RAPIDS/cuDF preview build request for Grace-Blackwell GB10 (DGX) — aarch64 kernels missing

Accelerated Computing Confidential Computing
, , ,
1

Case reference: Incident 251120-000525

Hi DGX team,

I ordered my DGX GB10 (Grace-Blackwell Superchip, aarch64) in May 2025. I just received it about 2 weeks ago. The public RAPIDS channels (24.12/nightly) do not contain GB10 kernels, so cuDF is unusable on this machine.

System details:

  • GPU: NVIDIA GB10 (Blackwell Superchip)
  • Driver: 580.95.05
  • CUDA: 13.0
  • OS: Ubuntu 24.04 (aarch64)

I need the internal/preview RAPIDS channel or .tar.bz2 build for Grace-Blackwell to run cuDF natively.

Could you please:

  1. Grant access to the preview channel, or
  2. Provide the private conda URL or build directly?

This is blocking GPU-accelerated dataframe work on a brand-new DGX system.

Thank you — happy to provide any additional details.

Best,
Adam Trina
NVIDIA Enterprise Customer

2

@adam344 you’ll want this in DGX Spark / GB10 User Forum - NVIDIA Developer Forums.

3

Hi Zia / sbellock,

Thank you for the reply.

I already posted in the public DGX Spark forum as requested:
https://forums.developer.nvidia.com/t/…), but I still need the actual preview RAPIDS build.

The public channels do not contain aarch64/GB10 kernels — cuDF is completely broken on my DGX GB10.

Could you please escalate this to the RAPIDS team and provide:

  • The private preview conda channel URL, or
    The .tar.bz2 build with GB10 support

This is a production DGX GB10 system — I need cuDF to work.

Support case: 251120-000525

Thank you,
Adam

4

DGX Spark (GB10) support is already available in public RAPIDS packages with version 25.10. No internal builds or prereleases are needed. You can reference Installation Guide - RAPIDS Docs for installation instructions. There are instructions for conda, pip, and Docker – choose whichever is best for your needs.

For example, the conda installation commands can be used like this:

$ conda create -n rapids-25.10 -c rapidsai -c conda-forge -c nvidia rapids=25.10 python=3.13 'cuda-version=13.0'

Here is a test I ran on a DGX Spark system.

$ conda activate rapids-25.10
$ python
Python 3.13.9 | packaged by conda-forge | (main, Oct 22 2025, 23:14:51) [GCC 14.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import cudf
>>> cudf.Series([1, 2, 3]).sum()
np.int64(10)

If you are having trouble, can you please share the error messages you are observing?