Nvidia’s open 6G tools pass 1 million downloads, with 45.6% less RAN power in a test

Nvidia says its open source 5G and 6G tools have passed 1 million downloads, and that developer adoption of its Aerial and Sionna software has grown tenfold in a year. In a blog post, it also reports four research results built on them.

The company open sourced its Aerial wireless software a year ago, joining Sionna and other tools aimed at 6G research. Nvidia says more than 8,000 research and academic institutions worldwide now use the platforms, which are available on GitHub.

Four university and operator projects, with their numbers

All the figures below come from Nvidia’s post and the projects it describes. None is accompanied by a paper or dataset in the announcement.

Robots. A team at King’s College London linked a robot arm with a camera to an Nvidia Jetson AGX Orin board that runs Sionna for both a software-defined 5G network and edge AI processing. The robot does not stream full images. It sends compact descriptions of objects, where they are and how they relate, an approach called goal-oriented semantic communications. In fruit-packing and cube-stacking tests with deliberately introduced faults, Nvidia says task completion time fell by up to 52.6% compared with sending periodic images, and success improved.

Receivers. SoftBank put a neural receiver into Nvidia’s Aerial CUDA-Accelerated RAN (ACAR) and tested it with a commercial 5G smartphone. With interference present, uplink throughput beat a conventional receiver at all 15 tested settings, by roughly 60% at best. Separately, ETH Zurich researchers used Aerial Testbed and found that tuning neural receivers on local measurements cut block error rates by more than a third compared with selected pretrained versions.

Power. The University of Leeds built a system that forecasts demand 10 minutes ahead and uses the forecast to decide when base stations can sleep. It checks that neighboring stations can cover affected devices before powering one down. On a portable 5G “network-in-a-box” powered by an Nvidia DGX Spark, plus a digital twin, the team reported a 45.6% cut in RAN power use against a full-power baseline while devices stayed connected.

Simulation. CTTC in Catalonia is adding GPU-accelerated radio channel models, running on ACAR, to its open source 5G-LENA simulator. In an early evaluation, runtime dropped from about three hours to 10-20 minutes depending on the integration, with the expected statistical behavior preserved.

Nvidia also points to 6G developer sessions at GTC Washington, D.C.

Who can use this today, and what it replaces

Anyone can download the software from GitHub, so a university lab does not need an Nvidia partnership to reproduce the work. What the tools change is the loop between stages: Nvidia describes them as linking simulation and AI training with experiments on real over-the-air systems, so a neural receiver trained in Sionna can be tried inside Aerial on live radio hardware. The result is faster iteration, but it also ties that workflow to Nvidia GPUs and platforms such as Jetson and DGX Spark, a dependency the post does not discuss.

What the results do not yet establish

The announcement is a curated showcase, so several gaps matter. The 52.6% figure is described as “up to”, and the robot tests used deliberately injected faults in lab tasks, not production warehouses. The SoftBank gain of roughly 60% is a maximum across 15 controlled interference settings, and the post gives no average. The CTTC speedup is labeled early, and the Leeds result is a comparison against a full-power baseline, not against the sleep schemes operators already deploy. Nvidia does not say whether any of the work has been peer reviewed, and the download and adoption counts are its own, with no breakdown of how many users are active or commercial.

There is also a timing question. According to RCR Wireless, citing the June 2026 3GPP RAN plenary, the Release 21 6G work item is due for approval in March 2027, with physical-layer and protocol freezes in late 2028 and a final specification freeze in March 2029. 3GPP has said that Release 20 is the study phase and Release 21 the normative one, per its Release 20 page. So the neural receivers and sleep schedulers in Nvidia’s post are being proven on 5G equipment ahead of a 6G air interface that is not yet defined. Whether such AI components end up in the standard, or stay as vendor-specific implementations, is open.

The robotics result also sits alongside Nvidia’s other work on machine control. Its VFlash research, which we covered earlier, decoded robot trajectories 5.1x faster, addressing the compute side of response time, while the King’s College work targets the network side.

For operators, the Leeds power figure is likely the most commercially relevant claim, since energy is a large share of RAN running cost. Nvidia frames the broader goal as AI-native 6G networks combining connectivity, computing and sensing on one software-defined platform. The evidence so far is a set of promising lab and testbed results rather than field deployments.