Testbed Landscape: From LuMaMi to Massive Beams
Why We Build Testbeds
Every massive MIMO result in this book has so far been derived under idealized assumptions. The favorable-propagation theorem assumes perfect channel estimates. The ZF precoder assumes perfect CSI at the transmit side. The capacity scaling proofs assume the noise is white complex Gaussian with a known variance. The cell-free duality assumes a shared time base across access points. Each of these assumptions is a statement about hardware — about what the radio front end, the clock distribution network, and the baseband processor are allowed to not do.
A testbed is an experimental rig where we stop allowing those things. We build an array with real LNAs that drift with temperature, real ADCs with quantization error, real oscillators with phase noise, and real Ethernet links between processors that add latency and jitter. We then try to make the theory work anyway. The gap between what the theory predicts and what the testbed measures is the engineering residual, and narrowing it is the hardest part of massive-MIMO research. Section 26.1 surveys what has actually been built so that later sections can dissect the impairments that drive the residual.
Definition: Massive MIMO Testbed
Massive MIMO Testbed
A massive MIMO testbed is a physical experimental platform with:
-
A base-station array of coherent RF chains — typically — feeding a uniform linear, uniform rectangular, or cylindrical antenna geometry;
-
A set of distinct user-equipments, either real commercial modems or synthetic transmitters that emulate them, that can be served simultaneously on the same time–frequency resource;
-
A real-time processing chain that ingests the ADC samples from every antenna at bandwidth , performs channel estimation and linear detection or precoding, and delivers user data within the slot deadline ;
-
An instrumentation layer that records raw I/Q samples, intermediate signals, and KPIs (throughput, BER, latency, SINR) so that measured performance can be compared against theory.
The combination of all four is what distinguishes a testbed from a laboratory demo. Each subsystem imposes tight constraints which, in the integrated system, cannot be solved independently.
LuMaMi
The Lund University Massive MIMO testbed, built at Lund University in collaboration with the National Instruments Communications Systems Research group. LuMaMi serves 10 mobile users with 100 coherent RF chains at GHz on 20 MHz LTE-TDD, and is the original reference implementation for real-time massive-MIMO processing. The software stack is LabVIEW FPGA on NI USRP-RIO hardware.
Argos / ArgosV3
The Argos family of massive-MIMO testbeds from Rice University's ASH (Argos Self-contained Hybrid) group. Argos was the original 64-antenna demonstration in 2012; ArgosV2 (2014) scaled to 108 antennas with argos-style calibration pilots; ArgosV3 scaled to 96/108 antennas at MHz using Rice's Faros SDR and became the open-source platform that most subsequent academic massive MIMO testbeds build upon. ArgosV3 formalized the argos calibration protocol that we treat in Section 26.3.
OpenAirInterface (OAI)
An open-source implementation of the 3GPP 4G-LTE and 5G-NR stacks maintained by the OpenAirInterface Software Alliance. OAI runs on commodity x86 servers driving SDR front ends (USRP, LimeSDR, AW2S, Benetel) and is the reference platform for experimental 5G NR RAN research. The OAI gNB supports real-time multi-user MIMO up to a few tens of antennas on the USRP N320/X310 class hardware.
srsRAN
An alternative open-source 4G/5G RAN implementation from Software Radio Systems (SRS). Unlike OAI, srsRAN targets a clean modular architecture with a strong focus on 5G NR gNB-CU/DU split and runs on commodity Linux servers. srsRAN is widely used in academic 5G deployments that need a reproducible baseline to benchmark experimental MIMO algorithms against.
Massive Beams
The CommIT group's spin-off startup from TU Berlin, commercializing scalable cell-free massive MIMO in the O-RAN framework. Massive Beams develops a distributed architecture with many low-complexity access points coordinated through a central unit that runs the precoding and scheduling algorithms described in Chapters 11--15 of this book. The company targets private 5G-Advanced and early 6G deployments, bringing the cell-free ideas out of the lab and into industrial trials.
Definition: Real-Time vs Offline Testbeds
Real-Time vs Offline Testbeds
Testbeds divide into two broad categories:
-
Offline testbeds capture raw I/Q samples on all antennas synchronously and dump them to disk. The MIMO processing runs afterwards on a workstation. This is the natural choice for channel-sounding experiments and measurement campaigns (GChannel Sounder). No slot-deadline constraint applies.
-
Real-time testbeds must complete all channel estimation, detection, and precoding within the slot deadline , closing the loop with transmit data in the next slot. This adds a compute-latency budget that offline testbeds do not face.
The compute-latency budget is what turns massive MIMO from a data-processing exercise into a system-engineering problem. Section 26.2 unpacks this budget.
Channel Sounder
A measurement instrument (often a testbed operated in offline mode) designed to capture the physical propagation channel with known reference waveforms. The output is a large database of measured channel responses indexed over time, frequency, and transmitter/receiver location. Sounding drives both the statistical channel models used in simulation (3GPP TR 38.901) and the AI/ML training datasets of Rel-18 and beyond.
Representative Massive MIMO Testbeds
| Testbed | Antennas | Carrier | Bandwidth | Stack | Notable Result |
|---|---|---|---|---|---|
| LuMaMi (Lund) | 100 | 3.7 GHz | 20 MHz | LabVIEW FPGA, USRP-RIO | First real-time 10-user massive MIMO |
| ArgosV3 (Rice) | 96--108 | 2.4 GHz | 20 MHz | Faros SDR, C++ baseband | Open-source calibration and MU-MIMO |
| HHI OpenAirNet | 64 | 3.5 GHz | 100 MHz | FPGA + x86, OAI gNB | 5G NR trials with field users |
| HHI KIARA | 128 (planar) | 26 GHz | 400 MHz | ZCU111, custom RTL | FR2 channel sounding and beam tracking |
| OAI USRP-N310 | 16--32 | 3.5 GHz | 40--80 MHz | Linux, x86, OAI | Reproducible gNB/UE lab setup |
| srsRAN + X310 | 16--64 | 3.5 GHz | 20--50 MHz | Linux, x86, srsRAN | Open baseline for SDR 5G NR |
| Massive Beams (TU Berlin spin-off) | Cell-free, many distributed APs | 3.5--6 GHz | MHz | O-RAN CU/DU, x86+FPGA | Commercial cell-free with reciprocity calibration |
Historical Note: LuMaMi and the First Real-Time Massive MIMO
2012--2015The first end-to-end real-time massive MIMO system was demonstrated by the LuMaMi team at Lund in 2014--2015. The project, led by Fredrik Tufvesson and Ove Edfors and built in close collaboration with National Instruments, used 100 coherent USRP-RIO channels feeding a rectangular patch-antenna array, plus ten mobile user nodes. The LuMaMi paper by Vieira, Malkowsky, Nieman, Miers, Kundargi, Liu, Wong, Öwall, Edfors, and Tufvesson (GlobeCom 2014) established the benchmark numbers that the whole community afterwards tried to match: real-time 10-user spatial multiplexing at a 20 MHz LTE-TDD carrier with ZF precoding computed inside the slot deadline. The subsequent 2017 IEEE Communications Magazine follow-up by Malkowsky et al. reported the full-system measurements and documented the discrepancies between theory and reality — discrepancies that motivate everything in this chapter.
Historical Note: Argos and the Birth of Reciprocity Calibration
2012--2016The Rice Argos project, running in parallel with LuMaMi, contributed the calibration procedure that made reciprocity-based downlink precoding practical. Shepard, Yu, and collaborators recognized early that raw TDD reciprocity is not enough: the uplink channel observed at baseband includes the receive RF chain, while the downlink transmission passes through the transmit RF chain. Their argos calibration protocol (retrospectively named after the testbed) injects pilot tones between reference antennas and solves for a diagonal correction. Concurrently, Rogalin, Bursalioglu, Papadopoulos, and Caire proved that relative calibration suffices — we need only the ratio of Tx and Rx chains, not their absolute values. That result, which we prove in Section 26.3, is the theoretical backbone of every modern reciprocity-based testbed.
Massive Beams: A TU Berlin Spin-Off
Massive Beams is the CommIT group's spin-off startup from TU Berlin, commercializing the cell-free massive MIMO architecture we developed through Chapters 11--15. The company targets private 5G-Advanced campus networks and is an early industrial advocate for bringing cell-free architectures into the O-RAN ecosystem. The technical differentiators are (i) a scalable user-centric clustering algorithm that keeps fronthaul load bounded as the network grows (Chapter 12), (ii) reciprocity-calibrated TDD operation with drift tracking (Section 26.3 below), and (iii) a distributed-SIC uplink receiver that handles the remaining pilot contamination after clustering. The hardware is deliberately low-complexity per AP — a design choice driven by the deployment economics of dense private networks.
OAI vs srsRAN for Academic Massive MIMO Research
Academic groups that want a reproducible 5G NR baseline have two realistic choices: OpenAirInterface and srsRAN. Both are open-source C/C++ implementations running on commodity Linux servers driving USRP-class SDRs. The tradeoffs are practical:
-
OAI has a longer history, supports a wider range of SDR front ends (including older NI USRP lines and the Benetel RRH), and integrates readily with the OAI 5G Core for end-to-end testing. The baseband pipeline is deeply optimized with AVX2/AVX-512 vectorization and can drive a 64-antenna TDD gNB on a single high-end Xeon, at the cost of a steeper learning curve.
-
srsRAN offers a cleaner modular architecture, a more aggressive embrace of the O-RAN CU/DU split, and a documentation and contributor base that is easier for students to join. It is the default choice for new academic 5G NR labs that want to iterate on the physical-layer code quickly without fighting the build system.
For experimental massive MIMO research, the trend since 2023 has been to use srsRAN for the RAN baseline and layer custom massive MIMO baseband code as external DSP modules.
- •
Both stacks currently support at most on a single compute node in real time
- •
Scaling to requires FPGA offload or a distributed CU/DU split
- •
Neither stack ships with a built-in reciprocity calibration subsystem — it must be added as a custom module
Generic Massive MIMO Testbed Block Diagram
Common Mistake: Demo vs Testbed
Mistake:
Equating a working laboratory demo with a full testbed. A demo that runs for thirty seconds with stationary users under controlled lab conditions is not evidence that the algorithms work in the field.
Correction:
A testbed must run continuously under field conditions — moving users, temperature variation, interference from other radios, day/night RF-chain drift — and must include the instrumentation to detect when it is misbehaving. Results from brief lab demos should be reported as "feasibility demonstrations"; only long-duration field measurements are "testbed results".
Common Mistake: The Reality Gap Is a Feature, Not a Bug
Mistake:
Treating the gap between theoretical capacity bounds and measured throughput as an embarrassing shortfall to be explained away.
Correction:
The gap is the most valuable result a testbed produces. It tells the community where the next research problems are. LuMaMi's reports did not match the asymptotic capacity predictions, and that is exactly why we now have the calibration, synchronization, and hardware-aware precoding literature that fills the second half of this book. Report the gap honestly and decompose it into its physical origins.
Key Takeaway
Testbed landscape. The community has converged on a small set of reference platforms — LuMaMi, ArgosV3, HHI testbeds, OAI, srsRAN, and the Massive Beams cell-free spin-off. Each combines an antenna array, a real-time baseband pipeline, and an instrumentation layer, and each reports a gap between theoretical massive MIMO predictions and measured performance. The rest of this chapter decomposes that gap into its four physical origins: fixed-point precision loss (Section 26.2), reciprocity-calibration residual (Section 26.3), synchronization offsets (Section 26.4), and the statistical variance of real channels (Section 26.5).
Quick Check
What is the canonical size of the LuMaMi testbed as reported in the Vieira-Malkowsky-Nieman-Miers 2014 paper?
64 antennas, 8 users, 20 MHz
100 antennas, 10 users, 20 MHz at 3.7 GHz
256 antennas, 16 users, 100 MHz at 28 GHz
128 antennas, 12 users, 40 MHz at 2.4 GHz
LuMaMi uses 100 coherent USRP-RIO RF chains to serve 10 single-antenna users simultaneously on a 20 MHz TDD carrier at 3.7 GHz — the reference numbers for every subsequent massive MIMO benchmarking paper.
Why This Matters: Testbeds Meet O-RAN and 5G-Advanced
Everything in this chapter is, physically, a small-scale O-RAN deployment. The split between the distributed radio units and the centralized processing matches the O-RAN CU/DU/RU decomposition; the reciprocity calibration procedure we study in Section 26.3 is being standardized as an M-plane operation in 3GPP Rel-18; the synchronization requirements of Section 26.4 map directly to the O-RAN fronthaul timing spec. A massive MIMO testbed is, increasingly, the only way to credibly test that an O-RAN deployment will actually meet its KPIs before a carrier commits to rolling it out across a country.