Skip to content

Add GPU architecture BLACKWELL#488

Merged
hyschive merged 10 commits intogamer-project:mainfrom
koarakawaii:main-GPU-micro-architecture
Feb 13, 2026
Merged

Add GPU architecture BLACKWELL#488
hyschive merged 10 commits intogamer-project:mainfrom
koarakawaii:main-GPU-micro-architecture

Conversation

@koarakawaii
Copy link
Contributor

@koarakawaii koarakawaii commented Nov 9, 2025

  • Some properties for cudaDeviceProp are deprecated in CUDA 13.0; I did not check all of them, but only modified those affecting the compilation. See https://docs.nvidia.com/cuda/cuda-runtime-api/structcudaDeviceProp.html#structcudaDeviceProp for details.
    • Can consider switching the cudaDeviceGetAttribute API; but again, I did not check whether it covers all the properties we queried for cudaDeviceProp
  • Support latest cufftDx
    • Modify macro NONE to OPTION_NONE in GAMER to be compatible with cufftDx's .hpp file.
  • Add configure file spock_intel_RTX5090.config for RTX5090 on spock26 and GB10 for gx10, which uses private NVIDIA HPC SDK modules instead of public CUDA on the log-in node. The path should be directing to public CUDA 13.0 when available, if we want to install it.
    • CUDA 12.1 also works for current driver version 580.95.05, but I am not sure we want to fall back to CUDA 12.1 if a more advanced version has been released.
    • cufftDx on gx10 is under private path.

@hyschive hyschive requested a review from xuanweishan November 15, 2025 13:57
@hyschive hyschive added enhancement gpu GPU issues and implementation labels Nov 15, 2025
@xuanweishan
Copy link
Contributor

@koarakawaii Do you think we need to install cuda13 and nvidia hpc to spock:/software? So that the config no longer need to depend to your own file.

@koarakawaii
Copy link
Contributor Author

@xuanweishan : I think that will be great. I guess we will need it in the future anyway~

@xuanweishan
Copy link
Contributor

It seems good to me.
I tried tests on both eureka and spock:

  • eureka 2080super(cuda12)
  • spock 3080ti(cuda12)
  • spock 5090(cuda13)
  • spock 3080ti(cuda13)(With modify config file for 3080 ti)
    All works properly.
    I think this PR is ready to merge after resolved the conflicts.

@koarakawaii
Copy link
Contributor Author

@xuanweishan : conflicts have been solved~

Copy link
Contributor

@xuanweishan xuanweishan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@koarakawaii
I've test it with following settings

  • eureka
    • cuda11
  • spock
    • cuda12
    • cuda13 on ordinary node
    • cuda13 on RTX5090 node
  • GX10
    • cuda13
      It would be ready after update the path.

@koarakawaii
Copy link
Contributor Author

koarakawaii commented Feb 11, 2026

@xuanweishan Thank for all the testing~ I have updated the cufftDx path to the public one.

Copy link
Contributor

@xuanweishan xuanweishan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@koarakawaii Thanks for the update
@hyschive This PR is ready to go.

Copy link
Contributor

@hyschive hyschive left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@koarakawaii @xuanweishan Looks good. Thanks for the contributions and review!

@hyschive hyschive merged commit 8b15bce into gamer-project:main Feb 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement gpu GPU issues and implementation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants