Umber Networks reference drawings for the rig: the traffic ground truth measurement
plane, the programmable 2×2 MIMO signal path with AP/RRH selection and the
calibrated transfer-matrix model, and the MAC load generator that produces the
contending 802.11 load.
Traffic Ground Truth
This drawing shows the measurement plane of the rig: the two iperf2 traffic endpoints, the
System A and System B devices under test, and the conducted channel between them. Both
endpoints sit on a PTP common time base carried on a dedicated sync network, so one-way
delay is a directly measured quantity. The conducted channel and the MAC-load injection appear as single blocks; their
RF construction and calibration are specified in the reference drawing below.
Umber NetworksBuilt by Umber and VaunixProgrammable 2×2 MIMO Signal PathReference drawingSystem A AP: MikroTik C53UiG+5HPaxD2HPaxDSeptember 2026
MAC Load Generator
This drawing shows how the contending 802.11 load is produced: the Ramsey STE5125M shielded
enclosure holding the 45-node ESP32 farm on three levels, the Cambrionix hubs with per-port
power control, the RPi5 farm host and internal Ethernet switch, and the three hanging
dipoles. Dipoles 1 and 2 run on interior jumpers to their bulkheads, and the injected load
level is set by the Vaunix attenuator outside the shield; dipole 3, mounted on a fixed
interior frame, runs direct to its bulkhead as the fixed reference path. The external jumpers land
on the MAC-load combiner inputs shown in the reference drawing.
2×2 MIMO / Two Spatial Streams
The four independently programmable Attij / φij paths implement a full
2×2 complex channel matrix. The rig therefore supports two MIMO spatial streams when the
programmed channel matrix is full rank and sufficiently well conditioned.
A simple two-stream validation setting is a near-diagonal channel:
keep Att11/φ11 and Att22/φ22 strong while placing the cross paths
Att12/φ12 and Att21/φ21 substantially lower. As the programmed matrix becomes
ill-conditioned or rank-deficient, the receiver may fall back from two spatial streams
to one.
Installed System A AP
MikroTik C53UiG+5HPaxD2HPaxD
IPQ-6010 · 2.4/5 GHz · 802.11a/n/ac/ax · 1200 Mbit/s · USB 3.0 Type A
Product/listing reference: W127080453
This is the conventional AP connected to the System A side of the two SPDT switches.
Its two RF chains connect to the conducted Ant 1 / Ant 2 paths shown in the reference drawing.
Single-path RF loss calculator
Calculates one physical Attij / φij path from the Ant1/Ant2 source input to the selected
System A AP port or to one System B RRH branch. It includes the theoretical divider/
combiner split loss, excess insertion loss, the programmable attenuator setting,
phase-shifter insertion loss, SPDT loss, and optional cable/connector loss.
Source → System A AP
0.0 dB
Estimated output: 0.0 dBm
Source → one System B RRH
0.0 dB
Estimated output: 0.0 dBm
Component assumptions / edit calculation
Editable conservative default; published range is 0.5–1.5 dB.
Default from Vaunix product-guide value; replace with your calibration.
Public LPS-802 typical value; verify the LPS-802-4 build.
Defaults are engineering starting points, not a substitute for through-calibration.
In particular, replace component values with measured insertion loss at the operating
frequency when the rig is assembled.
This is a scalar loss calculation for one Attij / φij path. When multiple channel-matrix paths
are present simultaneously, their complex amplitudes add at the 4-way combiner, so the
actual r1/r2 level depends on H(f), phase, and the active sources—not just the individual
path-loss sums.
What this rig can and cannot validate
The drawings above specify what the rig is. This specifies what a result from it is evidence for, so that a number measured on one band is never quietly promoted into a claim about another.
What the two systems are
System A is a conventional access point, with its own queues, EDCA and rate control. System B is a UAX-8, an Umber Airtime Switch with eight radio heads: one concentrator holding the packets and every scheduling decision, reaching the heads over fronthaul. The drawings label it UAX-8 for short.
The bands are the binding limit
The System A access point is a 2.4 and 5 GHz device, and the forty-five load nodes are ESP32-C5 parts, which are 2.4 and 5 GHz with no 6 GHz radio at all. Their 802.11ax operation is 20 MHz only; 40 MHz is available to them for legacy 802.11n. So the MAC load this rig injects is a 2.4 and 5 GHz load at 20 MHz, and nothing measured here is direct evidence about 6 GHz behaviour or about high-efficiency contention at 40, 80 or 160 MHz.
What a result here is evidence for
Measured directly on this rig
Claim
On what basis
One-way delay and its distribution under contention, against a real air interface
Directly measured, both endpoints on a common time base
Grant-to-air latency against known, repeatable MAC load
At 2.4 and 5 GHz, with a 20 MHz load
Behaviour of one spatial channel under a programmable 2x2 matrix, including rank
That a controlled H sustains one stream or two, and how that tracks conditioning
What it is not evidence for
Belongs to another rig, or to a building
Claim
Where it has to come from
What fraction of links in a real building are rank two
Deployment survey. A conducted matrix cannot supply the distribution
6 GHz contention, channel plan or DFS-free behaviour
No 6 GHz radio in the load farm or in System A
High-efficiency contention at 40, 80, 160 or 320 MHz
Load nodes are 20 MHz for 802.11ax
Ninety-six head capacity, memory load or failure behaviour at scale
The concentrator load rig
The isolation matrix of a building
Conducted channel, not a building
The Sonde is characterised against the receiver
Not against a transmitter a metre away. Every Sonde test records the same row: how many transmitters were active, what the intended receiver ground truth was, what the reference capture saw, where the Sonde sat relative to that receiver, and whether the Sonde detected, decoded and returned correct metadata. Each row also carries band, channel width, PHY family, MCS, spatial streams, guard interval and the H-matrix setting, because a detection result without those is not a result.
PTP is a common time base, not a zero
One-way delay is only as good as the residual offset and jitter between the two endpoints, so that residual is measured rather than assumed and is carried alongside every one-way figure the rig publishes. A one-way delay quoted without its clock uncertainty is a number with no error bar, which is the thing the ground truth paper argues against on every other axis.
The same measurement from a phone
The rig's traffic plane is two iperf2 endpoints, and the same tool runs on Android, which is how a reader takes a lighter version of these measurements without a rig: one-way delay against a stated clock, bounceback under working load, and the L4S marking checks. It is not a substitute for the conducted channel or the injected MAC load, and nothing measured on a handset carries the repeatability the rows above depend on. It is the honest small version of the same discipline.
Get iperf 2 on Google Play