[Sionna-RK] Running SRK oai-nr-ue-cuda on x86 + RTX 4090 (UE-only host)

[rcbarke]

Hello,

I’m experimenting with Sionna-RK (SRK) on DGX Spark with USRP B210 SDRs, following the repo and build workflow here:
https://github.com/NVlabs/sionna-rk/tree/main

The standard SRK build produces the oai-nr-ue-cuda container, which provides CUDA-accelerated L1 for the OAI NR-UE. The documentation suggests the NR-UE container can be run on a second host (UE-only), similar to the provided Quectel HW-UE example.

My question is about platform support for a UE-only deployment:

Is it supported (or practical) to run oai-nr-ue-cuda on a separate x86_64 Ubuntu 22.04/24.04 host with an RTX 4090, using the B210 as the RF front end, to keep SRK’s softUE flexibility + CUDA L1 offload?
Or are SRK softUE + USRP deployments currently limited to Jetson/DGX Spark targets, meaning the realistic x86 option today is OAI’s native x86 softUE (without SRK’s CUDA L1 integration)?

If x86 + 4090 is not a supported path today, would you recommend rrunning the UE from a second DGX Spark or SRK-compatible Jetson device to retain the CUDA L1 pipeline?

Best,
Ryan

[SebastianCa]

Hi @rcbarke,
yes, you can run the soft-UE on x86 using the standard OAI configuration. There is limited GPU acceleration on the UE side (the CUDA LDPC decoder can be reused); most of our tutorials focus on the gNB side.

[rcbarke]

Thank you for all of your help, Sebastian! I'll utilize our OAI deployment automation, then see if I can instrument it to deploy the LDPC tutorial. It would be nice to retain offload for the heavier operations, though not a necessity in single UE benchmarking.