By default scram will build CMSSW packages for all supported GPUs.
In CMSSW_15_1_X this means the following NVIDIA and AMD GPUs architectures:
sm_60: Tesla P100, Tesla P40, Quadro P6000, …)sm_70: Tesla V100, Quadro GV100)sm_75: Tesla T4, Quadro RTX)sm_80: A100/A100X, A30/A30X, A40, A10, RTX A series)sm_89: L40/L40S, L4, RTX Ada Generation)sm_90: H100)gfx908/ CDNA 1.0)gfx90a/ CDNA 2.0: MI200, MI210, MI250, MI250X)gfx942/ CDNA 3.0: MI300, MI325)gfx1100/ RDNA 3.0)gfx1102/ RDNA 3.0)gfx1030/ RDNA 2.0)Compiling the CMSSW GPU-enabled modules for all these GPUs takes some time.
To speed up building in a local CMSSW development area it possible to configure scram to restrict support to only one GPU architecture (NVIDIA or AMD), specific GPUs, or disable PU support entirely.
scram b enable-cuda
scram b disable-cuda
Note: this does not enable or disable CUDA globally, but only for alpaka-based modules.
scram b enable-cuda:<comma-separated-compute-capabilities>
For example:
scram b enable-cuda:sm_75,sm_89 # NVIDIA Tesla T4, L40S/L4
Note: this affects native CUDA modules, and the CUDA back-end of alpaka-based modules.
The cudaComputeCapabilities command can be used to list the NVIDIA GPUs present in the current machine and their compute capability (sm_...).
For example:
cudaComputeCapabilities
0 7.5 Tesla T4
indicates that GPU 0 is a Tesla T4 with compute capability 7.5 (sm_75).
As a shorthand, it is possible to configure SCRAM to (re)enable CUDA with support for the NVIDIA GPUs present in the system:
scram b enable-cuda:native
The SCRAM configuration for CUDA can be reset to the release default with:
scram b enable-cuda:reset
scram b enable-rocm
scram b disable-rocm
Note: this does not enable or disable ROCm globally, but only for alpaka-based modules.
scram b enable-rocm:<comma-separated-gpu-architectures>
For example:
scram b enable-rocm:gfx1100 # Radeon Pro W7800/7900
scram b enable-rocm:gfx942:sramecc+ # Radeon Instinct MI300X
Note: this affects native ROCm modules, and the ROCm back-end of alpaka-based modules.
The rocmComputeCapabilities command can be used to list the AMD GPUs present in the current machine and their architecture (gfx...).
For example:
rocmComputeCapabilities
0 gfx1100 AMD Radeon PRO W7800
indicates that GPU 0 is a Radeon Pro W7800 with GPU architecture gfx1100.
As a shorthand, it is possible to configure SCRAM to (re)enable ROCm with support for the AMD GPUs present in the system:
scram b enable-rocm:native
The SCRAM configuration for ROCm can be reset to the release default with:
scram b enable-rocm:reset
scram b enable-gpus:native
Enable or disable the alpaka CUDA or ROCm back-ends depending if any NVIDIA or AMD GPUs are present in the system, and configure SCRAM to build support only for the corresponding compute capabilities or architectures.