# AI-RAN on DGX Spark **URL:** \ **Category:** Aerial Forum\ **Created:** [December 12, 2025, 10:57pm UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502 "2025-12-12T22:57:51Z")\ **Posts on this page:** 6\ **Page:** 1
**Author:** ![rcbarke1](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/rcbarke1/32/353327_2.png) [@rcbarke1](https://forums.developer.nvidia.com/u/rcbarke1)\ **Post date:** [December 12, 2025, 10:57pm UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502/1 "2025-12-12T22:57:51Z")
Hello all, I aim to be precise in sharing this, as it is not a formal resource, though it is pertinent to my PhD research and something I wanted to extend to the community. Clemson University and I are actively exploring DGX Spark’s potential for edge computing applications within several of our research projects. I created the below repository to explore integrating the DGX Spark with the NVIDIA AI-RAN stack. It contains preliminary smoke tests, logs, academic summaries, discovered limits, and known working deployment paths for the AI-RAN stack on DGX Spark (`Sionna`, `Aerial CUDA Accelerated RAN [ACAR]`, `AODT`). You can find my notes here: [https://github.com/rcbarke/ai-ran-dgx-spark](https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Frcbarke%2Fai-ran-dgx-spark&data=05%7C02%7Crcbarke%40clemson.edu%7Cac1e693206da4e54b3e008de39c771bc%7C0c9bf8f6ccad4b87818d49026938aa97%7C0%7C0%7C639011725946683715%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=8hZWSW7F550lEwamkQuh%2BjAl9njxJtZe2Rk7jgZ0naI%3D&reserved=0) My goal in sharing this resource is to aid anyone testing the new platform. I want to set realistic expectations that these are preliminary findings on state-of-the-art hardware, anticipate partial compatibilities. The repository and this post are a comprehensive sweep across the stack, though the two most relevant interests on Spark are `Sionna-RK` and `ACAR` for `O-RAN` and `AI-RAN` studies. **——** **Documentation and Utilities** 1. Created and documented a `first-boot/`process for provisioning and enabling multiple members of my local team to SSH into unique logins. 2. Created several helpful references from NVIDIA’s `DGX Spark` architectural docs, `REMOTE_ACCESS`, and the`THREE_COMPUTER_FRAMEWORK`, with citations and references to source docs. 3. Created the beginnings of use case mappings between similar components, to the best of my knowledge from 6G Developer Sessions, Academic Work, and the Three Computer Framework. 4. Scripted NGC utilities in `first-boot/`, helpful for deploying several AI-RAN resources. **12-12-25 Stack Summary** 1. `Sionna 1.2.1`’s standard installation workflow currently conflicts with the GB10’s base TensorFlow build (reference `Sionna/install_sionna`) 1. Sionna 1.2.1 declares the dependency `tensorflow!=2.16,!=2.17,>=2.14`, so `pip` tries to pull **TF 2.20.0**. 2. The GB10 is only supported today by NVIDIA’s own wheel `nvidia-tensorflow==2.17.0+nv25.2` installed from NGC: `pip install “nvidia-tensorflow[horovod]” –extra-index-url=https://pypi.ngc.nvidia.com/` 3. This can be partially overcome to install `Sionna PHY` and `Sionna SYS`, though `Sionna-RT` fails to install due to a dependency conflict within the `mitsuba` wheel. 4. `Sionna-RK` wasn’t tested given `Sionna-RT`’s installation issues. 5. Replicated the [Sionna LPDC Decoding Tutorial](https://nvlabs.github.io/sionna/rk/tutorials/ldpc_cuda/index.html) and successfully completed a 1,000 run ablation. 2. `AODT` is not currently supported on DGX/ARM, see [this related post](https://forums.developer.nvidia.com/t/aerial-omniverse-digital-twin-installer-403-unauthorized-on-ngc-dell-pathway/354367/2). 1. Refer to the [formal installation documentation and recommended architectures also discussed on 6G Developer Day.](https://docs.nvidia.com/aerial/aerial-dt/index.html) 2. All AODT logs in this repository are shared for experimental purposes only, as I anticipate the intense commercial mathematical computation might require a stronger processor than the`GB10`. 3. `ACAR`partially deploys in pieces, but not enough to connect an end-to-end front haul link presently. 1. My goal was a `PyAerial` + `cuMAC` + `cuBB` + `RU-Emulator` for real-time simulated scheduling study evaluations over a realistic `7.2x FHI` and ray-traced channels. 2. `TEST_VECTOR` generation works, though for ARM devices such as DGX Spark, TVs must be generated on a separate x86 machine due to `MATLAB` compatibilities, then copied over. 3. `PyAerial`won’t deploy due to a hard hardware conformance check in the container. I’m uncertain if the `TensorFlow` builds align. 4. `RU-Emulator` and `TEST_MAC` don’t deploy, because DGX Spark at a platform level does not currently support `GPU Direct RDMA` and therefore only partially supports `DOCA`. [See the DGX Spark FAQ](https://forums.developer.nvidia.com/t/dgx-spark-gb10-faq/347344), though the DGX Spark forums show [preliminary individual experimentation](https://forums.developer.nvidia.com/t/enabling-gpu-direct-rdma-for-dgx-spark-clustering/352051). 5. `cuPHY`’s [SCF L2 Adapter Standalone](https://docs.nvidia.com/aerial/cuda-accelerated-ran/latest/quickstart_guide/running_cubb-end-to-end.html#running-scf-l2-adapter-standalone) works, though `testMAC` fails due to `DOCA`. 6. `cuMAC` can be hacked into deploying (`aerial/acar/CUMAC.md`), though the results aren’t intelligible without the rest of the stack. ---
**Author:** ![jixu](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/jixu/32/136703_2.png) [@jixu](https://forums.developer.nvidia.com/u/jixu)\ **Post date:** [December 16, 2025, 12:40am UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502/2 "2025-12-16T00:40:18Z")
Hi @rcbarke1 Great Notes! Thanks a lot for Sharing your notes in the Forum! ---
**Author:** ![rcbarke1](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/rcbarke1/32/353327_2.png) [@rcbarke1](https://forums.developer.nvidia.com/u/rcbarke1)\ **Post date:** [December 16, 2025, 12:46am UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502/3 "2025-12-16T00:46:34Z")
Hi, you’re welcome. I hope this helps get ahead of inevitable compatibility questions, and look forward to following changes here. I look forward to seeing what DGX Spark can do for AI RAN. ---
**Author:** ![rcbarke1](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/rcbarke1/32/353327_2.png) [@rcbarke1](https://forums.developer.nvidia.com/u/rcbarke1)\ **Post date:** [December 19, 2025, 6:34pm UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502/4 "2025-12-19T18:34:04Z")
**12-19-2025** : `Sionna-RK v1.1.0`, released today, addressing the original `Sionna 1.2.1` installation quirks. The original notes are still relevant for users attempting standalone `Sionna` installations, though SRK now installs a custom kernel for DGX Spark. This resolves the `tensorflow` and `mitsuba` conflicts I documented with the base `DGX OS`, in addition to moving the `OpenAirInterface patch files` for SRK to `OAI 2025w.34`; resolving a `CUDA` dependency conflict originating from a prior 2024 OAI branch. Working with `Sionna` on DGX Spark is now as simple as following the [Sionna Research Kit Quickstart Guide](https://nvlabs.github.io/sionna/rk/quickstart.html). ---
**Author:** ![system](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/system/32/68080_2.png) [@system](https://forums.developer.nvidia.com/u/system)\ **Post date:** [January 2, 2026, 6:34pm UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502/5 "2026-01-02T18:34:48Z")
Closed on January 2, 2026, 6:34pm UTC This topic was automatically closed 14 days after the last reply. New replies are no longer allowed. ---
**Author:** ![nhedberg](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/nhedberg/32/136703_2.png) [@nhedberg](https://forums.developer.nvidia.com/u/nhedberg)\ **Post date:** [January 7, 2026, 11:20am UTC](https://forums.developer.nvidia.com/t/ai-ran-on-dgx-spark/354502/6 "2026-01-07T11:20:19Z")
Hi Ryan, Can you check if a custom kernel was installed on your DGX Spark? The scripts indicate that that step is skipped and when I ran it to confirm I get: ```auto ./scripts/build-custom-kernel.sh This script requires elevated privileges. It will ask for password on the first call to sudo. This is required to install dependencies and the compiled kernel. Use --dry-run to see the operations instead. This script is only supported on Tegra platforms. It is not needed in DGX Spark. Exiting… ./scripts/install-custom-kernel.sh This script requires elevated privileges. It will ask for password on the first call to sudo. This is required to install the compiled kernel and its modules, and to modify the boot sequence. Use --dry-run to see the operations instead. This script is only supported on Tegra platforms. It is not needed in DGX Spark. Exiting… echo “Reboot to load the new kernel and continue the installation.” Reboot to load the new kernel and continue the installation. ``` Thanks, Nick