nGRG OpenLab — Open AI-RAN Research Testbed

Mission

Give AI-RAN research a common base — not a scoreboard.

OpenLab exists so that researchers anywhere can run real Open RAN experiments on a stable, credible, reproducible platform under O-RAN nGRG governance. We deliberately do not rank vendor platforms against each other. What we provide is a neutral stage: open stacks and vendor tooling exercised in the open, with results others can re-run rather than merely read.

Open baseline

OpenAirInterface anchors the lab as the primary research stack — fully open under LGPL/CeCILL, academically reproducible, and vendor-neutral by construction.

A neutral stage for vendors

NVIDIA and Intel each contribute AI-RAN enablement on equal footing. They compete by making research easier — not by being ranked in a report the lab publishes.

Reproducible by default

Every stack is source-built and scripted, from protocol code to channel model. Experiments ship as dataset, model, and code — not as claims.

Flagship Direction

AI-on-Open-RAN

One open, disaggregated, GPU-accelerated platform serving as both the substrate AI runs on and the system AI improves. These three threads run in parallel — they are facets of the same question, not stages of a process.

Thread · AI enhances RAN

Learning inside the radio

Replace or augment RAN functions with learned components — neural receivers, ML-driven scheduling, beam management delivered as xApps over E2.

SionnaPyTorchFlexRICE2SM-KPM / RC

Thread · AI runs on RAN

Sharing the compute

Co-locate edge AI inference with vRAN workloads on the same accelerators, then measure what actually happens to isolation, latency, and scheduling under contention.

CUDAOAI gNBOpen5GSGPU partitioning

Thread · Shared platform

Orchestrating the trade-off

Treat AI and RAN as competing tenants of one compute fabric, and build reproducible benchmarks for the energy-versus-throughput decisions that follow.

5G-LENAns-3energy xApptelemetry

Platform

A full O-RAN stack, open top to bottom.

Control, radio, core, and an AI/channel plane — all software, all source-built, running end-to-end without SDR or over-the-air hardware. Everything below is deployed and verified today.

Layer · ControlO-RAN / RIC

FlexRICnear-RT RIC xAppsKPM · RC · MAC · TC E2 service modelsKPM v3 · RC · E2AP v3

Layer · RAN L1–L3gNB & UE

OpenAirInterfacegNB + nrUE · RFsimbaseline srsRAN ProjectgNB · ZMQ srsRAN 4GsrsUE / eNB / EPC UERANSIMgNB / UE

Layer · Core5G Core & EPC

Open5GS5G SA + EPC · 20 NFs Subscriber DBMongoDB

Plane · AI & ChannelPHY / ML / SysSim

Sionnalink-level PHY + ray tracing PyTorchCUDA · GPU 5G-LENAns-3 · NR system-level MATLAB 5Glink-level cross-check

Layer · FabricCompute

Intel Xeon 6+ NVIDIA RTX 5090 AMD Threadripper+ NVIDIA RTX 5090 NVIDIA CUDAGPU acceleration

The fabric spans both x86 vendors — Intel and AMD — each paired with an NVIDIA GPU, which is what makes the neutral-stage claim structural rather than rhetorical. All radio is software-simulated; nothing here has yet touched an antenna.

Roadmap

From software testbed to live radio to new silicon.

Phase 0 Complete

2026 Q2

Infrastructure, OAI deployment, and the shared CI and registry that make every later result reproducible.

Phase 1 Current

2026 Q3

The triad joined up — OAI on NVIDIA GPU and Intel CPU — plus the first AI-RAN proofs of concept.

Phase 2 Planned

2026 Q4

First live RF via O-RU, and reproducible multi-scenario benchmarks used as method rather than headline.

Phase 3 Planned

2027 H1

ARM, RISC-V, and accelerator cards join the fabric, alongside the first major nGRG research output.

What comes out

Everything the lab produces is meant to be re-run.

ResearchnGRG Research Reports and peer-reviewed papers. The lab is measured by publishable findings, not by utilization.
Reference designsReproducible AI-RAN prototypes shipped as dataset, model, and code — enough for another group to rebuild the result independently.
UpstreamContributions returned to OpenAirInterface, O-RAN SC, and other upstream projects, so the lab strengthens the commons it depends on.
StandardsFindings fed into O-RAN Alliance specification work where the evidence warrants it.

Access

Open application, members first.

Send a short proposalWhat you want to investigate, which parts of the stack you need, and roughly how much compute and for how long.
Scoping conversationWe confirm the experiment fits the platform as it stands today, and agree what reproducible output looks like.
Get your environmentAn isolated tenancy on the fabric, with the stack pre-built and your repository and CI already wired up.
PublishResults land as an nGRG report or paper, with dataset, model, and code released alongside.

Current status

The platform is in Phase 1 build-out. The software stack, shared CI, and compute fabric are running; multi-tenant onboarding is being staged as capacity comes online.

O-RAN Alliance member organisations are prioritised. Academic groups outside the Alliance are welcome to enquire.

Host Tongji University Governance O-RAN nGRG