You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
docs/example: wmma:: (mma.sync) on SM 12x — the MMA chapter only covers WGMMA (SM 90) and tcgen05 (SM 100) #1824
The "Matrix Multiply Accelerators" chapter opens with:
cuda-oxide provides access to two generations of matrix accelerators: WGMMA on Hopper (SM 90) and tcgen05 on Blackwell
(SM 100).
A reader on consumer/workstation Blackwell — compute capability 12.0 — concludes that tensor cores are unavailable to them. They
are not: cuda_device::wmma (mma.sync, ldmatrix, movmatrix) works there, and the book contains no wmma:: code at all
(grep -r "wmma::" cuda-oxide-book → 0 hits at 348895e4).
Evidence
A #[kernel] calling wmma::mma_m16n8k16_s32_s8 compiles and runs on compute capability 12.0 (NVIDIA RTX PRO 4500
Blackwell, sm_120, CUDA 13.4), and returns the value an int8 m16n8k16 product implies: with uniform fills of A and B
every output element is 16 * A * B (K = 16), whatever the fragment layout, so the check needs no layout agreement.
wmma is listed "Full" in appendix/supported-features.md.
tcgen05 genuinely is datacenter-only (no TMEM/DSMEM on 12.0) — which is exactly why the mma.sync path needs its own documented
entry rather than being implied by the tcgen05 section.
Ask
Either a wmma::-level example for SM 12x (a 16x8x16 tile, the fragment register layouts, ldmatrix_x1 feeding mma_*), or
one sentence in the chapter saying that consumers on SM 12x use cuda_device::wmma — or cute-rs, whose fmha_*_sm120a_cute
kernels do this — and that tcgen05 does not apply to them.
Environment
348895e4, CUDA toolkit 13.4 (V13.4.92), NVIDIA RTX PRO 4500 Blackwell, compute capability 12.0 (sm_120).
Activity