What is compute capability in CUDA?

The version number that says which features a GPU has, written 7.5 in docs and sm_75 on the nvcc command line.

It is a property of the silicon, not of anything you install. A Tesla T4 is 7.5 forever; an A100 is 8.0, an RTX 5090 is 12.0. The runtime reads it back as major and minor, the docs write it with a dot, and nvcc wants the dot gone, which is where -arch=sm_75 comes from. Two other numbers get mistaken for it: the toolkit nvcc --version prints, and the highest CUDA runtime your driver accepts, which is what the nvidia-smi header shows. Day 3 prints all three together so you can see which one moved.

The number is a feature gate, and the row matters more than the marketing name of the generation.

Feature	7.5	8.0	8.9	9.0	10.0	12.0
`cp.async`	no	yes	yes	yes	yes	yes
FP8	no	no	yes	yes	yes	yes
Thread block clusters	no	no	no	yes	yes	yes
Distributed shared memory	no	no	no	yes	yes	yes
TMA, single CTA	no	no	no	yes	yes	yes
TMA multicast, in hardware	no	no	no	yes	yes	no
`wgmma`	no	no	no	`sm_90a` only	no	no
tcgen05	no	no	no	no	yes	no

The bold cells are the ones most pages get wrong. A consumer Blackwell card at 12.0 does have clusters, distributed shared memory and single-CTA TMA. It has no wgmma, which the PTX target notes pin to sm_90a and which therefore exists on no Blackwell at all, and no tcgen05. A bigger number is not a superset either: 12.0 is a GeForce card and 10.0 is a B200.

Get the number wrong in either direction and you get one of two failures. Ask for something under your toolkit's floor and nvcc stops before it compiles anything: nvcc fatal : Unsupported gpu architecture 'sm_60', with three spaces before the colon, which matters when you paste it into a search box. Build above your card instead, -arch=sm_80 on a T4, and it compiles, links and starts, then the first launch returns error 209, no kernel image is available for execution on the device. -arch=sm_75 survives newer cards because it also emits compute_75 PTX for the driver to JIT. -arch=native does not: nvcc generates code only for the GPUs it can see, and "no PTX program will be generated for this option".

Measured

On a Tesla T4 (driver 595.84, CUDA 12.6, V12.6.85, built with nvcc -O3 -arch=sm_75), day 3's devicequery read compute capability 7.5 off the device and 750 out of the binary, captured 2026-08-30 on the project's verification node. Same number, two notations: major.minor from cudaGetDeviceProperties, and the three-digit __CUDA_ARCH__ nvcc defines during the device pass. Full run behind How to set up CUDA.

Code

The card's number is easy to read. The binary's number only exists inside device code, so a kernel has to write it out:

__global__ void reportCompiledArch(int* out, size_t n) {
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i < n) {
#if defined(__CUDA_ARCH__)
        out[i] = __CUDA_ARCH__;
#else
        out[i] = 0;
#endif
    }
}

The #else branch is not dead code. The same file is compiled once for the host, where __CUDA_ARCH__ is undefined.

Where you meet this

Day 3, how to set up CUDA, reads it off your card and picks the flag. Day 0 uses it to choose hardware, day 69 revisits it when one binary must serve several cards, and day 78 lands on the corrected row above. Both failures have pages: nvcc fatal: Unsupported gpu architecture and no kernel image is available for execution on the device.

Sources

Feature support per compute capability, Table 29: https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/compute-capabilities.html (checked 2026-08-29)
-arch shorthand, the sm_75 default and native: https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#gpu-architecture-arch (checked 2026-08-30)
Which capability your card has: https://developer.nvidia.com/cuda-gpus (checked 2026-08-29)
wgmma requires sm_90a: https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#asynchronous-warpgroup-level-matrix-instructions-wgmma-mma (checked 2026-08-29)
No multicast on GeForce, so the cluster shape is fixed to 1x1x1: https://github.com/NVIDIA/cutlass/blob/main/media/docs/cpp/blackwell_functionality.md#cluster-size (checked 2026-08-29)

Byline