The include/header files are the most relevant authority. My guess is there is some incorrect info in the docs. You may wish to file a bug to confirm/update.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Ampere single-bit MMA missing from Linux SDK (v11.6)? | 2 | 575 | March 31, 2022 | |
| Implement all supported matrix shapes for wmma::bmma_sync | 5 | 443 | May 6, 2026 | |
| Error or incomprehension, MMa ptx mixed precision Bfloat16 rtx3080 | 19 | 2819 | April 12, 2021 | |
| How to choose from so many `mma.m*n*k*` instructions? | 2 | 614 | June 19, 2024 | |
| Questions about mma instruction with Nvidia ptx | 1 | 313 | July 15, 2024 | |
| PTX instruction `mma` not lowered to tensor core related SASS instruction | 2 | 1534 | March 22, 2022 | |
| Support of binary MMA | 1 | 241 | July 25, 2025 | |
| BMMA on CC 7.2 for hamming distance | 4 | 826 | December 4, 2020 | |
| Direct access to Volta HMMA instruction | 9 | 5579 | December 19, 2017 | |
| The number of IMAD instructions blow up after changing to m16n8k16 mma | 10 | 506 | November 14, 2025 |