Is GROMACS currently unavailable on the DGX B200 with CUDA 12.8 (Ubuntu 24.04)?
I am encountering the following error:
WARNING: An error occurred while sanity checking device #0. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #1. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #2. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #3. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #4. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #5. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #6. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
WARNING: An error occurred while sanity checking device #7. An unhandled error from a previous CUDA operation was detected. CUDA error #209 (cudaErrorNoKernelImageForDevice): no kernel image is available for execution on the device.
I’d be grateful for any feedback or suggestions.
Thank you in advance.
GROMACS 2024 should be compatible with newer hardware but was not tested on NVIDIA B200. Could you share the output of gmx -version and nvidia-smi?
Also, keep in mind that this Friday, we’re planning to make the GROMACS 2025.0 release, which was tested on Blackwell devices and has some nice performance tweaks for this hardware. You can try the 2025 release candidate in the mean time; feedback would be appreciated :)
Building on this, I have a 5090 which should also be a Blackwell, and I got the same error as the original post, so I arrived here. I used the suggested flags but in the make I get this error
CMake Error at libgromacs_generated_nbnxm_cuda_kernel_pruneonly.cu.o.Release.cmake:280
both with CUDA12.8 and 12.9. Any idea what it might be? Because I would probably need to install both GROMACS version 23.0 and 24.5. On a side note, I installed GROMACS 25.1 and it worked out of the box, no additional flags!
Currently, CC 12.0 is not supported by GROMACS 2024-series. The patch is trivial (src/gromacs/gpu_utils/cuda_arch_utils.cuh part of MR !4933); it will be in GROMACS 2024.6 (just like the fix for the issues above), which we plan to release eventually.
For now, on RTX 50xx, please use GROMACS 2025-series. Applying the patch above is at your own risk :)