Does PyTorch Already Support the NVIDIA GB200 GPU?

1

Hi,

Does PyTorch already support the NVIDIA GB200 GPU? I couldn’t find any binary that mentions support for it.

If not, are there any workarounds, such as building PyTorch from source to enable compatibility?

Thanks!

2

GB200 isn’t released yet so could you describe how a workaround would be used?

3

Hi,

Thanks for pointing that out! I actually have access to a GB200 and need to conduct some workload research on it.

I’m exploring how to get PyTorch to work with this hardware. Specifically:

  1. Would building PyTorch from source with a specific CUDA version be a viable option, assuming the appropriate driver and CUDA are installed?
  2. Since the GB200 might share similarities with the GH200, could it be possible to configure PyTorch to treat it as a GH200 for compatibility?
  3. Are there general guidelines for enabling GPU support for new hardware in PyTorch?

Any advice or suggestions on how to proceed would be greatly appreciated.

Thanks in advance!

4

I would recommend to get in touch with your NVIDIA contact to access software stack for hardware preview which you apparently received.
Official support for GB200 will land after the release of the platform and SW stack.

5

Thank you so much for your thoughtful response and helpful recommendations. I truly appreciate your help!