Core concept
Volta: FP16 tensor cores. Ampere: BF16 + TF32 + INT8. Hopper: FP8. Blackwell: FP4.
Advertisement
How it works
Programmed via WMMA API or auto by cuDNN/cuBLAS. Manual use complex.
Advertisement
Trade-offs + gotchas
All modern ML. cuBLAS + cuDNN + Transformer Engine use automatically.