CUDA compute capability

CUDA compute capability is the major.minor version number NVIDIA gives each GPU to define the hardware features and instructions it supports. A GPU of compute capability 12.0 has major version 12 and minor version 0, and its streaming multiprocessors carry the version sm_120. That sm_XX label names the build target passed to the nvcc compiler and tags the binary it produces.

NVIDIA's CUDA GPUs page lists the A100 at 8.0, the RTX A6000 at 8.6, the RTX 6000 Ada at 8.9 and the H100 at 9.0. The B200 is 10.0, the B300 is 10.3, and the GeForce RTX 5090 and RTX PRO 6000 Blackwell are 12.0. A binary built for 8.6 loads on 8.6 and 8.9 GPUs but not on 8.0, and no binary loads across major versions. A B200 binary therefore will not load on an RTX PRO 6000 Blackwell. CUDA 13.0 removed offline compilation for compute capability below 7.5, so software for older GPUs must be built with CUDA 12.9 or earlier.

Before buying, check that prebuilt software lists the card's sm_XX target. If it does not, the build needs embedded Parallel Thread Execution (PTX) code that compiles at run time, or a rebuild from source.

Sources

Last verified October 7, 2026.