From 311d3be18e5b86afab9c7efc962de010f9eef3d6 Mon Sep 17 00:00:00 2001 From: Leopold Talirz Date: Wed, 22 Mar 2023 21:52:53 +0100 Subject: docs: mention cuda multi-arch capability (#36321) --- lib/spack/docs/build_systems/cudapackage.rst | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/lib/spack/docs/build_systems/cudapackage.rst b/lib/spack/docs/build_systems/cudapackage.rst index f586b71952..15fd6d7f02 100644 --- a/lib/spack/docs/build_systems/cudapackage.rst +++ b/lib/spack/docs/build_systems/cudapackage.rst @@ -28,11 +28,14 @@ This package provides the following variants: * **cuda_arch** - This variant supports the optional specification of the architecture. + This variant supports the optional specification of one or multiple architectures. Valid values are maintained in the ``cuda_arch_values`` property and are the numeric character equivalent of the compute capability version (e.g., '10' for version 1.0). Each provided value affects associated ``CUDA`` dependencies and compiler conflicts. + + The variant builds both PTX code for the _virtual_ architecture + (e.g. ``compute_10``) and binary code for the _real_ architecture (e.g. ``sm_10``). GPUs and their compute capability versions are listed at https://developer.nvidia.com/cuda-gpus . -- cgit v1.2.3-60-g2f50