or add RRHs without bringing down the entire building's wireless network. PCIe provides native support for this through hot-plug capability, which is standard in enterprise server platforms and can be leveraged for Fi-Wi deployments.

13.5.1 Hot-plug fundamentals

PCIe hot-plug allows physical insertion and removal of endpoint devices (RRHs) while the system is running:

• Hardware signaling: Hot-plug controllers detect physical presence via dedicated signals (PRSNT# pins) or through link training state changes.
• Surprise removal: Modern PCIe supports surprise removal where endpoints disappear without software notification—critical for field environments where RRHs may lose power or have cable disconnections.
• Software enumeration: The concentrator's PCIe root complex can dynamically enumerate new RRHs when they appear and gracefully handle removal of existing RRHs.

13.5.2 RRH insertion flow

When a new RRH is connected or powered on:

1. Physical detection: PCIe hot-plug controller detects the new device via link training.
2. Enumeration: Concentrator OS (Linux) enumerates the new PCIe endpoint:
   • Reads device configuration space
   • Assigns bus/device/function (BDF) identifier
   • Allocates BAR (Base Address Register) resources for DMA rings and control registers
3. Driver initialization: Fi-Wi driver:
   • Initializes DMA rings for this RRH
   • Establishes time synchronization (PTP)
   • Reads RRH capabilities (bands, spatial streams, MAC features)
   • Configures initial RF parameters (channel, power, SSID)
4. RF group integration: Concentrator control plane:
   • Determines which airtime domain this RRH should join (Section 7)
   • Updates queue group assignments
   • Begins CSI collection and spatial stream analysis
   • Steers traffic to/from the new RRH as appropriate

Time from physical insertion to active traffic forwarding: typically 1–5 seconds, depending on link training, driver initialization, and RF group discovery.

13.5.3 RRH removal flow

When an RRH is removed (planned maintenance, failure, or surprise disconnection):

1. Detection: PCIe hot-plug event or surprise removal detected:
   • Link training failure indicates physical disconnection
   • Transaction timeouts indicate unresponsive endpoint
   • Software-initiated removal for planned maintenance
2. Traffic rerouting: Concentrator immediately:
   • Stops scheduling TXOPs to the removed RRH
   • Redirects downlink traffic to other RRHs in the same airtime domain (Section 9)
   • Continues receiving uplink traffic from remaining RRHs in the domain
3. Queue cleanup: Driver:
   • Flushes any outstanding DMA transactions
   • Returns in-flight packets to group queues for retransmission via other RRHs
   • Releases DMA ring memory and BAR resources
4. RF group adjustment: Control plane:
   • Removes RRH from airtime domain membership
   • Recomputes spatial stream structure without this RRH
   • Adjusts beacon power and grouping for remaining RRHs if needed
   • Updates capacity estimates for the affected domain

Impact on active connections: minimal to none for STAs served by multi-RRH domains. Traffic seamlessly fails over to remaining RRHs within the same RF group. For isolated single-RRH cells, removal causes brief disconnection until STAs reassociate with neighboring cells.

13.5.4 Operational advantages

Hot-plug capability provides critical operational benefits:

• Zero-downtime maintenance: Individual RRHs can be serviced, upgraded, or replaced without affecting other RRHs or requiring a maintenance window.
• Gradual deployment: Buildings can be brought online room-by-room as RRHs are installed, rather than requiring all hardware to be in place before service begins.
• Failure resilience: Hardware failures are automatically detected and isolated. The concentrator continues serving the building with reduced capacity rather than failing entirely.
• Capacity scaling: New RRHs can be added to increase coverage or capacity in high-demand areas without disrupting existing service.
• Troubleshooting: Suspect RRHs can be temporarily removed for testing or swapped to isolate problems without taking the network offline.

13.5.5 Design considerations

To fully support hot-swap in production deployments:

• PCIe root complex requirements: The concentrator must use a PCIe root complex with hot-plug controller support (standard in server-class platforms).
• Driver state management: The Fi-Wi driver must handle surprise removal gracefully:
  • No kernel panics or resource leaks on sudden RRH disconnection
  • Proper cleanup of DMA mappings and IOMMU entries
  • Idempotent initialization for re-inserted RRHs
• Power delivery: RRHs should support Power-over-Ethernet (PoE) or similar remote power so they can be powered on/off without physical access to the device.
• Fiber link quality: Optical transceivers must reliably signal link state changes for accurate hot-plug detection. Poor fiber connections can cause spurious removal/insertion events.
• Testing: Hot-plug paths must be tested under realistic failure scenarios:
  • Power loss during active traffic
  • Fiber disconnection mid-TXOP
  • Rapid insertion/removal cycles
  • Multiple simultaneous RRH failures

13.5.6 Contrast with traditional APs

Traditional distributed APs handle failures differently:

• Independent failure domains: Each AP is autonomous. When one fails, only its coverage area is affected, but STAs must reassociate and rebuild transport-layer connections.
• No graceful failover: Distributed APs cannot seamlessly redirect traffic to neighbors because they lack shared state and centralized control. STA reassociation takes hundreds of milliseconds to seconds and disrupts active connections.
• Controller-managed systems: Even with controllers, AP failures are detected via keepalive timeout (seconds), not microsecond-scale PCIe events. Recovery is slow and disruptive.

Fi-Wi's PCIe hot-plug, combined with multi-RRH airtime domains and centralized queues, enables sub-second failover with minimal packet loss—a qualitative improvement over traditional Wi-Fi high-availability approaches.

13.5.7 Integration with L4S and queue management

Hot-swap events interact cleanly with Fi-Wi's L4S and queueing architecture:

• Group queue preservation: When an RRH is removed, its airtime domain's group queue remains intact. Packets simply drain via remaining RRHs instead of the removed one.
• ECN marking continuity: L4S marking continues uninterrupted because it happens at the group queue, not at individual RRHs. Transport-layer congestion control sees no disruption.
• Capacity adjustment: The concentrator dynamically adjusts capacity estimates for the affected domain, which naturally influences L4S control loop behavior without requiring special handling.
• Timestamp coherence: Even during RRH insertion/removal, timestamp-based sojourn time measurements remain valid because they reference the concentrator's PTP time base, not per-RRH clocks.

This separation—queues and control in the concentrator, timing-critical MAC in hot-swappable RRHs—is precisely what enables graceful hardware lifecycle management while maintaining the control-theoretic cleanliness that L4S requires (Appendix A).

14. Hardware Architecture: The Workstation Concentrator vs. The Legacy AP

To understand why Fi-Wi achieves deterministic latency where traditional Wi-Fi fails, we must look beyond the protocol and into the physical architecture of the devices. The feasibility of the "Cut-Through" RRH design relies on the upstream link being non-blocking. Fi-Wi achieves this by replacing the internal switching fabric of legacy APs with the massive PCIe lane overprovisioning of a workstation-class Concentrator.

14.1 The Legacy Bottleneck: Anatomy of a Traditional AP

A traditional Enterprise Access Point is functionally a "Router-on-a-Stick." It forces high-speed wireless traffic through a series of internal serialization bottlenecks before the software ever sees the packet.

Architectural Flaws in Legacy APs:

1. The GMII Choke: The interface between the internal Switch and the CPU is a serialized bottleneck (typically GMII/SGMII). High-speed bursts from Wi-Fi 6E/7 radios can saturate this single link, causing invisible backpressure inside the SoC.
2. Triple Buffering: A packet is buffered at the Switch (Point A), then in system RAM (Point B), and finally in the Hardware DMA Ring (Point C). This "Store-and-Forward" chain destroys the precise timing required for L4S.
3. Opaque Switching: The internal switch operates autonomously. The CPU has no visibility into the depth of the switch's internal buffers, meaning latency accumulates invisibly before the OS can measure it.

14.2 The Fi-Wi Solution: The 92-Lane Fabric

Fi-Wi eliminates the internal switch, the GMII link, and the autonomous CPU. By utilizing high-end workstation silicon (e.g., AMD Threadripper Pro or Intel Xeon W-3400 series), the Concentrator provides 92 to 128 native PCIe lanes directly from a CPU with 24 to 96 high-performance cores.

The 92+ lanes of PCIe eliminate the need for an internal ethernet switch anywhere in the datapath.

14.3 Dedicated Resources and Determinism

By mapping each RRH (or small groups of RRHs) to dedicated root ports on the CPU, Fi-Wi achieves a Non-Blocking Architecture:

• Zero Switching Contention: A read request from RRH #1 does not compete with RRH #2 for switch buffer credits. They travel on separate physical copper traces all the way to the CPU's I/O die.
• Massive Bandwidth Headroom: 92 lanes of PCIe Gen4/5 offer >150 GB/s of aggregate bandwidth. Even if all 50 RRHs retry simultaneously (a "Synchronized Collapse" event), the aggregate demand (~10 GB/s) consumes less than 7% of the host's I/O capacity.

This guarantees that the host DRAM behaves like Deterministic Ultra-Low Latency Memory rather than a shared network resource. This stability is the physical foundation that allows the software-defined queues (Section 15) to operate with microsecond precision.

15. Hardware Queues and the Software Advantage

15.1 The Hardware Queue Problem

Traditional Wi-Fi APs use hardware DMA (Direct Memory Access) rings to meet strict 802.11 MAC timing requirements—SIFS and DIFS deadlines measured in microseconds. While this solves the timing problem, it creates a cascade of architectural constraints that Fi-Wi explicitly avoids.

Hardware queues are expensive to implement in silicon. Each queue requires dedicated SRAM for descriptor storage, control logic for pointer management and overflow handling, and power even when idle. Current chip design limits traditional APs to hardware queues at L2 or MAC—typically the four WMM access categories (AC_VO, AC_VI, AC_BE, AC_BK) per radio * N stations.

While sufficient for basic priority handling, this fundamental constraint prevents the sophisticated per-flow scheduling that modern high-density networks require:

15.2 The DMA Ownership Constraint

An equally significant problem is that once packets are enqueued to hardware DMA rings, the CPU cannot access them without causing race conditions. This "ownership transfer" creates fundamental limitations:

This prevents:

• Real-time ECN marking: Must mark speculatively at enqueue time based on guessed future queue depth, not actual sojourn time at transmission. This leads to oscillation and poor L4S performance.
• Adaptive packet inspection: Cannot implement policies like "inspect deeply when queue exceeds threshold" because packets in queue are inaccessible. All inspection must happen in the fast path, adding latency or requiring expensive hardware DPI engines.
• Dynamic flow reclassification: If a flow's characteristics change (HTTP connection becomes video stream), packets already in hardware queues cannot be moved. Only future packets can be reclassified.
• Queue debugging: Network operators cannot see which flows are filling queues, packet ages, or causes of head-of-line blocking. Only aggregate counters are visible.

15.3 Compensating Hardware

Because hardware queues are limited and packets become inaccessible after DMA, traditional AP vendors must add compensating hardware functionality to address these fundamental architectural limitations:

This compensating hardware represents substantial additional silicon area, design complexity, and verification effort. More critically, hardware-based solutions are fundamentally limited to fixed thresholds and simple policies that were designed into the chip. They cannot implement sophisticated algorithms like CoDel, PIE, or adaptive per-flow policies that require complex state and frequent updates.

15.4 Fi-Wi's Architectural Solution

Fi-Wi escapes these constraints through architectural separation:

RRH: Timing without queuing

RRH silicon implements only timing-critical functions (MAC/PHY, synchronization) with zero hardware queues. Packets arrive from the concentrator milliseconds before transmission, stay in simple descriptor rings briefly, then transmit. No autonomous queuing or scheduling logic.

Concentrator: Unlimited software queues

All queues live in concentrator DRAM. Because the concentrator operates at TXOP granularity (~600 µs) rather than SIFS granularity (16 µs), it has time for software scheduling. Queue structures are simple data structures in memory— vastly cheaper than dedicated silicon:

Packet ownership until last moment

The critical difference: packets remain in concentrator DRAM (software-accessible) until milliseconds before transmission. The scheduler can:

• Compute sojourn time at TXOP instant: now() - pkt->enqueue_time
• Mark ECN based on actual delay, not speculation
• Inspect packet contents and apply adaptive policies
• Reclassify flows by moving packets between queues
• Export queue state to monitoring tools

RRH only owns packets for ~1 ms while transmitting a TXOP—too brief to constrain the system.

15.5 Economic and Strategic Impact

Beyond the direct silicon cost advantages, Fi-Wi gains strategic advantages that compound over time:

• Continuous improvement: Deploy new AQM algorithms, better classifiers, or ML-based scheduling via software updates—no hardware changes required.
• Operational excellence: Full queue visibility enables root-cause analysis, capacity planning, and identification of misbehaving clients—impossible with opaque hardware queues.
• Research integration: As networking research advances, Fi-Wi deployments benefit immediately through software deployment, not years-long silicon redesign cycles.

15.6 Architectural Principle

Fi-Wi's approach follows a clear design principle:

This separation is not arbitrary. It's driven by fundamental constraints: hardware is expensive, inflexible, and opaque; software is cheap, updatable, and inspectable. By placing intelligence in software and only timing-critical functions in hardware, Fi-Wi achieves both the performance of hardware-accelerated systems and the flexibility of software-defined networking—advantages that traditional distributed-AP architectures cannot replicate due to their need for autonomous per-AP decision-making at microsecond timescales.

16. Adaptive Control via Machine Learning

The Fi-Wi architecture's centralized observability enables machine learning to optimize MCS transition dynamics on a per-site basis. Unlike autonomous APs that operate on partial, local state, the Concentrator observes the complete state-transition graph for all RRHs under a single clock. This section describes how Fi-Wi combines physics-based models with adaptive learning to optimize performance.

16.1 The MCS State Graph as a Probability Current Network

The MCS state graph from §3.13 can be formalized as a probability current network, where each node represents a PHY configuration state (MCS index, spatial stream count) and edges represent transitions between states. The system's behavior follows probability flow dynamics:

16.2 What Gets Learned: The Transition Rate Matrix

Machine learning in Fi-Wi optimizes the transition rate matrix W based on telemetry that is only observable in a centralized architecture. For each potential transition from state i (MCS_i, SS_i) to state j (MCS_j, SS_j), the learned rate depends on:

The learned transition rate function takes the form:

This learned function answers: "Given the current state and observed conditions, what is the optimal next MCS/SS configuration to meet the L4S latency target while maximizing achievable throughput?"

16.3 Physics-Informed Learning

Fi-Wi uses physics-informed machine learning that combines Shannon capacity theory with learned corrections. This hybrid approach provides explainability, sample efficiency, and principled generalization.

The transition rate decomposes into two components:

W_physics: The physics baseline uses Shannon capacity to establish theoretical bounds. For each MCS index, the required SNR is known from 802.11 specifications (e.g., MCS 11 requires ~30 dB). The base transition rate is the probability that current SNR exceeds the threshold given measured CSI.

W_learned: The learned correction factor captures deviations from ideal conditions on a per-station basis, as different spatial stream capabilities and local RF environments require station-specific adaptation:

• Building attenuation: Concrete walls vs drywall, multipath reflections
• Interference patterns: Neighboring networks, non-Wi-Fi devices (Bluetooth, microwave)
• Temporal dynamics: Rush hour density vs overnight, seasonal changes
• Client behavior: Application mix (VoIP vs bulk), protocol compliance variations

This approach uses residual learning: the physics model W_physics provides the coarse steering (the "prior"), while the ML model learns the residual error Δ specific to the site. This guarantees the system never performs worse than a standard physics-based model, even before site-specific training converges. The ML correction is additive (or multiplicative) to a known-good baseline.

This decomposition provides three advantages:

1. Explainability: When W_learned deviates significantly from 1.0, the system can flag anomalies and explain why performance differs from theory.
2. Sample Efficiency: The physics prior means the ML model only needs to learn corrections rather than the full mapping from scratch.
3. Generalization: The base model W_physics is universal. Site-specific W_learned factors can be initialized from similar deployments and fine-tuned with site-specific data.

16.4 Training Data from Centralized Observability

The Concentrator's complete state visibility provides labeled training examples that are impossible to obtain in distributed AP systems. Each scheduling decision creates a training tuple:

Over time, the Concentrator accumulates thousands of these labeled examples across varying conditions. The ML model learns patterns such as:

• "When queue sojourn exceeds 700 µs AND client density > 10, stepping down from MCS 9→7 reduces latency by 40%"
• "In specific rooms at certain times, interference spikes require proactive MCS reduction to maintain PER < 5%"
• "4-stream Mu-MIMO (SS=4) only succeeds when κ < 10 dB; above this threshold, SS=2 is optimal"

This supervised learning is only possible with centralized observability. As detailed in Appendix H, autonomous APs lack:

• Global CSI: Each AP sees only its own channel, not cross-RRH interference
• Cross-AP operational state: Cannot observe other APs' band selection, power levels, or scheduling decisions, leading to conflicting local optimizations
• Queue visibility: Queue depth and sojourn time are hidden in firmware
• Deterministic replay: Cannot reproduce exact conditions to verify decisions
• Causal observability: High inference distance between PHY decisions and transport-layer effects

It's worth noting that supervised learning doesn't require perfect ground truth labels to be effective—even relative quality assessments ("better" vs "worse") can drive learning. However, Fi-Wi's complete observability provides significantly richer training signals: precise measurements of queue impact, throughput changes, and latency effects that enable more efficient learning compared to the partial observability available to autonomous systems.

16.5 Transfer Learning Across Sites

Fi-Wi's ML strategy uses transfer learning to balance generalization across sites with site-specific optimization:

Base Model (Cross-Site Training):

A foundational model is trained across multiple deployment sites to learn universal patterns:

Site-Specific Adaptation:

When deployed to a new site, the base model is augmented with learned corrections:

Continuous Adaptation:

The system continues to adapt using online learning with safety constraints:

• Conservative exploration: Only explore new MCS transitions when queue depth is low
• Immediate rollback: If a transition causes PER spike or latency increase, revert and mark as high-risk
• Periodic retraining: Update W_learned based on accumulated telemetry
• Anomaly detection: Alert when learned corrections deviate significantly from baseline

16.6 The Learning Feedback Loop

Fi-Wi's ML capability creates a feedback loop that improves system performance over time:

This loop is unique to centralized architectures. Autonomous APs cannot generate ground truth labels without queue observability. Coordinated AP systems (where APs share summaries via a controller) see effects (latency, ECN) but not causes (queue growth, retry timing, aggregation depth) due to high inference distance.

Fi-Wi's centralized state graph provides the causal observability that machine learning requires. The probability current framework gives this learning a rigorous mathematical foundation: we are learning the transition rate matrix of a physical system governed by conservation laws.

16.7 The Multi-RRH Advantage: Learning the Spatial Network

The presence of multiple concurrent Radio Heads (RRHs) serves as the primary multiplier for the Fi-Wi machine learning capability. It transforms the learning problem from optimizing a single isolated link into optimizing a spatially coupled network. While a traditional AP optimizes a local objective function (its own throughput), the Fi-Wi Concentrator utilizes concurrent RRHs to construct a global view of the RF environment.

This multi-RRH architecture impacts the learning model in three critical ways:

1. Global RF State Visibility ("The Super-Eye")

In traditional systems, an AP is blind to the interference seen by its neighbors. In Fi-Wi, the Concentrator aggregates real-time telemetry from all RRHs simultaneously.

This creates a Global RF State Matrix composed of:

• Cross-RRH RSSI/SNR: Seeing how a client's signal is received across multiple rooms.
• Beamforming Feedback: Aggregating the standard 802.11ax/be compressed V-matrices (CSI) from sounding frames.
• Interference Maps: Correlating noise floor spikes across the floorplan.

This state matrix is sparse, time-aliased, and derived from standards-compliant telemetry rather than continuous per-packet baseband capture.

The model learns not just that "Client A has a weak signal," but specifically that "Client A is weak on RRH 1, strong on RRH 2, and creates -80 dBm interference on RRH 3." This global observability enables the prediction of building-wide interference patterns invisible to single-cell learners.

2. Expanded Action Space (Selection & Redundancy)

Because Fi-Wi treats multiple RRHs as an active redundant set, the ML engine has a broader action space than a standard rate-control algorithm. It learns not only how to transmit (MCS and scheduling decisions) but which RRHs are eligible transmitters for a given packet.

• Downlink Path Selection: The model predicts which RRH is most likely to deliver a packet successfully to a specific client location at a specific time of day, minimizing retries.
• Uplink Diversity Optimization: The system learns which set of RRHs provides the best Selection Diversity for a given client location, optimizing the receive path to maximize SNR.

3. Phase 2 Capability: Eigenstructure & Rank Expansion

Note: This capability requires the hardware-synchronized FPGA architecture (Phase 2).

With sub-nanosecond synchronization, the ML engine will be able to resolve the true distributed Eigenstructure of the environment—the "shape" of available RF paths across distributed radios. This allows for Rank Expansion, where the system resolves more spatial streams (Eigenvectors) than a single physical AP could support, scaling capacity approximately with the number of RRHs, subject to channel rank and geometry.

16.8 Operational Calibration: Zero-Occupancy Sounding

To ensure the physics-informed model converges accurately, Fi-Wi employs a specific operational strategy: Zero-Occupancy Sounding.

As described in Section 16.5, the site-specific transfer function is composed of static building characteristics (H_static) and dynamic temporal variations (Δ_temporal). To disentangle these variables, the system schedules automated channel sounding during hours of minimum occupancy.

• Schedule: Configurable based on vertical (e.g., 03:00–04:00 for hospitality, weekends for office).
• Frequency: Typically runs nightly, or triggers automatically if the system detects significant performance drift (e.g., after a renovation).

This establishes a stable baseline "Zero State" for the learning model, ensuring that subsequent online learning is optimizing for dynamic changes rather than relearning the static environment. This separation dramatically improves offline RL dataset conditioning by preventing the model from relearning static structure while adapting to temporal dynamics.

16.9 Bounded Model Validation During Idle Periods

While the primary learning mode is offline (using historical data), the centralized Concentrator architecture enables a hybrid approach: opportunistic, bounded model validation during predicted idle periods.

Idle Period Detection

Because the Concentrator has global visibility of queue states across all RRHs in an Airtime Domain, it can predict when the RF channel will be underutilized—a capability fundamentally unavailable to autonomous APs that see only their local queues.

• Queue Monitoring: When all per-client queues fall below threshold and no new flows are active
• Traffic Pattern Learning: Historical data reveals predictable idle periods (e.g., 2-4 AM, lunch hours in office buildings)
• Duration Estimation: The system predicts how long the idle period will last based on queue derivative trends

Safe Validation Protocol

During high-confidence idle predictions, the system can perform controlled validation and calibration—not arbitrary exploration:

• Model Validation: Test the next-best admissible action within the model's confidence bounds to validate prediction accuracy
• MCS Probing: Send synthetic frames at alternative MCS values to clients with stale channel estimates
• Calibration Updates: Perform extended channel sounding to refresh H_static estimates

These activities refine the offline model without introducing risk to production traffic.

Production Traffic Protection

Validation is strictly bounded to prevent interference with real traffic:

• Immediate Abort: If real packets arrive during validation, the system immediately reverts to production mode
• Time Budget: Validation limited to maximum 5% of channel time per domain
• Confidence Gating: Only validate when >90% confident of idle duration (based on queue derivative trends)
• Domain Isolation: Validation in one Airtime Domain does not affect others
• Flow Exclusion: No validation is performed on flows marked latency-critical or L4S-sensitive

This hybrid approach provides the safety of offline learning with the adaptability of continuous refinement, exploiting natural traffic lulls that autonomous APs cannot collectively identify.

16.10 Architectural Comparison: Why Autonomous APs Cannot Learn

Machine learning for MCS optimization is fundamentally enabled by Fi-Wi's centralized architecture and impossible in distributed AP systems:

This observability gap is not a vendor implementation issue—it is an architectural limitation. Autonomous APs cannot generate high-quality training labels without queue observability.

17. Concentrator Fast Path: DPDK, DMA, and Queue Determinism

The preceding sections established the architecture of the Fi-Wi concentrator: why kernel bypass is required (§3.5), centralized packet memory (Section 5.4), group queues as the sole AQM bottleneck (Section 5.3), microsecond timestamps written into the Fi-Wi shim header (Section 5.2), and ML-driven MCS selection running continuously against that centralized data (Section 16). This section explains how the concentrator executes that pipeline with the determinism the architecture requires — maintaining a single observable bottleneck per airtime domain, applying ECN marks at the right moment, and keeping the RRH free of scheduling logic.

17.1 Why a Kernel-Bypass Data Plane

The Fi-Wi concentrator's latency and determinism targets strongly favor a kernel-bypass data plane. A conventional interrupt-driven kernel path would reintroduce jitter at exactly the point where the architecture is trying to remove it.

L4S requires ECN marks to be applied at the group queue on the same time scale as a single 802.11 TXOP. The Linux kernel's softirq-based packet path introduces interrupt coalescing and scheduler contention that accumulates across bursts. More fundamentally: every packet that transits the kernel stack competes with arbitrary OS activity for CPU time. The queue depth is not directly visible to userspace without a syscall; the marking decision cannot be co-located with the queue measurement in the same cache line.

Fi-Wi's concentrator data plane therefore runs via DPDK (Data Plane Development Kit): tight busy-poll loops on dedicated cores, with no interrupt-driven jitter. All packet operations — receive, classify, AQM mark, forward — execute in a cache-resident loop that preserves the single-bottleneck, fully-observable queue structure that the rest of the architecture depends on.

17.2 The Memory Model: IOMMU, VFIO, and Hugepages

DPDK allocates all packet buffers (mbufs) from hugepages, eliminating TLB misses during packet processing. Each airtime domain's group queue is a logically contiguous region within this space. The pool is allocated once at startup; no per-packet memory allocation occurs on the fast path.

Each SFP+ NIC is bound to the vfio-pci driver. The system IOMMU enforces DMA isolation: a card can only reach the memory regions explicitly registered with it at startup. This gives the concentrator two properties simultaneously:

• Safety: A malfunctioning RRH cannot issue PCIe DMA transactions to arbitrary concentrator memory. The IOMMU enforces the boundary in hardware.
• Performance: IOMMU mappings are established once when the mbuf pool is created. From that point, rx_burst and tx_burst are zero-copy: the NIC DMA engine writes received frames directly into pre-registered mbuf space and reads transmit frames from the same space, with no per-packet kernel involvement.

17.3 Airtime Domains as Hardware Queue Partitions

DPDK exposes each NIC's hardware receive queues independently. Fi-Wi uses this to achieve a direct, lockless mapping from PCIe port and queue index to airtime domain — the same logical grouping described in Section 7. Each lcore owns a fixed set of (port, queue) pairs. Because ownership is exclusive, there are no locks on the fast path and no shared state between lcores during steady-state forwarding.

17.4 The L4S Marking Loop

The AQM marking step is deliberately minimal. The DPDK data plane does not run a full queue scheduler — that is the outer control loop's responsibility (Section 6). The inner loop does one thing: read sojourn time from the shim header (Section 5.2) and set the ECN CE codepoint if the threshold is exceeded.

// Per-packet in the rx → tx burst loop:
uint64_t sojourn_ns = now_tsc() - pkt->t_ingress;
if (sojourn_ns > THRESHOLD_NS) {
    rte_ipv4_l4s_mark(pkt);                       // in-place, no copy
    fiwi_meta(pkt)->ecn_flags |= ECN_CE_APPLIED;
}
rte_eth_tx_burst(out_port, queue_id, &pkt, 1);

Because t_ingress is written by the same lcore at enqueue, no cross-core communication is needed to compute sojourn time at dequeue. The marking decision is local to the polling thread. This is what Section 5.3 means when it says AQM runs "exactly where the integrator lives": the integrator is the group queue, the group queue is an mbuf ring in hugepage memory, and the marking loop touches that ring on every poll cadence with no additional indirection.

17.5 Fault Isolation via IOMMU Groups

In a multi-card concentrator, each SFP+ card appears in its own IOMMU group, which means each card can be bound to VFIO independently and the IOMMU enforces that one card's DMA cannot reach another card's memory regions. In a deployment with multiple SFP+ cards, the IOMMU topology provides natural fault isolation at the card boundary: a PCIe error or runaway DMA event from one RRH is contained within its card's group and cannot corrupt the packet memory of an adjacent airtime domain. This is a hardware guarantee, not a software policy.

17.6 What DPDK Does and Does Not Solve

The kernel-bypass data plane is not a complexity cost — it is the mechanism that justifies the RRH's simplicity. Because the concentrator runs a deterministic, observable pipeline that applies AQM, tracks sojourn time, and manages all descriptor posting without OS intervention, the RRH never needs to make a queuing or scheduling decision. It remains a pure DMA client, exactly as the silicon cost argument in Section 5.4 requires.

Incumbent distributed APs have no equivalent. Because each AP operates autonomously, it must run its own Linux network stack, its own qdisc, and its own firmware scheduler. The CPU carrying that stack is the dominant gate cost per RRH (Section 5.4, silicon cost table). A centralized DPDK pipeline eliminates that requirement across every RRH simultaneously — not by optimizing the AP implementation, but by removing the architectural condition that forces the CPU to exist there in the first place.

That said, DPDK solves a specific problem: it gives the concentrator a deterministic, observable, zero-copy execution path in which queue state, ECN marking, and packet steering remain under unified software control. It does not solve the radio-side interface. Per-packet MCS selection, EDCA parameter control, and TX-outcome metadata from the Wi-Fi silicon remain the next required interface boundary — the point at which concentrator intelligence must reach into the RRH to close the control loop. DPDK is the precondition; radio-side per-packet programmability is what completes it.

17.7 DualPI2 Baseline: Control Law and Queue Structure

Section 17.4 described the minimal ECN marking step — reading queue state and applying a CE mark in the fast path. That sketch is sufficient to illustrate where marking occurs, but it elides the control structure that makes L4S coexistence with legacy traffic work: the dual-queue coupled AQM defined in RFC 9332.

This section defines the baseline DualPI2 control law as it would be realized inside the DPDK polling loop. Fi-Wi preserves this dual-queue topology, coupling mechanism, and PI-based control structure, but Section 17 replaces the underlying congestion signal with Airtime Debt (D_i), grounding the controller in predicted wireless service time rather than raw queue occupancy.

17.7.1 The Two Queues

Each airtime domain maintains two logically independent mbuf rings in the concentrator's hugepage pool: an L4S queue for scalable congestion-control flows (senders marking with ECT(1)), and a Classic queue for legacy RFC 3168 flows and unmarked traffic. Classification happens at ingress on the fast path, before the packet is enqueued, and costs a single bitfield check on the IP ECN field:

// Ingress classification — per-packet, inline in the rx burst loop
uint8_t ecn = (pkt_ip->type_of_service & 0x03);
bool is_l4s = (ecn == 0x01 || ecn == 0x03);   // ECT(1) or CE — scalable sender

fiwi_meta(pkt)->queue_class = is_l4s ? QUEUE_L4S : QUEUE_CLASSIC;
enqueue_to_domain(pkt, domain_id, fiwi_meta(pkt)->queue_class);

Both queues drain toward the same transmit burst for that airtime domain. The scheduler services the L4S queue with a strict low-latency budget and the Classic queue at a rate that saturates the domain's aggregate share, matching the DualPI2 service model from RFC 9332.

17.7.2 The Coupling Mechanism

The key property of DualPI2 is that the two queues are not independent. The Classic queue's drop probability pc — computed by a PI controller from a congestion signal representing pressure at the shared bottleneck — also governs the L4S queue's ECN marking probability via a coupling factor k (default 2 in the Linux sch_dualpi2 reference implementation).

// Outer control loop — runs on a slow timer cadence (~16 ms), same lcore,
// non-preemptive. Not per-packet.
double signal_classic = ewma_update(&domain->classic_signal,
                                    ring_depth(QUEUE_CLASSIC));
double p_c = max(0.0, K_PI * (signal_classic - TARGET_CLASSIC));  // PI controller

double p_l = COUPLING_K * p_c;   // Coupled L4S marking probability

// Applied per-packet in the L4S dequeue path:
double p_l_step = (sojourn_L4S_ns > THRESHOLD_L4S_NS) ? 1.0 : p_l;
if (rte_rand_u64() < (uint64_t)(p_l_step * (double)UINT64_MAX))
    rte_ipv4_l4s_mark(pkt);      // Set ECN CE in-place, no copy

In a conventional queue-based implementation, signal_classic would be an EWMA of Classic queue depth. In Fi-Wi, that queue-derived signal is replaced as the PI controller input by Airtime Debt (D_i), a forward estimate of wireless service time. The DualPI2 control law, coupling mechanism, and dual-queue topology remain unchanged; only the input signal changes.

Queue depth is a lagging indicator in Wi-Fi because contention, retries, and variable PHY rates consume airtime without necessarily appearing in buffer occupancy. Airtime Debt provides a forward-looking signal that better matches the true wireless bottleneck while preserving the DualPI2 coexistence structure required for L4S and Classic traffic to share the medium.

17.7.3 Per-Domain State and the fiwi_update Interface

Each airtime domain carries its own DualPI2 state alongside the fiwi_rrh_state struct (Section 17.5). Because each lcore owns a fixed set of domains exclusively (Section 16.8), this state is never shared across cores — no locks, no atomics, no cache-line bouncing on the fast path.

The telemetry path (Section 17.8) delivers ground-truth airtime measurements back to the lcore via a lockless ring carrying fiwi_update objects. The struct is defined here because it originates in the DPDK fast-path layer and is consumed by it; Section 17.8 populates it from Netlink/vendor telemetry events:

/**
 * fiwi_update — telemetry record posted by the Netlink callback,
 * consumed by the DPDK lcore during its scheduling loop.
 * Allocated from fiwi_update_pool (rte_mempool); returned after use.
 */
struct fiwi_update {
    uint8_t  type;          /* AIRTIME_RECONCILE (only type currently defined) */
    uint32_t rrh_id;        /* RRH index, validated < FIWI_MAX_RRHS before enqueue */
    uint64_t actual_us;     /* Hardware-path-to-status interval (ground truth) */
    uint64_t expected_us;   /* Forward estimate: T_phy + T_agg at enqueue time */
    uint32_t retry_us;      /* Observed retry airtime from telemetry metadata */
};

17.8 Multi-RRH lcore Topology and Control Ownership

The Umber concentrator runs on a workstation-class host with a Threadripper PRO processor and multiple PCIe-connected RRHs. This section describes how DPDK lcore assignments map onto that hardware topology to preserve cache locality, single-writer semantics, and deterministic fast-path execution.

Each lcore owns both the DualPI2 control state (Section 17.7) and the Airtime Debt estimator (Section 18) for its assigned RRHs. This ensures that congestion estimation, scheduling, and ECN marking operate within a single execution context.

17.8.1 RRH Assignment

17.8.2 Control and Data Flow

Each RRH lcore applies its per-domain DualPI2 loop as described in Section 17.7, with Airtime Debt (D_i) serving as the PI controller input in place of queue depth. This presents a single, airtime-grounded congestion signal per domain to the L4S control loop.

Downlink traffic is classified at ingress and directed to the appropriate airtime domain. The owning lcore performs scheduling, ECN marking, and transmission. Uplink traffic follows the reverse path toward the WAN interface.

Because each lcore exclusively owns its RRHs and associated Airtime Debt state, congestion estimation, scheduling, and ECN marking operate without cross-core coordination. This preserves deterministic fast-path behavior.

18. Airtime-Assisted ECN: Airtime Debt as the Congestion Signal

Fi-Wi does not infer congestion from queue depth alone. The bottleneck is the wireless medium, and the relevant state variable is the time required to successfully transmit packets over that medium. The system replaces the queue sojourn-time inputs of traditional PI² controllers with Airtime Debt (D_i), converting a stochastic medium into a controlled service process.

18.1 The Bottleneck is Airtime, Not a Queue

In traditional L4S systems, ECN marking is derived from queue sojourn time, which assumes a stationary service rate. These assumptions fail in Wi-Fi because service time varies per client based on PHY rates, contention, and retries. Fi-Wi replaces backward-looking buffer metrics with a forward model of wireless service time. The Concentrator maintains this model continuously and makes scheduling decisions on predicted service outcomes, not observed queue growth. This approach provides the AQM with a signal that has a more stationary distribution than raw queue depth over a variable-rate medium, improving marking coherence and L4S stability.

18.2 Airtime Debt Model (Per RRH)

For each RRH (i), the Concentrator maintains a real-time Airtime Debt (Di):

D_i = A_i + C_i + R_i

• A_i (Backlog Airtime): Predicted transmit duration for staged and in-flight packets.
• C_i (Contention Penalty): Predicted medium access delay (µs), informed by driver telemetry and controlled via EDCA scheduling (Section 5.1.4).
• R_i (Retry Penalty): Predicted retransmission overhead (µs) derived from recent PER history (Section 16.4).

18.3 Measuring Ground Truth (Hardware-Path-to-Status)

The "Ground Truth" for airtime consumption is measured as the interval from descriptor posting into the hardware transmit path to TX Status (hardware completion signal via driver/vendor-specific telemetry events such as RRH-driver TX status reports). This interval captures the full service duration, including the full wait for TXOP eligibility (AIFS + backoff), aggregation delay, and all hardware-level retransmission attempts.

18.4 Predicted Sojourn Time (S_i)

For any packet, the Predicted Sojourn Time (S_i) is a forward estimate of delivery time:

S_i(packet) = D_i + T_service(packet)

The Tservice calculation is decomposed into: Tagg (aggregation hold time) + Tphy (modulation time at current MCS) + Tretry (statistical retry overhead). This estimate is packet- and client-specific; it is not a constant service quantum.

18.5 Implementation: DPDK Fast Path State

The Concentrator tracks RRH state in hugepage-backed memory. The DPDK lcore is the sole writer of fiwi_rrh_state; telemetry updates are applied via per-RRH lockless ring buffers to preserve single-writer semantics and microsecond-level determinism.

struct __rte_cache_aligned fiwi_rrh_state {
    uint32_t rrh_id;
    uint64_t D_i;            /* Total airtime debt (A+C+R) */

    /* Component Estimates (microseconds) */
    uint64_t A_i;            /* Total scheduled airtime (queued + in-flight) */
    uint32_t C_i;            /* Estimated contention delay */
    uint32_t R_i;            /* Estimated retry penalty */

    /* Feedback & Synchronization */
    uint64_t last_update_us;     /* Timestamp of last lcore application */
    uint64_t last_tx_status_us;  /* TSC of last hardware completion */
    uint32_t moving_avg_per;     /* Recent PER (Section 16.4) */
};
    

Di is recomputed in the DPDK fast path after each update to Ai, Ci, or Ri. The loop updates Ai when packets are assigned to an RRH and decrements it upon TX completion using telemetry feedback.

18.6 Authoritative Congestion Signaling

Airtime Debt replaces physical queue depth as the authoritative input for the Dual-Queue AQM, providing a single, authoritative congestion signal across all RRHs without relying on a shared physical buffer.

• L4S Marking: Applied if Si > Tlow. This bypasses traditional sojourn measurements to signal congestion at the true wireless bottleneck.
• Classic Drop: Airtime Debt (Di) replaces queue depth as the input to the PI controller defined in Section 17.7. This preserves the Dual-Queue AQM structure while grounding the control signal in predicted wireless service time rather than buffer occupancy.

18.7 Slow-Path Observability

While Di provides fast-path control, the system monitors Airtime Utilization (U_air = ΔTX_DURATION / Δt) as a slow-path observability metric. This metric is used to identify external interference patterns and long-term capacity shifts in the airtime domain, calibrating the confidence weights applied to the Ci and Ri estimators.

18.8 Telemetry Feedback: Netlink Calibration

The following logic processes TX_STATUS events from the RRH driver. Completion data is retrieved from a pre-allocated mempool and posted to a per-RRH lockless ring to reconcile state without lcore contention.

/* Telemetry Path (Netlink Callback) */
static int fiwi_handle_mt76_telemetry(struct nl_msg *msg, void *arg) {
    struct nlattr *attrs[MT76_ATTR_MAX + 1];
    nla_parse(attrs, MT76_ATTR_MAX, genlmsg_attrdata(nlmsg_data(nlmsg_hdr(msg)), 0),
              genlmsg_attrlen(nlmsg_data(nlmsg_hdr(msg)), 0), NULL);

    if (!attrs[MT76_ATTR_TX_DURATION] || !attrs[MT76_ATTR_RRH_ID])
        return NL_SKIP;

    uint32_t rrh_id = nla_get_u32(attrs[MT76_ATTR_RRH_ID]);
    if (rrh_id >= FIWI_MAX_RRHS) return NL_SKIP;

    struct fiwi_update *update;
    if (rte_mempool_get(fiwi_update_pool, (void**)&update) < 0) return NL_SKIP;

    update->type = AIRTIME_RECONCILE;
    update->rrh_id = rrh_id;
    update->actual_us = nla_get_u64(attrs[MT76_ATTR_TX_DURATION]);
    update->retry_us = nla_get_u32(attrs[MT76_ATTR_RETRY_DURATION]);
    update->expected_us = estimate_service_time(msg);

    rte_ring_enqueue(rrh_update_rings[rrh_id], update);
    return NL_PROCEED;
}
    

18.8.1 Telemetry Application (DPDK lcore)

The DPDK lcore closes the control loop by draining the update ring. It decrements the backlog and calibrates penalties to ensure the Airtime Debt remains an accurate representation of physical medium pressure.

/* DPDK lcore: apply telemetry updates */
static inline void
fiwi_apply_updates(struct fiwi_rrh_state *rrh, struct rte_ring *ring)
{
    struct fiwi_update *upd;
    while (rte_ring_dequeue(ring, (void**)&upd) == 0) {
        /* 1. Discharge processed backlog */
        rrh->A_i = (rrh->A_i > upd->actual_us) ? (rrh->A_i - upd->actual_us) : 0;

        /* 2. Update contention estimate (drift from expected modulation time) */
        uint32_t drift = (upd->actual_us > (upd->expected_us + upd->retry_us)) ?
                         (upd->actual_us - upd->expected_us - upd->retry_us) : 0;
        rrh->C_i = (rrh->C_i * 7 + drift) >> 3;

        /* 3. Update retry penalty */
        rrh->R_i = (rrh->R_i * 7 + upd->retry_us) >> 3;

        /* 4. Recompute total Airtime Debt (D_i) */
        rrh->D_i = rrh->A_i + rrh->C_i + rrh->R_i;

        rrh->last_tx_status_us = rte_get_tsc_cycles();
        rte_mempool_put(fiwi_update_pool, upd);
    }
}
    

18.9 Visualization: The Airtime Debt Control Loop

Figure 17-1: The Fi-Wi recursive control loop for stabilizing stochastic wireless service.

Diagram Overview: Closing the Feedback Loop

Figure 17-1 synthesizes the technical components of the Airtime Debt model into a continuous functional loop. The architecture separates the Speculative Forward Path (Fast Path) from the Calibrated Feedback Path (Telemetry Path).

1. Forward Service Model (Prediction): Every ingress packet triggers a per-STA calculation of Tservice. This is not a global constant; it is a client-specific sum of aggregation hold time (Tagg), PHY modulation time (Tphy), and predicted retry overhead (Tretry) based on that STA's specific RF context.

2. Debt Update & Marking Decision: The predicted Tservice is added to the RRH's Ai (Backlog). If the resulting Predicted Sojourn Time (Si) exceeds Tlow, an ECN CE mark is applied immediately in the DPDK fast path. This provides the "Virtual Backpressure" that stabilizes L4S senders.

3. Ground Truth Calibration (Correction): As the packet is dispatched via DMA, the hardware records the precise interval from descriptor posting into the hardware transmit path to TX Status completion. The Telemetry Path calculates the Drift—the delta between the forward prediction and physical reality.

4. Estimator Refinement: This drift is fed back into the EWMA filters for Ci (Contention) and Ri (Retries). This ensures that subsequent predictions for the same STA or RRH domain are corrected for changing medium pressure, effectively regularizing the stochastic nature of the 802.11 medium.

19. Summary

The core idea of Umber’s Fi-Wi architecture is to make a building full of Wi-Fi radios behave like a large number of predictable, low-latency, cellularized bottlenecks (often cell-per-room) that integrate cleanly with L4S, and to avoid Wi-Fi collapse in the regime that matters most for users: tail latency.

We do that by:

• Time-synchronizing RRHs and concentrator to microsecond accuracy
• Centralizing packet memory and queues in the concentrator
• Adding a Fi-Wi shim header with timestamps and queue metadata
• Doing all L4S AQM/ECN marking at group queues, where the wireless delay is actually created
• Using shared state to dynamically group RRHs into airtime domains (cells), considering both interference and spatial-stream / eigenvector structure
• Using multiple RRHs per room and per STA as an active redundant RF set, analogous to MLO but building-wide
• Leveraging CSI and learning to predict capacity, interference, and collapse risk
• Enabling, where appropriate, coordinated multi-RRH operation and dynamic point selection (Section 10)
• Keeping RRHs within a strict power/thermal envelope, while using PCIe Gen3 x1 over fiber to connect them as low-latency DMA clients of central packet memory

Compared to a building filled with independent APs, Fi-Wi provides:

• Lower and more predictable latency, especially in the tails
• More stable and effective L4S behavior
• A better aggregation–latency tradeoff
• Centralized, cellularized control with shared state that scales across many RRHs and rooms
• RF-layer redundancy, richer spatial streams, and fast failover with a clean single-bottleneck view for L4S
• A practical path to cell-per-room density with manageable power, thermals, and hardware complexity

---

Appendix A: 802.11 Backoff Timing & Collapse Dynamics

This appendix explains the precise behavior of the 802.11 CSMA/CA backoff algorithm, why the freeze/resume mechanics create strong nonlinearities under load, and how this drives the collapse behavior discussed in Sections 2 and 6. We also include reference diagrams, accurate pseudocode, and probability scaling that shows why birthday-paradox collisions appear long before PHY saturation.

A.1 Overview

The 802.11 MAC is built around two core mechanisms:

• Carrier Sensing (CSMA/CA) — sense-before-transmit.
• Random Backoff — avoid collisions by randomized slot countdown.

These mechanisms interact in a way that works beautifully for light to moderate station counts, but begins to break down sharply once multiple stations become backlogged. Collapse is not a "bug"; it is the mathematically expected outcome under high concurrency.

A.2 Backoff Decrements Only During Idle SlotTime

When a station has a frame to send, it chooses a random integer:

B ← Uniform[0, CW]

where CW is the contention window. The counter decrements only when:

• The PHY senses the channel idle for the entire SlotTime (9 µs typical).
• NAV = 0 (no virtual carrier-sense in effect).
• The station has already observed one full AIFS interval.

If any of these conditions break during a SlotTime boundary, backoff does not decrement.

Diagram A-A — Backoff Countdown with Idle Slots and Freezes

This "idle-slot-only" decrement rule is the source of nonlinear timing behavior.

A.3 Freeze Conditions: Physical Busy + NAV Busy

The backoff counter freezes immediately under either condition:

• PHY busy: Energy detect (ED) threshold exceeded — another station is transmitting.
• NAV busy: The local station believes a scheduled TXOP is ongoing (Duration field).

NAV counts down in microseconds, not slot units, so a NAV may span dozens or hundreds of SlotTimes, creating long frozen periods.

Diagram A-B — NAV Freezes Backoff for Entire Duration

A.4 Full Backoff State Machine

The following pseudocode describes the real 802.11 backoff and retry machine:

# Variables
B   = random integer in [0, CW]
CW  = CWmin initially, doubled on failures
NAV = virtual carrier sense (µs timer)
Slot = 9 microseconds (typical)
AIFS = access category-specific inter-frame space

while True:

    wait_until( medium_idle() and NAV == 0 )
    wait(AIFS)  # must see idle for entire AIFS

    # Backoff countdown
    while B > 0:

        if medium_idle() and NAV == 0:
            wait(Slot)
            if medium_idle() and NAV == 0:
                B -= 1      # decrement only if entire slot was idle
        else:
            # Freeze B until another idle AIFS appears
            wait_until( medium_idle() and NAV == 0 )
            wait(AIFS)

    # Backoff fully expired, attempt TX
    transmit()

    if ack_received():
        CW = CWmin
        B = random(0, CW)
    else:
        CW = min(2 * CW, CWmax)
        B = random(0, CW)

The critical detail: multiple stations freeze and resume their counters in lock-step after every long TXOP or NAV, making collisions statistically inevitable as station count grows.

A.5 Collision Probability and the Birthday Paradox

Each station independently picks a backoff slot in [0, CW]. The probability that no two stations choose the same slot is:

P(no collision) = (CW+1)! / [(CW+1 - n)! · (CW+1)^n]

where n = number of active contenders. Therefore:

• Even with CW=15 (CWmin for AC_BE), collisions explode once n > 6–8.
• Increasing CW reduces collisions but explodes delay variance.
• Freeze/resume sync makes large groups of stations “compete” on the same few idle slots.

Diagram A-C — Collision Probability vs. Number of Stations

This is the MAC-level reason collapse begins long before PHY capacity is reached.

A.6 Why Collapse Appears as 2–3 ms TXOP Tails

Once collisions become frequent:

• Stations reach retry limits and enlarge CW exponentially.
• Longer CW → longer idle-time waits → bigger aggregates formed.
• Bigger aggregates → longer TXOPs (1.5–3 ms typical in collapse).
• Long TXOPs → other stations see multi-ms starvation → tail latency balloons.

Diagram A-D — TXOP Length as Collapse Indicator

A single 3 ms TXOP already violates the bottleneck-delay budget required by L4S (≈250–300 µs). With multiple stations taking such TXOPs, service gaps can reach 10–50 ms for unlucky flows.

A.7 Multi-Station Synchronization Example

The following diagram illustrates how multiple stations become phase-aligned:

This synchronization is why the birthday paradox applies so strongly in Wi-Fi.

A.8 Why Fi-Wi Breaks the Cycle

Fi-Wi addresses the “every station fends for itself” randomness by:

• Centralizing queueing and airtime allocation.
• Eliminating autonomous EDCA decisions for concentrator-controlled downlink transmissions.
• Coordinating RRH TXOP lengths and schedules.
• Ensuring no two RRHs “race” via random backoff.
• Driving TXOPs toward safe, L4S-friendly durations (≈250 µs).
• Reducing contention pressure for legacy uplink clients through EDCA parameter tuning, coordinated protection windows, and transport-layer pacing; cooperative 802.11ax/be clients can be further shaped through trigger-based mechanisms.

Thus Fi-Wi converts the AP-controlled portion of Wi-Fi from autonomous contention into a centrally scheduled wireless fabric, while bounding and influencing the remaining legacy CSMA/CA behavior.

---

Appendix B: Channel State Information (CSI) and Learning-Enhanced Fi-Wi

This appendix describes how Fi-Wi can use Channel State Information (CSI) from each RRH, together with learning models (e.g. LSTM or TCN), to improve grouping, scheduling, redundancy, and control beyond what is possible with queue-based feedback alone.

B.1 What CSI provides in a Fi-Wi context

Concept: What is CSI?
Imagine shouting in a complex room. You hear echoes bouncing off walls, furniture, and people. If you analyze those echoes, you can map the environment.

In Wi-Fi, Channel State Information (CSI) is that map. It describes exactly how the radio wave traveled from the transmitter to the receiver—including all the bounces (multipath), fading, and phase shifts caused by the physical environment.

• RSSI (Signal Strength): Tells you how loud the signal is. (One number).
• CSI (Channel State): Tells you the shape of the signal distortion. (A matrix of complex numbers).

Traditional APs throw this data away after decoding the packet. Fi-Wi sends it to the Concentrator, allowing the system to "see" the RF environment and mathematically calculate how to steer beams or combine signals.

Wi-Fi Sensing: Because physical objects reflect radio waves, any movement in the room changes the CSI pattern. By monitoring these changes over time, Fi-Wi can detect presence—such as a person walking or a pet breathing—turning the network into a ubiquitous sensor without cameras.

Modern 802.11 chipsets can export CSI per subcarrier or per resource unit: complex-valued estimates of the channel between an RRH and a station (STA). In a Fi-Wi deployment, each RRH periodically reports:

• Per-subcarrier CSI magnitude/phase (or compressed variants)
• Signal-to-noise ratio (SNR), RSSI, and noise floor estimates
• Selected MCS, PHY rate, and per-TXOP success / retry / PER statistics
• Spatial-stream usage and beamforming feedback (if applicable)

Thanks to centralized time synchronization and packet memory, the concentrator can align CSI reports with:

• Specific flows and STAs
• Specific group queues and airtime domains
• Queue states and ECN marking decisions at those times

This gives Fi-Wi a rich per-domain, per-STA time series:

• CSI(t) — the RF channel state (including spatial eigenstructure)
• MAC(t) — TXOP, MCS, retries, aggregation size, spatial-stream usage
• Queue(t) — group queue depth, marking probability p_k
• Outcome(t) — achieved throughput, delay, PER, “collapse events”

B.2 What we want to predict

Using this data, Fi-Wi can learn models to help answer questions such as:

• Short-horizon capacity prediction: Given recent CSI + MAC statistics, what is the effective service rate C_eff for this RRH or airtime group in the next Δt?
• Interference and grouping: Which RRHs are likely to interfere strongly in the next Δt (e.g. due to mobility, reflections, or external interferers)? Which ones share useful eigenmodes?
• Scheduling decisions: For a given flow/STA, which RRH (or band/channel) and which spatial-stream configuration is likely to give the best service (throughput vs. delay) in the next window?
• Collapse risk: Are we heading into a region where adding more load will cause retries, PER, and tail latency to spike?

These predictions can feed directly into:

• Queue grouping and re-grouping (airtime domains)
• Per-group AQM parameter adaptation (e.g., queue target, gains)
• RRH selection, channel selection, and per-STA scheduling
• Active redundancy policies (which RRHs should listen or transmit in parallel)
• Spatial-stream / eigenmode choices within each airtime domain

B.3 Example model: LSTM / TCN

One reasonable approach is to use a sequence model such as an LSTM or Temporal Convolutional Network (TCN) per airtime domain:

Input features (per timestep):
  - queue depth q_k
  - marking probability p_k
  - throughput, PER, retries
  - per-RRH CSI summary (e.g. dominant eigenvalues/eigenvectors)
  - beacon power settings, channel, bandwidth

Outputs:
  - predicted effective capacity C_eff,k+1
  - predicted collapse risk score
  - recommended group reconfiguration / beacon adjustments (optional)

A higher-level policy layer then uses these predictions to:

• Re-balance load across domains
• Merge or split airtime domains
• Adjust beacon powers to reshape cell edges
• Tweak PI2 / L4S parameters per domain

The key point is that Fi-Wi has access to the joint state across all RRHs—queues, CSI, MAC outcomes, and beacon configuration—so learning can be done on a true building-scale view rather than a per-AP snippet.

B.4 The Non-Linear Control Policy (Feature Vectors)

While the PI² controller (Section 6.2) provides a robust baseline using linear control theory, the wireless medium is inherently non-linear. A small drop in SNR can cause a discrete, non-linear step-down in MCS, cutting capacity by half in microseconds. A linear controller often reacts too slowly to these step-changes.

Because the Concentrator terminates both the MAC (Inner Loop) and L4S (Outer Loop), it possesses a complete, global view of the system state. This allows Fi-Wi to implement a Non-Linear Marking Signal derived from a rich real-time feature vector:

Optimization Objective: Efficiency vs. Latency
The system uses this vector to solve the fundamental Wi-Fi trade-off: Aggregation Efficiency vs. Serialized Latency.

• The Physics: PHY efficiency increases with aggregation size (amortizing the preamble/SIFS overhead). However, larger aggregates typically imply deeper queues and higher latency.
• The Logic: By correlating MCS with Arrival Rate, the control policy can dynamically adjust the marking threshold.
  • High MCS Scenario: The policy knows the "pipe" is wide. It relaxes the marking threshold, allowing the "Inner Integrator" (Section 6.3) to accumulate more packets. This naturally builds larger, more efficient aggregates (e.g., 64 MPDUs) without violating the L4S latency budget because they drain instantly.
  • Low MCS Scenario: The policy knows the "pipe" is narrow. It aggressively marks via the ARM Policer, forcing the sender to slow down. This keeps the aggregate size small (e.g., 4 MPDUs), preventing the massive "sawtooth" latency spikes that occur when sending large aggregates at low data rates.

This creates a Non-Linear Marking Signal that optimizes Throughput per Microsecond of Latency, rather than simply targeting a fixed queue depth.

---

Appendix C: Latency Hiding via Scatter-Gather DMA

Early architectural models of C-RAN often assumed a "Store-and-Forward" approach, where full packets must be buffered at the edge to meet timing. Fi-Wi eliminates this inefficiency by leveraging the natural physics of the 802.11 air interface. We utilize a Scatter-Gather DMA engine with Preamble Hiding to enable a "Thin RRH" design with minimal local SRAM.

C.1 The "Preamble Shield" Physics

The critical timing constraint in Wi-Fi is the transition from "Decision to Transmit" to "Energy on Air." However, the 802.11 PHY does not transmit user data immediately. Every transmission begins with a PHY Preamble (PLCP) and MAC Headers.

The Insight: The transmission of the Preamble and Headers takes roughly 20–40 µs (depending on PHY generation). The round-trip time to fetch payload data over 100m of PCIe-over-Fiber is roughly 2–5 µs.

Consequently, the fetch latency is completely "hidden" behind the transmission of the headers. The payload data arrives at the RRH's small FIFO well before the PHY is ready to modulate it.

C.2 Scatter-Gather Architecture

Instead of a large packet buffer, the Fi-Wi RRH implements a Scatter-Gather DMA engine that composes frames on the fly from two distinct memory regions:

1. Template RAM (Local RRH SRAM): Stores 802.11 MAC headers, PLCP headers, and delimiter signatures. This memory is small (< 16 KB), fast, and populated by the Concentrator during the descriptor posting phase.
2. Payload Buffer (Remote Concentrator DRAM): Stores the actual 802.3 Ethernet payloads. These remain in the host server's memory until the exact moment of transmission.

C.3 The Transmit Sequence

1. Descriptor Posting: The Concentrator posts a descriptor to the RRH. This descriptor points to the header in Local RAM and the payload in Remote DRAM.
2. Contention: The RRH MAC performs EDCA backoff. No data is moved during this phase.
3. TX Trigger: When backoff reaches zero, the MAC immediately begins transmitting the Preamble from Local RAM.
4. Just-in-Time Fetch: Simultaneously with the Preamble start, the DMA engine issues a read request to the Concentrator for the payload data.
5. Cut-Through: Data returns from the fiber, flows into a small speed-matching FIFO (e.g., 4 KB), and flows directly into the PHY serialization path immediately following the header.

C.4 Solving the Retry Timing (SIFS)

A common objection to C-RAN is the SIFS deadline (16 µs) required for retries. If a transmission fails, the station must retransmit immediately.

With Scatter-Gather, the RRH does not need to buffer the packet for retries. If a NACK occurs, the MAC simply resets the Scatter-Gather engine. It re-transmits the Preamble (from Local RAM) while re-issuing the DMA fetch (from Remote RAM). Because the fiber latency (5 µs) is significantly shorter than the SIFS + Preamble duration, the data again arrives in time.

C.5 Architectural Benefits

• Zero Copy: Packets are never copied CPU-to-Buffer or Buffer-to-Buffer. They move from Host DRAM to Air.
• Minimal SRAM: The RRH does not need MBs of RAM to buffer aggregates. It only needs KB for headers and FIFOs. This significantly reduces leakage power and die cost.
• Infinite Retries: The system is not limited by local buffer size for retries; it relies on the massive capacity of Host DRAM.

---

Appendix D: 802.11ax/be Features and Fi-Wi Integration

Modern Wi-Fi standards — particularly 802.11ax (Wi-Fi 6/6E) and 802.11be (Wi-Fi 7) — introduce features that appear to address some of the same problems as Fi-Wi: uplink scheduling, spatial reuse, and multi-AP coordination. This appendix clarifies how these features relate to Fi-Wi's architecture, where they're complementary, and why they don't eliminate the need for Fi-Wi's centralized data-plane approach.

Key takeaway: 802.11ax/be features like trigger frames and multi-AP coordination are valuable enhancements that Fi-Wi can leverage when client support is available, but they operate at a different architectural level (per-AP MAC features vs. building-scale data-plane unification) and cannot replace Fi-Wi's core innovations: centralized queues, shared state, L4S marking coordination, and dynamic RF grouping across the entire building.

D.1 Trigger Frames and Uplink Scheduling

802.11ax introduced trigger frames (TF) to enable centralized uplink scheduling. Instead of clients contending for the channel using stochastic EDCA backoff, the AP sends a trigger frame that grants specific clients permission to transmit on specific OFDMA resource units (RUs) or spatial streams at a specific time.

What trigger frames provide:

• Uplink OFDMA: Multiple clients can transmit simultaneously on different frequency resource units within the same channel, controlled by the AP's trigger frame.
• Uplink MU-MIMO: Multiple clients can transmit simultaneously on different spatial streams, again coordinated by the AP.
• Deterministic uplink airtime: The AP controls which clients transmit when, reducing contention and collision probability on the uplink.

How trigger frames align with Fi-Wi:

Trigger frames match Fi-Wi's philosophy of centralized scheduling rather than distributed contention. In a Fi-Wi deployment where RRHs support 802.11ax and clients support uplink OFDMA/MU-MIMO, the concentrator can:

• Generate trigger frames via RRHs to schedule uplink transmissions from multiple STAs
• Coordinate uplink RU allocation across an airtime domain to minimize collision and maximize spatial reuse
• Use shared state (CSI, queue levels, traffic patterns) to make smarter trigger scheduling decisions than an isolated AP could
• Treat uplink trigger-based transmissions as part of the same controlled airtime resource managed by the group queue

Reality check — client support in 2025:

While 802.11ax was ratified in 2019, uplink OFDMA support remains inconsistent. Crucially, trigger frames only control 802.11ax/be clients; legacy devices (iPhone 11, older IoT) are invisible to this schedule. These legacy clients cannot parse the trigger, so they continue to contend via random EDCA, acting as unmanaged interference sources. In contrast, Fi-Wi's reception diversity (Section 9.1) enhances uplink reliability for all clients, regardless of generation, by combining signals from multiple RRHs.

D.2 Why Trigger Frames Don't Eliminate the Need for Fi-Wi

A natural question: "If 802.11ax APs can use trigger frames for uplink scheduling, why do we need Fi-Wi's centralized architecture?"

Answer: Trigger frames address only a small subset of the problems Fi-Wi solves, and even for uplink scheduling, they provide per-AP control, not building-scale coordination.

What trigger frames do NOT provide:

1. Centralized queues across APs: Even with trigger frames, each AP maintains its own independent downlink and uplink queues. There's no shared queue state, no unified bottleneck, and no coordinated ECN marking across APs.
2. Shared state: Trigger-capable APs still operate autonomously. They don't share CSI, retry statistics, airtime usage, or queue metrics. Each AP makes trigger scheduling decisions based only on its local view.
3. Coordinated L4S marking: There's no mechanism in 802.11ax for multiple APs to coordinate ECN marking or present a single logical bottleneck to L4S. Each AP marks (or doesn't mark) independently.
4. Dynamic RF grouping: 802.11ax APs don't dynamically reconfigure which radios share airtime resources based on interference, CSI structure, or collapse risk. They're fixed islands.
5. Tail latency control: Trigger frames help with uplink efficiency, but they don't address the fundamental problem of hidden queues, uncontrolled aggregation, and tail latency blowup under load across a multi-AP building.

D.3 OFDMA Resource Units and Airtime Domains

802.11ax OFDMA subdivides a channel into resource units (RUs). In Fi-Wi, an airtime domain is a logical entity representing a shared RF resource. OFDMA RUs provide finer-grained subdivision of that airtime resource.

Conceptually:

• Without OFDMA: An airtime domain on channel 36 has one resource: the full 80 MHz channel. Only one transmission can occur at a time within that domain.
• With OFDMA: The same airtime domain can be subdivided into RUs. The concentrator's scheduler can allocate RU1 to STA-A, RU2 to STA-B simultaneously.

This does not change the fact that all RRHs in that airtime domain share a single group queue and marking point. It simply allows the service process to be more efficient.

D.4 BSS Coloring and Spatial Reuse

802.11ax BSS coloring allows STAs to distinguish between intra-BSS frames (same color) and inter-BSS frames (different color), enabling more aggressive spatial reuse.

Relationship to Fi-Wi RF grouping: Fi-Wi's dynamic RF grouping (Section 7) serves a similar but more sophisticated purpose. Fi-Wi uses richer information (CSI, retry statistics, airtime) to decide grouping, not just RSSI thresholds. In a Fi-Wi deployment, the concentrator can assign BSS colors to RRHs strategically: RRHs in the same airtime domain get the same color, while isolated domains get different colors.

D.5 802.11be (Wi-Fi 7) Multi-AP Coordination

802.11be (Wi-Fi 7) introduces multi-AP coordination features that appear to move in Fi-Wi's direction:

• Coordinated Spatial Reuse (C-SR): Multiple APs coordinate their transmissions to reduce mutual interference.
• Coordinated Beamforming (C-BF): APs share CSI and coordinate beamforming weights to improve SINR.
• Joint Transmission (JT): Multiple APs transmit the same data to a client simultaneously, providing diversity or increased SNR.

How these relate to Fi-Wi: These features acknowledge the problem of autonomous APs but approach it incrementally. 802.11be uses distributed AP-to-AP messaging, which limits scale and speed. Fi-Wi centralizes the data plane, enabling deeper coordination than distributed messaging can achieve.

D.6 Deployment Strategy: Mixed Client Populations

A key advantage of Fi-Wi's architecture is that it degrades gracefully with mixed client populations and doesn't require forklift client upgrades.

Client capability tiers in a 2025 deployment:

1. Legacy 802.11ac and earlier: No trigger frame support, no OFDMA, no BSS coloring.
   • Fi-Wi provides: centralized downlink queuing, L4S marking, reception diversity on uplink, beacon shaping to reduce contention.
   • Result: Significantly better latency and stability than traditional multi-AP, even without 802.11ax features.
2. 802.11ax with partial features: May support downlink OFDMA, BSS coloring, some power save enhancements, but not uplink OFDMA or uplink MU-MIMO.
   • Fi-Wi provides: All of the above, plus downlink MU-OFDMA where beneficial, coordinated BSS coloring across RRH groups.
   • Result: Better spatial reuse and efficiency, still robust to clients that don't support full 802.11ax.
3. 802.11ax with full features: Supports uplink OFDMA and uplink MU-MIMO via trigger frames.
   • Fi-Wi provides: All of the above, plus trigger-based uplink scheduling, uplink MU-OFDMA for small packets, coordinated uplink/downlink airtime management.
   • Result: Bidirectional sub-millisecond latency control, maximum airtime efficiency.
4. 802.11be (Wi-Fi 7): Adds MLO, 320 MHz channels, 4096-QAM, possibly multi-AP coordination support.
   • Fi-Wi provides: Can leverage MLO via concentrator coordination (Section 14.3), wider channels for capacity, and potentially integrate with 802.11be multi-AP features while maintaining superior shared-state coordination.
   • Result: Cutting-edge performance while maintaining backward compatibility.

Deployment strategy:

• Day 1: Fi-Wi delivers core benefits (low latency, L4S, collapse avoidance) with whatever client mix exists, including large populations of 802.11ac and partial-802.11ax devices.
• Ongoing: As clients are naturally refreshed, Fi-Wi automatically takes advantage of newer capabilities without requiring RRH hardware changes.

D.7 Summary: 802.11ax/be as Enhancements, Not Replacements

802.11ax and 802.11be introduce valuable features — trigger frames, OFDMA, BSS coloring, multi-AP coordination — that align with Fi-Wi's centralized control philosophy and can enhance Fi-Wi deployments when clients support them. However:

1. These features do not eliminate the need for Fi-Wi's architecture. They provide per-AP enhancements and limited inter-AP coordination, but they cannot create the unified data plane, shared state, and building-scale control that Fi-Wi provides.
2. Fi-Wi is designed to work with or without them. Core benefits (centralized queues, L4S marking, tail latency control) are independent of client 802.11ax/be support.
3. Fi-Wi leverages them when available. As client capabilities improve, Fi-Wi automatically benefits from trigger-based uplink scheduling, OFDMA efficiency, and other enhancements without requiring architectural changes.

In short: 802.11ax/be features make Fi-Wi better, but Fi-Wi solves problems these standards cannot address within the constraints of the distributed-AP model. Fi-Wi is not "better APs" — it's a different architecture that happens to integrate well with modern Wi-Fi standards as they evolve.

---

Appendix E: ASIC Evolution to Complexity

E.1 Why ASICs accumulate legacy complexity

Unlike software, ASICs cannot easily “refactor away” unused features. Removing blocks typically requires re-verifying entire subsystems, while adding blocks often requires verifying only the new logic. This asymmetry encourages accumulation:

• Removal risk > addition risk: removing a feature may impact existing customers in hard-to-observe ways; adding a feature does not threaten prior deployments.
• Career incentives favor addition: “I added feature X” is a visible win; “I removed 50K gates” is nearly invisible but carries blame if anything breaks.

Over many product generations, this leads to RTL codebases that only grow. Legacy modulation modes, preambles, power-save FSMs, calibration paths, and debug hooks persist long after their practical value has disappeared.

E.2 Real costs of legacy bloat

This accumulated complexity has tangible costs:

• Die area: more logic means a larger die, lower yield per wafer, and higher unit cost.
• Power: unused or rarely used blocks still consume leakage and often dynamic power.
• Verification: more RTL means more corner cases, more regressions, and longer bring-up.
• Timing closure: deep, interdependent datapaths make timing more difficult and fragile.

E.3 How Fi-Wi changes the design equation

Fi-Wi’s architecture separates the system into:

• Concentrator: complex, software-defined control plane and data plane intelligence.
• RRH: thin, timing-critical RF endpoint with minimal state.

This separation dictates where complexity must live. RRHs implement only what must be fast and deterministic: RF front end, PHY processing, minimal MAC TX/RX, DMA, PTP synchronization, and PCIe-over-fiber transport. All high-level behavior (queueing, L4S policy, aggregation strategy) lives in the concentrator.

E.4 Economic and engineering leverage

For a modern Wi-Fi chip at an advanced node, even a modest reduction in unnecessary logic can translate into significant savings: smaller die, lower power, simpler verification, and faster time to market.

E.5 Design principle for Fi-Wi RRH silicon

The guiding principle for Fi-Wi RRH design is:

Complexity belongs in the concentrator; only latency-critical functions belong in RRH silicon.

Concretely, this means: no autonomous AP queueing/scheduling logic, no legacy PHY/MAC support beyond what Fi-Wi needs, and no embedded firmware CPU managing per-station behavior at the edge.

---

Appendix F: A Day in the Life of a Packet (The "Preamble Shield" in Action)

To truly understand Fi-Wi, we must follow a single packet through the system at the microsecond scale. This narrative illustrates how the Workstation Concentrator (Section 14) and the Scatter-Gather RRH (Appendix C) collaborate to trick the physics of latency.

F.1 The Scenario

F.2 The Downlink Race (The "Preamble Shield")

T = 0 µs (Arrival): The video packet arrives at the Concentrator's NIC. The CPU timestamps it immediately.

T = 2 µs (The Decision): The Concentrator's software scheduler inspects the packet.

• Queue Check: It places the packet in the "Room 304 Airtime" queue.
• L4S Check: Current sojourn time is low (50 µs). No ECN mark applied.
• RF Check: Recent CSI shows RRH-A has the best vector to Alice.

T = 10 µs (The Setup): The scheduler posts a DMA Descriptor to RRH-A via PCIe.
Note: The payload data (1500 bytes) stays in the Concentrator. Only a 16-byte pointer moves to the edge.

T = 50 µs (The Trigger): RRH-A's LBT logic sees the airtime is clear. It begins the transmission sequence. This is where the magic happens:

The Race Against the PHY:
Action 1: RRH-A starts transmitting the 802.11 Preamble (PLCP) from its local SRAM. This takes 20 µs of airtime.
Action 2: Simultaneously, RRH-A issues a PCIe Read Request to fetch the payload from the Concentrator.

The payload must travel 200m up the fiber and back before the Preamble finishes transmitting.

T = 52 µs (The Fetch): The Read Request hits the Concentrator's PCIe controller. Because of the 92-lane non-blocking fabric (Section 14), there is zero switching delay.

T = 55 µs (The Return): The payload data flies back down the fiber.

T = 58 µs (The Handover): The payload data arrives at RRH-A's FIFO. The PHY is just finishing the last symbol of the Preamble.

T = 59 µs (Seamless Serialization): The PHY seamlessly switches from transmitting the Preamble to transmitting the payload. To the air, it looks like one continuous stream. The 200-meter fiber latency effectively vanished because it was hidden behind the mandatory PHY training sequence.

F.3 The Uplink Journey (Diversity & Sensing)

T = 200 µs: Alice sends a TCP ACK.

T = 204 µs (The Multi-Stat): Both RRH-A and RRH-B hear the ACK.

• RRH-A sees a strong signal (-45 dBm).
• RRH-B sees a weak reflection (-72 dBm).

T = 210 µs (The Race Up): Both RRHs push the packet + CSI metadata to the Concentrator.

T = 215 µs (The Deduplication): The Concentrator sees two copies of Sequence #104. It discards the weak one from RRH-B but keeps the CSI data to update the "Sensing Model" (detecting that someone is standing near RRH-B, blocking the line of sight).

F.4 Contrast with Legacy Wi-Fi

If this were a traditional AP:

• Downlink: The packet would be buffered in a per-TID queue inside the AP hardware. It might wait behind a large aggregate for another user (Head-of-Line blocking). ECN marking would be speculative (based on enqueue time), likely missing actual congestion events. The "Preamble Shield" trick is impossible because the AP CPU can't fetch data reactively—it must buffer it proactively.
• Uplink: Only the associated AP would hear the ACK. If Alice moved slightly and blocked the signal, the packet might be lost, requiring a MAC-level retry (adding 2-5ms latency). Fi-Wi's diversity combining prevents this retry.

F.5 Edge Cases and Advanced Scenarios

RRH Failure: If RRH-A fails during the prefetch (e.g., power loss), the concentrator detects the link loss immediately via PCIe link state. Because the packet payload never left Concentrator DRAM, the scheduler simply re-posts the descriptor to RRH-B. No packet is lost, and TCP does not see a drop.

Congestion: The scatter-gather pipeline depth allows the Concentrator to queue up the next descriptor while the current one is transmitting. This allows back-to-back TXOPs (SIFS spacing) without idle gaps on the air, even with the fiber latency.

Coordinated Transmission: The Concentrator can schedule RRH-A and RRH-B to transmit concurrently to spatially separated clients. It analyzes the CSI matrix to determine if spatial isolation is sufficient (>25 dB cross-coupling attenuation). If yes, both RRHs transmit simultaneously using standard 802.11 frames. If interference is detected, the Concentrator schedules sequential TXOPs. This dynamic decision happens per-packet based on real-time CSI.

F.6 Summary: The Packet's Perspective

From the packet's view, Fi-Wi provides uplink diversity, per-flow fair queuing, accurate ECN marking, and speculative DMA that hides PCIe latency. The packet experiences the network as a transparent, zero-wait pipe.

F.7 The Critical Insight: Timing vs. Intelligence

Fi-Wi separates timing (RRH hardware) from intelligence (Concentrator software), bridged by the speculative DMA prefetch pipeline. This allows the hardware to meet strict microsecond deadlines while the software retains the flexibility to run complex scheduling, L4S, and spatial multiplexing logic.

---

Appendix G: The Strategic Case for Fiber Infrastructure

The upfront cost of installing fiber is often the primary friction point for C-RAN adoption ("The Fiber Tax"). However, this framing ignores the physics of modern signaling and the macroeconomics of construction. Fi-Wi's reliance on fiber is not a tax; it is a strategic asset conversion.

G.1 The Physics of 100G (The Copper Wall)

We are hitting a hard physical limit with copper cabling. At modern data center speeds (100Gb/s), signal loss in copper is so high it is characterized in dB per inch.

• High-Frequency Attenuation: 100Gb/s signaling requires Nyquist frequencies of 25-50 GHz. At these frequencies, skin effect and dielectric loss in twisted-pair copper make long-distance transmission physically impossible.
• The Reach Limit: In the data center, 100G copper DAC cables are limited to 1-3 meters. To run 100G+ speeds through a building's walls (50-100m), copper is effectively dead.
• Future-Proofing: Fiber defies this gravity. It carries 100G, 400G, or 800G over hundreds of meters. Installing Cat6A today guarantees a "rip-and-replace" expense tomorrow.

G.2 Labor Rate Hedging (Inflation Proofing)

In low-voltage construction, the cost of cabling is dominated by labor (often 70-80%), not material.

• The Inflation Trap: Construction labor rates rise over time. Relying on copper commits the owner to re-cabling cycles at future, higher rates.
• The Fiber Hedge: Installing fiber today locks in today's labor rates for a permanent infrastructure asset.

G.3 Asset vs. Consumable

Unlike HDMI or Copper Ethernet—which are purpose-built cables engineered for a single generation—fiber is a raw transport medium. It is a "pipe for light" that supports Ethernet, DWDM, and PCIe-over-Fiber simultaneously.

While cable standards have cycled (Cat5e → Cat6 → Cat6A), they remain tethered to the legacy RJ45 connector. This physical interface is rapidly becoming obsolete. Fi-Wi recognizes that the connection is what matters, not the physical port. In this architecture, the 802.11 wireless interface becomes the new connector. By installing fiber once as a permanent asset and treating Wi-Fi as the universal 'plug' inside the room, the building infrastructure is 'one and done'. This finally breaks the cycle of physical obsolescence.

Appendix H: Centralized Observability and the ML Advantage

Fi-Wi's centralized architecture provides observability that is difficult or impractical to achieve in distributed AP systems. This appendix presents the Observability Matrix—a systematic comparison of what telemetry is directly observable, partially observable, or hidden across different measurement approaches. This complete visibility is the prerequisite for effective machine learning (Section 16) and deterministic L4S control.

The Observability Gap

Traditional Wi-Fi deployments rely on tools that provide only partial visibility into system state. Operators attempt to infer problems from symptoms (latency spikes, ECN marks, throughput degradation) without directly observing root causes (queue growth, retry timing, MCS selection under interference). This inference distance—the number of steps between observable effects and hidden causes—makes control systems less stable and limits the effectiveness of machine learning.

The table below compares observability across six measurement approaches. The legend indicates:

Direct: Directly measurable with microsecond-resolution timestamps

Partial: Partially observable or requires inference

Not Observable: Hidden or cannot be reliably measured

Observability Matrix

Telemetry / Metric	ESP32-C5 RF sensor	RPi 5 Monitor mode	RPi 5 L4S node	tcpdump Packet capture	iperf2 L4S	Fi-Wi Concentrator
Energy detect / CCA
Channel busy time
NAV / medium reservation
CSI / channel matrix
MCS / GI / NSS
PER / retry counts
RSSI / SNR
Queue depth
Sojourn time
ECN marks
One-way delay (OWD)
Responsiveness
Throughput / goodput
Deterministic playback

Provenance: the ESP32-C5 and RPi 5 columns reflect the Sonde capture system as built and code-audited (IDF v6, 2026-06-29) — see Appendix S for the field capability audit, NDJSON schema, and capture architecture. In particular: the ESP32-C5 exposes no hardware CCA counter, so channel busy time is a promiscuous-mode approximation (Partial) and energy detect / CCA is not observable from firmware; MCS / GI / NSS decode from he_siga1 per cur_bb_format (Partial, pending fix TB-D6); and retry is visible as a per-frame bit but never as a TX-side count (Partial). This matrix is the living reference for Sonde telemetry: capability changes land in Appendix S first, then update these cells.

Critical Observations

Queue Depth and Sojourn Time (highlighted rows):

These metrics are essential for L4S congestion control and machine learning. Traditional tools (tcpdump, Wi-Fi packet capture) cannot directly observe queue state because it exists inside firmware or kernel layers. While synchronized ingress and egress packet captures could theoretically infer queue depth through timing correlation, this approach requires nanosecond-precise time synchronization across physically separated capture points, perfect packet correlation despite potential losses, and still cannot observe firmware-internal retry queues, aggregation buffer states, or PHY scheduling decisions. External sniffers see the explosion (the packet hitting the air), but they cannot see the fuse burning (the packet sitting in the driver queue). Only centralized queueing architectures expose these values with direct microsecond-resolution timestamps.

MCS / GI / NSS (PHY Configuration):

Monitor-mode packet capture can partially infer MCS from radiotap headers, but this only shows what was transmitted—not the decision process, CSI data, or PER history that informed the choice. The Fi-Wi Concentrator has direct access to the complete decision state.

Deterministic Playback (bottom row):

This capability enables machine learning. Deterministic playback means the Concentrator can reproduce its own decision sequence from a log file: packet arrivals, queue transitions, scheduling decisions, MCS selections, and RRH transmission commands. While actual RF outcomes depend on station behavior and channel conditions that may vary, the Concentrator can replay its control decisions under the logged RF environment to evaluate alternative strategies offline and verify whether different MCS/scheduling choices would have improved performance. This is only possible when all Concentrator-controlled components operate under a single clock with complete state visibility. Distributed systems cannot reconstruct this causal chain from partial packet traces because they lack visibility into queue state, retry logic, and the decision-making process itself.

Why This Enables More Effective Machine Learning

Section 16 describes how Fi-Wi uses machine learning to optimize MCS transition rates. The observability matrix demonstrates significant practical advantages that Fi-Wi's centralized architecture provides for ML training:

• Direct Queue Observability: Can label training examples based on actual queue impact and sojourn time measurements
• Global CSI Visibility: Can learn cross-RRH interference patterns from centralized measurements
• Deterministic Replay Capability: Can evaluate whether alternative decisions would improve outcomes under the same conditions
• Low Inference Distance: 1-2 steps between PHY decisions and observable transport-layer effects versus 5-10 steps in distributed systems

Fi-Wi's centralized architecture provides these observability advantages. The Concentrator's event log becomes a high-quality training dataset where every state transition is labeled with measured outcomes under consistent instrumentation. While autonomous AP systems could attempt ML-based rate adaptation using the partial observability available to them, Fi-Wi's richer telemetry—particularly queue visibility, global CSI, and deterministic replay—enables significantly more effective learning and optimization.

Coordination Shares Outcomes; Fi-Wi Centralizes Causes

Coordinated AP systems can share summaries (throughput, ECN marks, interference reports) but cannot share hidden internal state (queue depth, firmware retry logic, aggregation decisions). This creates inference distance—the controller sees effects but not causes. Fi-Wi eliminates inference distance by removing autonomous decision-making from the edge. Queues, scheduling, and PHY selection are centralized under a single clock, producing an observable state graph where causes are explicit, replayable, and directly controllable. This architectural difference translates to measurably better ML training data quality.

Appendix I: Channel Width Orchestration and Service Time Variance

The Fi-Wi architecture treats channel width as a dynamic control parameter managed by the Concentrator. While 802.11be (Wi-Fi 7) emphasizes 320 MHz peak PHY rates, Fi-Wi's orchestration engine strategically selects 40 MHz channel widths in high-density environments to ensure Service Time Stationarity and the stability of the L4S control loop.

I.1 The Contention-Domain Collapse of Wideband Channels

In shared-spectrum MDUs (Multi-Dwelling Units), the theoretical gain of wider channels is often negated by contention-domain collapse. In a CSMA/CA environment, a transmission opportunity (TXOP) requires the entire bonded channel to be idle. In a 6-AP overlapping scenario with 50% aggregate airtime occupancy, the probability of finding all sub-bands simultaneously idle drops exponentially with bandwidth.

Under a simplified independent-sub-band occupancy assumption, a basic model suggests P(160 MHz idle) ≈ (P(40 MHz idle))^4, resulting in 4–16× fewer transmission opportunities. In practice, partial correlation between sub-bands moderates the exponent but does not eliminate the super-linear decline in idle probability. This leads to:

• Effective Airtime Utilization: 160 MHz utilization often falls below 10% due to fragmented TXOPs and CCA (Clear Channel Assessment) busy deferrals.
• NAV Extensions: A significantly higher probability of virtual carrier-sense freezes from neighboring uncoordinated cells occupying any portion of the wideband mask.

I.2 Queueing Theory and L4S Stability

From an M/G/1 queueing perspective, the performance of the L4S control loop depends on the stability of the service rate (μ). L4S stability requires frequent service opportunities and low variance in service time to prevent the decoupling of the sender's congestion window from the actual queue state.

• Service Time Variance (σ²): Wide channels in dense environments increase the variance of service time due to heavy-tailed CCA deferrals.
• Feedback Coherence: When a 160 MHz channel freezes for 100ms, it represents 2.5–20 RTTs of stale feedback for typical broadband RTTs (5–40 ms). Typical PI² parameterizations (Section 6.1) assume feedback coherence within a small number of RTTs; exceeding this threshold leads to oscillatory behavior.
• The 40 MHz Advantage: By providing a stationary service rate, 40 MHz ensures continuous ECN marking and stable RTT sampling, allowing the L4S loop to remain synchronized with the physical medium.

I.3 Link Adaptation and Spectral Robustness

Narrower channels reduce the probability that partial-band interference (e.g., unmanaged IoT bursts) forces a full MCS downgrade across the entire bonded width. This allows the Concentrator to maintain stable link adaptation and a predictable drain rate, avoiding the chaotic rate-shifting common in 160 MHz deployments.

I.4 Orchestration: Width as a Control Variable

Fi-Wi is not anti-wideband; channel width is an orchestrated variable. The system expands width opportunistically when contention is low to leverage PHY gains and contracts it to 40 MHz when deterministic latency is required. This prioritizes spatial reuse and airtime isolation over maximum burst rate—the fundamental technical unlock for Fi-Wi’s cell-per-room model.

I.5 Capacity Density: Throughput Under a Latency SLO

Fi-Wi optimizes Capacity Density under a Latency SLO, rather than peak PHY on a single link. In dense OBSS environments, wide channels reduce spatial reuse; narrower channels increase the number of bounded contention domains. Consequently, aggregate goodput per area increases even if per-link PHY decreases.

Metric Definition: Low-Latency Goodput Density (ρ_LL)

ρ_LL [Mbps / 1,000 sq ft] = (Σ Goodput_i) / Area | subject to p95 OWD ≤ 20ms

Where Goodput_i is the application-layer payload throughput delivered while maintaining the p95 one-way delay (OWD) constraint. The 20ms threshold reflects the target for interactive L4S applications.

Example Calculation (1,000 sq ft section of a 10,000 sq ft floor):

Assumptions: 50% aggregate offered load per BSS, default EDCA parameters, and no explicit inter-AP coordination in the autonomous case.

• 160 MHz Autonomous: 1 AP (Wide coverage) × 1.2 Gbps peak × 10% airtime efficiency ≈ 120 Mbps aggregate. Under heavy-tailed service intervals, a large fraction of packets violate the 20ms p95 constraint. When filtered by the latency SLO (ρ_LL definition), usable goodput is reduced to ~12 Mbps in this model.
• 40 MHz Fi-Wi: 8 RRHs (Cell-per-room) × 400 Mbps peak × 40% airtime efficiency ≈ 1,280 Mbps aggregate. Bounded service intervals allow nearly all traffic to meet the latency SLO, resulting in ~128 Mbps.

I.6 Application: Aligning Wireless Capacity to Gigabit WAN Service

To align with a Gigabit-class WAN service, the wireless architecture must match the aggregate wireline supply to orchestrated spatial demand. In a dense MDU, Contention Delay is 10–100× larger than serialization time. A single 160 MHz AP attempting to serve a Gigabit load creates a "fast but flaky" link that collapses under co-channel interference, delivering only a fraction of the ISP's provided capacity to real-time applications.

Fi-Wi resolves this by using 40 MHz orchestration to spread the Gigabit load across N coordinated spatial domains. This ensures that the building-wide Wireless Fabric can actually saturate a 1 Gbps WAN link with deterministic, multi-user goodput, rather than relying on single-device peak bursts that starve other users and destabilize shared airtime.

I.7 Aggregation Quantization and L4S Feedback Mismatch

L4S signals congestion at Layer 3 (IP ECN), but wideband Wi-Fi operates via massive Layer 2 A-MPDU aggregation to maintain PHY efficiency. This creates a fundamental control-loop mismatch:

• Feedback Sparsity: Massive 160 MHz aggregates transmit dozens of packets in a single burst. This results in "binary" feedback—either a massive burst of ECN marks or none at all—which prevents the PI² controller from calculating the smooth probability signal required for precise rate pacing.
• The Sawtooth Latency Trap: Wide channels induce a "fast but infrequent" cadence; the AP must buffer data longer to build efficient aggregates. This creates multi-millisecond stalls that look like random network noise to L4S senders, causing congestion window oscillation.
• Timing Incoherence: Traditional APs lose microsecond-level visibility once packets enter hardware aggregation pipelines. Fi-Wi's 40 MHz orchestration enforces smaller, frequent TXOPs (~250 µs), ensuring the MAC service frequency is significantly higher than the L4S control frequency, enabling accurate sojourn-time measurement and coherent ECN marking at the IP layer.

The Fi-Wi architecture addresses these challenges through its DualQ implementation (Section 6.2), which maintains separate queues for L4S and Classic traffic and performs per-packet sojourn time measurements at the Concentrator before entering the A-MPDU aggregation pipeline.

---

Comparison of Service Metrics (Dense MDU Contention Model)

Scenario: 2x2 MIMO, 6+ overlapping BSSIDs, shared unlicensed spectrum (5/6 GHz), 50% aggregate offered load, autonomous EDCA parameters. See Appendix J for full simulation parameters.

Metric	160 MHz (Autonomous CSMA)	40 MHz (Fi-Wi Orchestrated)
Peak PHY Rate (2x2, MCS 11)	~1.2 Gbps	~300-400 Mbps
Effective Airtime Utilization	<10% (Fragmented TXOPs)	30–50% (Planned reuse / Bounded domain)
Service Time Variance (σ²)	High (Heavy-tailed)	Low (Near-stationary)
Queue Service Interval (median)	Tens to >100 ms	5–15 ms (Stationary)
DualQ ECN Feedback Coherence	Sparse / Burst-marked	Continuous / Stable marking
Goodput Density (ρ_LL) (Mbps per 1,000 sq ft)	~12 Mbps (Overlapping contention domains)	~128 Mbps (8 RRHs, orthogonal 40 MHz channels)

Economic Conclusion: Under the Appendix J model assumptions, Fi-Wi's orchestrated 40 MHz architecture is projected to deliver approximately 10× higher usable goodput density compared to autonomous wide-channel deployments under realistic dense MDU conditions. This should be treated as a deployment hypothesis until validated with per-TXOP field telemetry against the stated baseline. The fundamental architectural advantage — that capacity scales with RRH density and spatial reuse rather than channel width alone — holds independent of the specific multiplier.

See Appendix J for detailed contention modeling and simulation methodology.

Appendix J: 10-Node MDU Simulation Methodology

This appendix details the Monte Carlo simulation and analytical models used to derive the Low-Latency Goodput Density (ρ_LL) metrics. The framework evaluates Fi-Wi's spatial capacity gains under realistic Multi-Dwelling Unit (MDU) contention scenarios.

J.1 Spatial and RF Environment Model

The simulation contrasts traditional wide-area coverage with Fi-Wi's localized orchestration.

Building & RF Assumptions:

• Geometry: 10,000 sq ft floor divided into 8 units (~1,250 sq ft each). Metrics are normalized to "per 1,000 sq ft" for comparative analysis.
• Path Loss Model: PL(d) = PL(d₀) + 10n log₁₀(d/d₀) + Xσ with n = 2.8.
• OBSS Overlap: Autonomous case assumes 6 neighboring BSSIDs audible at ≥ -62 dBm.
• Fi-Wi Isolation: 8 RRHs achieving >25 dB co-channel isolation through planned orthogonal reuse.

J.2 Contention and Backoff Logic

The simulation models 20 active stations (STAs) distributed across the 8-unit floor (average 2.5 STAs per unit). Service Time Variance (σ²) is calculated by observing the delay between TX_START and ACK_END across 10⁶ simulated TXOPs.

• Autonomous Case (160 MHz): All 20 STAs compete for the shared 160 MHz channel. Probability of transmission is modeled as: P(TX) = [1 - p_occ]^4, where p_occ is aggregate occupancy from overlapping neighbors.
• Fi-Wi Case (40 MHz): STAs are partitioned into 8 isolated airtime domains. P(TX) = 1 - p_local_occ, restricted to immediate room-level neighbors.

J.3 The ρ_LL Filtration Process

The Goodput Density is derived by filtering raw throughput through the 20ms p95 OWD constraint.

// Derivation for ρ_LL Calculation
for each packet i:
    delay_i = contention_delay + serialization_delay + retry_overhead
    if delay_i <= 20ms:
        accepted_payload += size_i
    else:
        dropped_from_goodput_metric++

ρ_LL = (accepted_payload) / (total_time * area)
  

J.3.1 Numerical Results and Derivation

The simulation produces the following goodput derivation for a 1,000 sq ft sections:

• 160 MHz Autonomous: 120 Mbps raw aggregate throughput. However, heavy-tail contention spikes cause 90% of packets to exceed the 20ms SLO. 120 Mbps * 0.10 = 12 Mbps ρ_LL.
• 40 MHz Fi-Wi: 1,280 Mbps raw aggregate throughput (across 8 orthogonal domains). Stationary service rate allows 99.8% of packets to meet the 20ms SLO. 1,280 Mbps * 0.998 = ~128 Mbps ρ_LL.

J.4 Traffic Model and Payload Composition

Traffic Type	% of Load	Constraint
Interactive (L4S/Gaming)	20%	Strict SLO subject
Streaming (4K Video)	50%	Freeze sensitive
Bulk (Background)	30%	Throughput focused

Appendix K: The Information-Theoretic Ceiling — Joint Multi-User Decoding

The uplink selection-combining approach described in §9.1 operates on post-FCS frames: each RRH independently demodulates, and the Concentrator picks the best surviving copy. This is the weak form of what Fi-Wi's architecture makes possible. The strong form, which is not part of the baseline design, is worth naming because it defines the information-theoretic ceiling of the entire Wi-Fi design space — and because Fi-Wi's physical architecture is the natural platform for reaching it.

K.1 The Single-User Ceiling Is Not the Real Ceiling

Conventional Wi-Fi design treats the single-user Shannon capacity of a channel as the fundamental capacity ceiling, and organizes the MAC (EDCA, NAV, CSMA/CA) around serializing access to stay beneath it. When two stations transmit simultaneously in the same contention domain, the MAC treats the event as a collision — an error to be avoided by backoff and carrier sense, discarded, and recovered through retransmission.

At the PHY layer, the same event is not an error. It is a linear superposition of two signals plus noise, of the form:

y = h₁x₁ + h₂x₂ + n

Multi-user information theory establishes that such superpositions are decodable in principle. The Slepian–Wolf theorem (1973) and the broader literature on multi-user channel capacity show that the aggregate capacity of a channel shared by multiple simultaneous transmitters is substantially greater than the single-user Shannon capacity, provided the receiver has access to multiple spatially-separated observations of the aggregate signal and operates at sampling rates sufficient to exploit the structure of the superposition. This decodability is not unconditional: it requires spatial diversity across receivers, tight synchronization, and sufficient computational capacity at the decoding point. §K.5 addresses these costs directly.

The magnitude of the gap between centralized and distributed processing is not merely theoretical. The cell-free massive MIMO literature, which models the cellular analog of Fi-Wi's architecture — many distributed single-antenna access points forwarding to a central processing unit — has quantified the advantage extensively. Ngo, Ashikhmin, Yang, Larsson, and Marzetta (2017) showed that a cell-free architecture delivers roughly a fivefold improvement in 95%-likely per-user throughput over a conventional small-cell (distributed AP) deployment under uncorrelated shadow fading, rising to tenfold under correlated shadow fading — precisely the regime that dense MDU deployments exhibit. Björnson and Sanguinetti (2020) sharpened this result: the large gains require centralized Minimum Mean Square Error (MMSE) combining operating on global signal knowledge, not distributed per-AP processing. Naive distributed combining leaves most of the available gain on the table. This directly corroborates the architectural prerequisite developed in §K.3: the gain is not a property of "having multiple APs" but of combining their observations at a single point with sufficient signal access.

These results are not directly transferable to Wi-Fi — the cell-free literature assumes licensed-band coordinated transmission, single-antenna APs, and time-division duplex operation, none of which map cleanly to unlicensed-band CSMA/CA. But they establish the scale of the architectural delta between centralized and distributed processing: it is measured in integer multiples, not percentage points. Any Wi-Fi architecture that forecloses centralized joint processing by construction is leaving this multiplier permanently on the table.

There is an important framing point that separates the information-theoretic argument from the MAC-centric reading of it. The ceiling is not reached by decoding collisions. That framing is an artifact of conventional Wi-Fi's MAC, which defines "collision" as a discrete event in the first place. The ceiling is reached by designing receivers that do not treat simultaneous transmissions as collisions at all — receivers that oversample well above the nominal signal Nyquist rate, exploit multipath structure in the time domain rather than filtering it out, and use channel codes strong enough to separate signals that superimpose additively. In this framing, what the MAC would have called a collision is simply more signal the receiver has access to. The argument is not "we can recover from collisions that conventional Wi-Fi cannot"; it is "the concept of a collision as a primitive is itself a design choice, and it is the wrong one if the goal is capacity." This is the distinction Reed draws sharply between EE receiver design (which has understood this for decades, even if it is taught poorly) and the CS-flavored reading of Shannon as a law of nature. Fi-Wi's architecture is neutral on this framing question in the baseline design, but it is the only Wi-Fi topology in which the EE-correct receiver design is even physically possible.

The takeaway is that conventional Wi-Fi is not capacity-limited in the Shannon sense. It is architecture-limited: its MAC collapses performance far below the PHY-layer capacity because it discards information the PHY layer could, in principle, recover.

K.2 Why This Is Not MIMO

The most common objection to the argument in K.1 is that modern Wi-Fi already uses multiple antennas and already supports multiple simultaneous users — through MIMO, MU-MIMO, UL-OFDMA, and 802.11be Multi-AP Coordination. Isn't the multi-user ceiling already being reached? The answer is no, and the distinction is not a matter of degree. Each of these techniques operates under architectural constraints that the Slepian–Wolf regime explicitly does not require.

Single-user MIMO uses multiple antennas at one transmitter and one receiver to create parallel spatial streams between a single pair of endpoints. The antennas are co-located — within a wavelength or few of each other on the same radio, sharing a single RF front-end and a single timing reference. MIMO raises the capacity realized between those two endpoints; it does not address the multi-user channel. It is a more efficient realization of the single-user Shannon ceiling, not an escape from it.

Downlink MU-MIMO (802.11ac and later) uses a multi-antenna AP to transmit different streams to different users simultaneously, using channel state information to pre-code each transmission so each user sees its own stream and the others as suppressed interference. The multi-user aspect is on the receive side of the downlink; the transmitter is still a single AP with a single co-located antenna array, orchestrating what every user receives. The AP is both the scheduler and the sole source of the aggregate signal.

Uplink MU-MIMO and UL-OFDMA (802.11ax) come closer to multi-user reception: multiple clients transmit simultaneously on AP-assigned resource units in response to a trigger frame, and the AP's multi-antenna array separates them. But the simultaneity is scheduled — the AP tells each client which RU and when to transmit, and all receive-side processing happens inside one AP with co-located antennas sharing a single RF front-end and one set of ADCs. The clients are cooperating with the AP's schedule. This is a coordinated multiple access scheme, not joint decoding of uncoordinated overlapping transmissions.

Multi-Link Operation (MLO) in 802.11be lets a single station maintain simultaneous links across multiple radios (different bands, different channels) and aggregate them at the MAC layer through a Multi-Link Device (MLD) abstraction. Each link below the MLD runs its own full PHY pipeline — independent channel estimation, equalization, decoding, and FCS check — and the MLD stitches the completed per-link results together above. MLO operates at a higher layer than this appendix is concerned with: it aggregates post-decode traffic across multiple already-completed PHY pipelines, increasing throughput by parallelism, but each constituent link still operates at the single-user Shannon ceiling on its own channel. MLO is not multi-user joint decoding any more than MU-MIMO is; it operates within the coordinated regime, just at the MAC aggregation layer above per-link PHY rather than within a single PHY chain.

802.11be Multi-AP Coordination (see Appendix D.5) is the closest thing in the standard to what this appendix describes. MAP-Co lets multiple APs coordinate over their backhaul to perform coordinated beamforming, coordinated OFDMA, or joint transmission. It is a genuine step toward exploiting multiple physical observation points. But it remains a coordination protocol: APs negotiate to avoid interfering with each other, or to jointly serve clients in controlled ways, through standardized inter-AP messaging. The receive-side processing for each frame still happens at one AP at a time. The multi-AP aspect is scheduling, not decoding.

The Slepian–Wolf regime is categorically different along three axes simultaneously:

• Antennas are spatially distributed, not co-located. Fi-Wi's RRHs sit meters to tens of meters apart — in different rooms, on different floors. This is distributed MIMO or cell-free MIMO in the cellular literature, and the channel-correlation, timing-discipline, and fronthaul requirements are structurally different from co-located arrays. Co-located MIMO exploits sub-wavelength phase differences between adjacent antennas; distributed reception exploits entirely different propagation paths through the environment.
• Transmissions are uncoordinated, not scheduled. MU-MIMO and UL-OFDMA require the AP to orchestrate who transmits when and on which resource unit. The regime Slepian–Wolf describes operates on uncoordinated simultaneous transmissions — transmitters need not be aware of each other, share a peer-to-peer schedule, or cooperate through any coordinating signal. This is a stronger capability than scheduled multi-user access: it applies to exactly the overlap events that conventional MAC design is built to avoid, including the collisions between stations that have no awareness of one another's presence.
• Separation happens at a central concentrator fed by fronthaul, not at one AP. MU-MIMO's receive-side separation happens inside one chip with one set of ADCs sharing one timing reference and one RF front-end. Fi-Wi's regime requires combining samples from physically separate radios over fiber at a concentrator with global state. This is the architectural prerequisite that no single-AP solution — however advanced its PHY — can satisfy, and it is the prerequisite that 802.11be MAP-Co deliberately avoids requiring, in order to remain deployable over conventional distributed-AP topologies.

	Single-user MIMO	DL MU-MIMO	UL MU-MIMO / OFDMA	802.11be MAP-Co	Fi-Wi joint decoding
Antennas	Co-located at AP	Co-located at AP	Co-located at AP	Per-AP arrays	Distributed RRHs
Transmitter coordination	N/A (1 Tx)	AP precodes all	AP schedules via trigger	APs negotiate over backhaul	None required
Receive processing	Single AP PHY	Single AP PHY	Single AP PHY	Per-AP, coordinated	Centralized concentrator
Capacity ceiling	Single-user Shannon	Shared single-user	Shared single-user	Coordinated multi-AP	Multi-user (Slepian–Wolf)
Key constraint	Channel rank	Channel rank	Scheduling overhead	Backhaul coordination	Topology (fronthaul + geometry)

The short version: standard MIMO makes a single ear more efficient at hearing one speaker. MU-MIMO and UL-OFDMA let that ear hear multiple speakers on a schedule. MAP-Co lets multiple ears negotiate who listens to whom. Fi-Wi turns the entire building into a synchronized array of ears that can decode transmissions whether they are coordinated or not.

K.3 The Three Architectural Prerequisites

The architectural distinction this appendix identifies can be stated as a single property: a Wi-Fi functional split is two crossings, not one. The signal boundary (continuous RF → sampled I/Q) decides how much of the received waveform survives the fronthaul, and therefore whether joint decoding across radios is even possible. The packet boundary (PHY → MAC) decides where the scheduler lives, and therefore whether the network is contention-based or coordinated. Every Wi-Fi architecture can be characterized by which of these two crossings it makes.

Conventional Wi-Fi crosses neither. Antenna to MPDU collapses into one chip. That is the architectural reason Wi-Fi stays a contention system, regardless of how many spatial streams or how much spectrum is added to the standard. PHY improvements raise the per-link single-user ceiling; they do not move the scheduler off the radio.

Cloud-managed Wi-Fi (Mist, Meraki, Ubiquiti) crosses the packet boundary but not the signal boundary. The control plane is centralized; the MAC/PHY is not. Each AP still decodes locally and still contends locally on the medium. The result is better dashboards, not a better MAC. The fronthaul carries management traffic and aggregated telemetry, not signal information.

5G crossed both more than a decade ago. That is what O-RAN Split 7.2x and CPRI Split 8 are for. Wi-Fi has not.

Fi-Wi crosses both. The signal-boundary crossing lets the Concentrator jointly decode across radios — the cell-free MIMO multiplier no single AP can reach. The packet-boundary crossing moves the scheduler off the radio and onto the fronthaul — deterministic airtime allocation instead of CSMA backoff. Either crossing alone is a half-measure. Together they are what turn Wi-Fi from a contention system into a scheduled one.

Reaching this ceiling requires at least three physical prerequisites — one for each half of the spatial-diversity requirement, plus a third for the sample-level access the Concentrator must have. Fi-Wi satisfies the first two by construction; the third is a silicon limitation, not an architectural one.

1. Multiple synchronized spatial observations. The receiver must have access to observations from antennas separated in space, with a common time reference precise enough for sample-level coherent processing of the aggregate signal. Standard 802.11 TSF at microsecond resolution is not sufficient for this — sample-level coherent processing requires alignment on the order of nanoseconds or below, depending on carrier frequency across receiving antennas, several orders of magnitude tighter than what TSF provides. Fi-Wi's distributed RRH topology supplies the spatial-diversity half of this prerequisite directly. The timing half is supplied by the Concentrator-resident master clock distributed to each RRH over the PCIe-over-Fiber fronthaul (§5.7), which provides the sample-level reference that vanilla TSF cannot. RRHs slave to this master rather than to their own local oscillators, eliminating the inter-RRH drift that independent AP clocks exhibit.
2. Centralized access to raw observations. Joint decoding cannot be performed at any single RRH — the information required to separate overlapping transmissions is distributed across receivers. The observations must be combined at a single processing point. Fi-Wi's PCIe-over-Fiber fronthaul removes this as a first-order constraint: the Concentrator has the bandwidth and architectural position to operate on data from every RRH simultaneously. The bottleneck that cripples distributed AP architectures — each AP committing to a local decode decision before anything leaves the box — is absent by construction.
3. Sample-level PHY access. Joint multi-user decoding — and more broadly, any receiver design that exploits the full information content of the aggregate signal — requires the Concentrator to operate on raw I/Q samples from each RRH, at sampling rates well above the signal's nominal Nyquist rate and without early demodulation or hypothesis collapse, with correlated timing metadata. This oversampling requirement is specifically the thing conventional narrowband 802.11 radio designs get wrong: they sample at rates adequate only to recover a single assumed transmitter, treating Nyquist as a rate bound rather than as a theorem about a specific signal model, and discarding the information that would let a better receiver exploit multipath structure or separate superimposed signals. COTS Wi-Fi silicon is built around this assumption by construction. The PHY pipeline commits to a single-user demodulation hypothesis early, and by the time the frame reaches the host it has already been reduced to either "decoded MPDU" or "FCS fail." The information required for joint decoding has been discarded before the fronthaul ever sees it.

K.3.1 Matrix Formulation

K.3.1.1 Signal-Chain Comparison

The following diagram compares three architectures across the full receive chain from antenna to WAN. Three distinct signal-domain regions are annotated at the top of the diagram: the continuous domain (analog signals as functions of time), the discrete z-domain (sample-level digital signals after the ADC), and the packet domain (post-decode MPDUs and PDUs as they cross transport links). The architectural distinction between Fi-Wi and conventional Wi-Fi APs lives entirely in how the signal flows through these domains, and where the transitions between them occur.

The 3GPP/O-RAN functional split notation, summarized in the legend at the bottom of the diagram, identifies the partitioning point between the radio unit and the centralized processing unit. Fi-Wi today operates internally at Split 6 (MAC-PHY boundary), with PHY decoding completed at the RRH or host before PCIe transport. Fi-Wi future operates at Split 8 or Split 7.2x, with raw I/Q samples or post-FFT/CP-removal symbols crossing PCIe-over-Fiber to the Concentrator for joint multi-user decoding. Conventional Wi-Fi APs operate at what we denote Split 8′ — a degenerate variant of Split 8 in which no fronthaul cut exists at all; the entire signal chain is contained within a single chip, with no architecturally-defined separation between radio unit and processing unit.

Three observations are worth drawing out from the diagram before the formal model in the next subsection.

The analog stages are identical across all three architectures. Antenna, RF front-end, mixer, and ADC operate identically in Fi-Wi today, Fi-Wi future, and a conventional Wi-Fi AP. This is the concession the appendix makes explicitly: the antenna-to-ADC path is not where the architectural distinction lives, and any claim that Fi-Wi has an analog-domain advantage over conventional radios would be wrong. The architectural distinction lives entirely in the post-ADC signal flow.

Fi-Wi today and Fi-Wi future share an identical topology. Both have three distributed RRHs with PCIe-over-Fiber transport to a Concentrator. What differs is what each RRH does with its post-ADC signal. Fi-Wi today runs the local PHY/MAC pipeline at the RRH or in the host CPU and ships decoded MPDUs across PCIe, operating internally at Split 6. Fi-Wi future runs only the fronthaul transport operator T_i(z) at the RRH and ships raw I/Q samples across PCIe, with all PHY processing including joint multi-user decoding happening centrally at the Concentrator. The migration from today to future is a silicon roadmap question — which PHY taps a future RRH chipset exposes, and at what sampling rate. The topology is already correct.

Conventional Wi-Fi APs cannot reach Fi-Wi future regardless of silicon evolution. The architectural property that forecloses joint decoding is not the device label "AP" but the absence of an externally-exposed fronthaul cut. An architecture without an externally-exposed fronthaul cut has no migration path to lower splits. Conventional Wi-Fi APs are an instance of this architecture; so are any RRH designs that complete PHY decoding internally before fronthaul transport. Fi-Wi today operates at Split 6 internally — the same internal cut location as a conventional AP — but with the fronthaul infrastructure sized for sample-level transport and a Concentrator architected to operate on M parallel sample streams. The migration to Split 8 or 7.2x is silicon-gated, not topology-foreclosed, because the topology is already correct for the future split. There is no equivalent of the migration arrow on the conventional AP row because there is no architectural path forward. A conventional AP with better silicon is still a conventional AP; it cannot become a low-split deployment without changing topology, which means changing the physical plant. This is the asymmetry the rest of K.3 is built to formalize.

K.3.1.2 Continuous-Domain Model

Let K denote the number of simultaneously transmitting stations and M the number of receivers (RRHs in Fi-Wi, APs in a conventional deployment). For station k with transmitted complex baseband signal s_k(t), the analog signal at the antenna of receiver i is

r_i(t) = Σ_k=1^K (h_ik ∗ s_k)(t) + n_i(t)

where h_ik(t) is the channel impulse response from transmitter k to receiver i, ∗ denotes convolution, and n_i(t) is the additive noise process at receiver i. The signal then passes through the analog RF front-end (LNA, mixer, baseband filter) and the ADC, producing the discrete-time sequence y_i[n] = ADC(r_i(t)) at the ADC output. This y_i[n] is the first quantity in the discrete z-domain — everything before the ADC is continuous-time analog.

The continuous-domain stage is identical across all three architectures in the diagram. The antennas, RF front-end, mixer, and ADC operate on r_i(t) the same way in Fi-Wi today, Fi-Wi future, and a conventional Wi-Fi AP. The architectural distinction this appendix identifies does not live in the analog signal chain, and Fi-Wi makes no claim of analog-domain advantage over conventional radios. What follows in K.3.1.3 operates entirely in the z-domain on y_i[n], which is where the architectural distinction does live.

K.3.1.3 Discrete-Domain Model

The architectural distinction between Fi-Wi and a conventional Wi-Fi AP is captured by a single per-RRH operator: the fronthaul transport operator T_i(z), which transforms the ADC output y_i[n] into whatever quantity the architecture delivers to the central processing point. The character of T_i(z) is what differs across the three architectures, and that character is what determines which subsequent processing is possible.

Define the per-RRH transported signal at the Concentrator as

z_i[n] = T_i(z) · y_i[n]

The aggregate received vector at the Concentrator at sample n is the M-dimensional stacking

R[n] = [ z₁[n], z₂[n], …, z_M[n] ]^T

The Concentrator combines this aggregate vector through a linear combining matrix W ∈ ℂ^K×M, producing K combined streams, one per intended user:

combined streams = W · R[n]

A decision function f(·) — the demapper, channel decoder, and FCS check — operates on the combined streams to produce the decoded streams

Ŝ[n] = f(W · R[n])

The matrix W is unconstrained in this formulation: it can be matched filter, zero-forcing, MMSE, regularized zero-forcing, or successive interference cancellation, depending on the channel model and the noise structure. Cell-free MIMO results (Ngo et al., 2017; Björnson and Sanguinetti, 2020) establish that under MMSE combining the achievable sum rate scales as a multiplier of the single-user ceiling, with the multiplier rising from approximately 5× under uncorrelated shadow fading to 10× under correlated shadow fading. Reaching this multiplier requires W to operate on the full M-vector R without per-receiver decoding decisions having been made first.

The architectural distinction between Fi-Wi and conventional Wi-Fi can be stated as a linearity property of T_i(z). The operator T_i(z) is linear when it preserves the algebraic structure of y_i[n] through to the Concentrator — that is, when no hard-decision step (FCS verification, MPDU framing, packet-level transport) intervenes between the ADC output and the input to W. Joint multi-user decoding requires T_i(z) linear; without linearity, the soft information that W needs to combine across receivers has already been collapsed into hard-decision form, and no downstream processing can recover it. The conventional Wi-Fi AP has T_i(z) nonlinear by construction, because the per-AP FCS decision is nonlinear; this is not a silicon limitation but a categorical property of the architecture.

This last condition is exactly where the three architectures diverge, and the divergence is encoded entirely in the character of T_i(z):

Fi-Wi future. T_i(z) is the fronthaul transport operator at sample-level depth — a near-identity mapping (with channel-modeling for PCIe-over-Fiber compression, framing, and finite delay) that delivers raw I/Q from RRH i to the Concentrator with the architectural information content preserved. The aggregate R[n] is in the z-domain, contains the full information from all M receivers, and admits any combining matrix W. This corresponds to 3GPP TR 38.801 Split 8 (raw I/Q over fronthaul) or O-RAN WG4 CUS.0 Split 7.2x (post-FFT, pre-decode symbols over fronthaul) — the regimes in which W retains the full optimization space.

Fi-Wi today. T_i(z) is the composition of the full local PHY pipeline (channel estimation, equalization, demapping, channel decoding, FCS verification) followed by PCIe packet transport. The output that reaches the Concentrator is no longer in the z-domain but in the packet domain — decoded MPDUs, not samples. The aggregate R[n] in this regime is a vector of decoded packets, not soft information, and the combining matrix W collapses to a per-RRH selection function. This is Split 6 — the MAC-PHY boundary — and it forecloses joint decoding because the soft information that W would need to operate on has been discarded by f(·) at each RRH before fronthaul transport. Selection diversity and centralized scheduling are still available; the cell-free MIMO multiplier is not. The migration path to Fi-Wi future is the silicon change that strips the local PHY pipeline out of T_i(z), leaving only the sample-level transport.

Conventional Wi-Fi AP. T_i(z) is the entire PHY/MAC pipeline executed inside one chip with no architecturally-defined fronthaul cut at all — the Split 8′ pattern. There is no R[n] vector at any aggregation point because no aggregation point exists; each AP terminates its own receive chain and emits packet-domain output to a LAN backhaul that does not connect to any peer AP's pre-decode samples. W is undefined because the inputs to it never coexist at any single processing point. This is the architectural foreclosure the Split 8′ legend entry names: it is not a question of which combining algorithm to use, because the data structure on which combining would operate cannot be assembled.

The matrix formulation makes precise what the diagram shows visually. The two crossings the K.3 introduction identified are formally the two structural changes that transform T_i(z) from the Split 8′ pattern (entire PHY/MAC in one chip, no cut) into the Split 8 / 7.2x pattern (sample-level transport only). The signal-boundary crossing strips the PHY pipeline out of T_i(z), so R[n] is in the z-domain rather than the packet domain. The packet-boundary crossing is the architectural fact that the scheduler now lives at the Concentrator, after f(·), rather than inside each RRH's chip. W's unconstrainedness depends on both crossings: without the signal crossing, R doesn't carry the right information; without the packet crossing, the medium access decisions that produce R are still being made locally by each radio's MAC, which means transmissions still collide rather than being scheduled.

K.3.2 Architectural Summary

Prerequisite	Fi-Wi implementation	Competitive status
Spatial observations	Distributed RRHs across the building	Co-located antenna arrays per AP; no inter-AP spatial diversity at decode
Shared time reference	Frequency-locked TSF (§5.7)	Independent AP clocks drift relative to each other
Data centralization	PCIe-over-Fiber fronthaul to Concentrator	Backhaul carries post-decode data only; no global observation point
PHY raw samples	Architecturally exposable; silicon-gated	Foreclosed by topology regardless of silicon evolution
Matrix-level condition	Ti(z) linear; W unrestricted	Ti(z) nonlinear; W block-diagonal

These three prerequisites together define a narrow architectural path — and only one Wi-Fi topology satisfies them simultaneously.

K.4 Topology vs. Silicon

That simultaneous satisfaction matters because topology is expensive to change and silicon is not. A distributed AP mesh — regardless of how advanced its PHY becomes — cannot reach the multi-user ceiling. The aggregate signal is never reconstructable at a single decoding point, because there is no single decoding point; each AP is architecturally committed to local decisions. No silicon revision fixes this. The ceiling is set by the topology.

Fi-Wi inverts this relationship. The topology is already correct: synchronized RRHs forwarding to a centralized Concentrator over fronthaul that is not bandwidth-limited. What remains is the silicon interface — a question of which PHY taps a future RRH chipset exposes, and at what sampling rate. That is a silicon roadmap conversation, not an architectural one. When RRH silicon evolves to expose the necessary interfaces, Fi-Wi reaches the multi-user ceiling; competing Wi-Fi architectures do not, because their topology has already foreclosed it.

K.5 Why the Baseline Design Does Not Pursue This Ceiling

The preceding subsections establish that Fi-Wi's topology is the only Wi-Fi architecture in which joint multi-user decoding is even well-defined. This subsection addresses the complementary question: given that the architecture keeps the door open, why does the baseline design (§9.1, §9.2) deliberately avoid this regime in favor of selection diversity and centralized scheduling? The answer is that joint decoding faces a hard real-time constraint that selection diversity does not, plus a complementary computational cost.

The SIFS deadline is the binding constraint. In 5 GHz and 6 GHz 802.11, SIFS is 16 μs — the interval between the last symbol of the received frame and the first symbol of the response (ACK or block ACK). This budget is hard: a transmitter that does not see the expected response within SIFS treats the frame as lost and retransmits. There is no forgiveness margin and no mechanism to negotiate a longer deadline. The 16 μs budget must accommodate the entire receive-to-response cycle: receive- side decoding, MAC processing, ACK frame construction, the RX/TX switch, and propagation back across the medium.

Joint multi-user decoding adds architectural overhead inside this budget. Samples from M RRHs must arrive at the Concentrator over PCIe-over-Fiber, the Concentrator must perform the W·R combining and the decoder f(·), and the resulting ACK must propagate back through the originating RRH before going on-air. Even at fiber-line speeds, the round-trip Concentrator transit consumes a meaningful fraction of the SIFS budget — order tens of nanoseconds per RRH over short-haul fiber, plus framing and processing overhead — which leaves correspondingly less time for the joint-decoding operation itself. Selection diversity, by contrast, completes the decode at the RRH where the frame arrived; the local decode-and-ACK loop happens entirely on the receive-side chip and the Concentrator participates only in scheduling, above the SIFS cycle rather than inside it.

Computational cost compounds the timing constraint. Optimal joint maximum-likelihood decoding of K simultaneous users has complexity exponential in K. Successive interference cancellation and iterative multi-user detection reduce this in practice, but still require full channel state information from every RRH that received a usable copy, joint estimation of carrier frequency offsets across radios that do not share an RF front-end, and tight sample-level timing alignment. For the small-K case (two to three overlapping transmitters) these techniques are tractable; for the dense-MDU case the baseline design is built for (§7, Appendix J), they are not — and even where they are tractable, fitting the computation inside the SIFS budget is a non-trivial silicon engineering problem.

L4S compatibility is a separable concern, not the binding constraint. The baseline architecture preserves L4S compatibility by emitting packets to the wired path in a manner that matches L4S queue management without disturbing the L4S signal. This is architecturally important — it is what lets Fi-Wi participate in modern low-latency Internet without compromising the L4S signal — but L4S queueing variance operates at the millisecond scale on the IP path, while SIFS operates at the microsecond scale on the link. The two budgets are orders of magnitude apart and apply at different points in the path. SIFS determines whether the link works at all; L4S variance determines how the working link behaves under load. Fitting joint decoding inside SIFS is the architectural problem that determines whether the joint-decoding ceiling is reachable in real time; L4S behavior is a downstream consequence that becomes relevant only after the receive-side decode has succeeded.

The baseline design makes a different choice. Rather than pursuing the information-theoretic ceiling directly, the baseline design uses centralized scheduling (§7, §8, §10) to prevent most of the overlap events that joint decoding would need to resolve, and uses selection diversity (§9.1) to recover from the overlap events that still occur. The three-way framing is:

Approach	Strategy for overlapping transmissions	SIFS-budget compatibility
Information-theoretic optimal	Decode the superposition	Requires fronthaul + compute inside 16 μs
Conventional Wi-Fi	Discard + retry	Local decode fits, but discards information
Fi-Wi baseline	Schedule to avoid; diversity-combine the rest	Local decode fits; scheduling above SIFS cycle

This is a pragmatic choice, not an architectural limit. The topology continues to support the information-theoretic ceiling as a future direction. If silicon evolves to expose the required interfaces and if the SIFS budget can be shown to accommodate fronthaul-mediated joint decoding for some workload class, Fi-Wi can pursue joint decoding selectively — for bulk-throughput traffic classes where SIFS-budget headroom is larger, for example. The baseline design does not depend on that evolution arriving — but the option itself is part of Fi-Wi's architectural value. Competing topologies do not have it to exercise.

K.6 What This Appendix Is and Is Not

This appendix is not a claim about what Fi-Wi implements in the baseline design. It is a claim about what Fi-Wi's architecture makes reachable that competing architectures do not. The baseline design uses selection diversity and per-packet steering — techniques that work with COTS Wi-Fi silicon today and deliver measurable gains over distributed APs on their own merits. Joint multi-user decoding is a research-horizon capability, in the same sense as coherent multi-RRH transmission (§5.7.7): architecturally enabled, silicon-gated, and in this case also gated by the latency/throughput tradeoff it would impose on L4S.

Strategic Observation: The real ceiling of Wi-Fi capacity is not the single-user Shannon limit that conventional MAC design assumes. It is the multi-user capacity that Slepian–Wolf and its successors establish as achievable when synchronized spatial observations can be jointly processed at a central point. The cell-free massive MIMO literature quantifies the architectural delta at roughly 5–10× in per-user throughput over distributed-AP topologies, with the largest gains in the correlated-shadowing regime characteristic of dense MDU deployments. Fi-Wi does not claim that ceiling in the baseline design, but its fronthaul architecture is the physical substrate on which claiming it eventually becomes a silicon problem. For distributed AP architectures it is not even that: their topology forecloses joint decoding regardless of how their silicon evolves, because the aggregate signal is never reconstructable at a single decoding point. The competition is not one chip generation behind on this ceiling. It is on the wrong side of an architectural boundary.

References

1. D. Slepian and J. K. Wolf, "Noiseless coding of correlated information sources," IEEE Transactions on Information Theory, vol. 19, no. 4, pp. 471–480, July 1973.
2. H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta, "Cell-Free Massive MIMO Versus Small Cells," IEEE Transactions on Wireless Communications, vol. 16, no. 3, pp. 1834–1850, March 2017.
3. E. Björnson and L. Sanguinetti, "Making Cell-Free Massive MIMO Competitive With MMSE Processing and Centralized Implementation," IEEE Transactions on Wireless Communications, vol. 19, no. 1, pp. 77–90, January 2020.