Run NVIDIA's Sionna Bench (channel-estimation benchmarks built on Sionna and PyTorch) on a GH200 from the shared content, and find the published results of a full reproduction on COSMOS.
Sionna Bench compares channel estimators, from classical LS and LMMSE to published neural estimators, on standard 5G channel models such as 3GPP UMi. It runs on Sionna and PyTorch.
On COSMOS it is shared content: a ready Python environment (Sionna 2.1.0, Sionna RT 2.2.0, PyTorch 2.11.0 for CUDA 13.0) and the Sionna Bench source, published read-only at /opt/content/sionna-bench. You load the GPU-compute base, source one file and run. The full reproduction of the paper's results ran on srv8.rrail in November 2026, and its results are published next to the environment.
After completing this tutorial you will be able to:
| Difficulty | Beginner |
|---|---|
| Estimated time | 15 min (the full reproduction: about 3 days) |
| Domain / sandbox | rrail (srv8.rrail or srv9.rrail) |
| Image | gpu-compute-gh200.ndz |
| Transmits? | no |
| Last verified | 2026-11-14, gpu-compute-gh200-20261114 and content sionna-bench 20261114 on srv8.rrail |
Background knowledge
Account & access
Devices / nodes
| Resource | Role | Qty | Notes |
|---|---|---|---|
srv8.rrail or srv9.rrail |
runs Sionna Bench | 1 | NVIDIA GH200 480GB (72 Arm cores, 96 GB GPU memory) |
Disk images
| Image | Load onto | Provides |
|---|---|---|
gpu-compute-gh200.ndz |
the GH200 server | driver, Docker, and the shared-content tools |
Software components
| Component | Version | Source |
|---|---|---|
| Sionna Bench | 0.1.0, commit e1d2fdb |
shared content sionna-bench 20261114 |
| Sionna / Sionna RT | 2.1.0 / 2.2.0 | same |
| PyTorch | 2.11.0+cu130 | same |
| Python | 3.12 | the image |
One server, no radio. The environment is read from the network share, where it was measured to run as fast as from a local copy (cold start 6.4 s vs. 7.0 s, test suite 112 s vs. 111 s), so nothing is copied at boot.
cosmos-content enable sionna-bench
cosmos-content sync
enabled sionna-bench
sionna-bench nfs (nothing marked copy)
source /opt/content/sionna-bench/env.sh
echo $SB_WORK
python -c 'import torch; from importlib.metadata import version as v; print(torch.__version__, torch.cuda.get_device_name(0), v("sionna"), v("sionna-rt"))'
/mnt/datadisk/work/root/sionna-bench
2.11.0+cu130 NVIDIA GH200 480GB 2.1.0 2.2.0
env.sh puts the shared environment on your PATH and sends every cache and output to $SB_WORK in your work area, because the content itself is read-only.
The fast tests cover every estimator and channel model without training:
cd $SB/src
pytest -q -p no:cacheprovider -m "not slow and not training and not system_variants" --basetemp=$SB_WORK/pytest
...
267 passed, 159 deselected, 14 warnings in 103.45s (0:01:43)
reproduce_results.py --all reruns every stage of the benchmark, including training the neural estimators. On srv8.rrail it ran from 2026-11-10 01:10Z to 2026-11-13 18:31Z, resumed once after the server lost power on 2026-11-13. It is CPU-bound: it keeps about one core and 22 % of the GPU busy. Its results are published, sealed with checksums:
ls /opt/aerial-data/sionna-bench/results/repro-20261113/out-repro/
ls /opt/aerial-data/sionna-bench/results/repro-20261113/out-repro/builtin/3gpp-umi/
balevi2020 baur2022 builtin coleri2002 delfeld2025 edfors1996 koller2022 li2020 lyu2026 melgar2022 wei2017
config.json LMMSE.npz LS (linear).npz LS (nearest neighbor).npz results.json
One directory per method (builtin holds the classical estimators), one per channel model below it. config.json describes the system (52 PRB, 15 kHz, 4 uplink users), and results.json lists each estimator with its settings and the .npz file that holds its results. MANIFEST.json in the release names the source commit, the packages and the GPU and driver it ran on.
To rerun one paper's stage yourself, write everything to your work area. The classical estimators (builtin) need no training:
cd $SB/src
python scripts/mimo_ofdm_chest/reproduce_results.py --paper builtin --prepare-assets \
--asset-root $SB_WORK/assets --output-dir $SB_WORK/out
It takes about 16 minutes (948 s on 2026-11-14) and ends with lines like:
Eb/N0 15.0 dB: BER=2.930369e-04, BLER=2.976190e-03, NMSE=5.325178e-04
[builtin] result saved to /mnt/datadisk/work/root/sionna-bench/out/builtin/3gpp-umi/LMMSE.npz
--help lists the papers. --all runs every stage; the neural estimators train for hours each.
The environment was built at its final path, /opt/content/sionna-bench, from the exact package list of the published reproduction (requirements.txt in the release). That path points to the network share; see how the GH200 content works.
PyTorch reports CUDA 13.0 on the GH200, and the test suite ends with 267 passed.
cd /opt/aerial-data/sionna-bench/20261114 && sha256sum -c --quiet SHA256SUMS && echo OK
OK
Copy what you want to keep from $SB_WORK off the server, then remove it (rm -rf $SB_WORK). Nothing runs in the background.
| Symptom | Cause | Fix |
|---|---|---|
Read-only file system |
something wrote into /opt/content/sionna-bench |
source env.sh first; write outputs under $SB_WORK |
cosmos-content enable says no published release |
the share is not mounted | ls /opt/aerial-data, then retry |
torch.cuda.is_available() is False |
not on gpu-compute-gh200 |
load the GPU-compute base |
Group: GPU computing · Image: gpu-compute-gh200.ndz · Content: sionna-bench 20261114 · Last verified 2026-11-14 on srv8.rrail