B200 Compute Capability Is sm_100 | Generated by AI
Question: Is the NVIDIA B200 compute capability sm_80?
Answer:
No. The B200 (Blackwell datacenter) is sm_100 (compute capability 10.0).
Quick mapping of the relevant architectures:
| GPU | Arch | Compute Capability | -arch flag |
|---|---|---|---|
| A100 | Ampere | 8.0 | sm_80 |
| RTX 30xx | Ampere | 8.6 | sm_86 |
| RTX 40xx (your 4070) | Ada Lovelace | 8.9 | sm_89 |
| H100 / H200 | Hopper | 9.0 | sm_90 / sm_90a |
| B100 / B200 | Blackwell | 10.0 | sm_100 / sm_100a |
| RTX 50xx (consumer Blackwell) | Blackwell | 12.0 | sm_120 |
So sm_80 is the A100 (and other GA100-class parts). The B200 needs sm_100, which requires CUDA 12.8+ to compile for.
Note the consumer/datacenter Blackwell split: datacenter B200 is sm_100, but consumer RTX 50-series Blackwell is sm_120 — different feature sets (e.g. the a variants like sm_100a expose arch-specific features like the 5th-gen Tensor Core / tcgen05 instructions used in cutting-edge GEMM kernels).
If you’re compiling for B200 on the MI300X box — you can’t, that’s AMD/ROCm. For the AMD MI300X the equivalent target is gfx942.