Why it matters
Custom high-perf matmul kernels require CUTLASS or equivalent. Understanding enables extreme optimization.
Advertisement
The architecture
Template hierarchy: device → grid → block → warp → mma.
Composability: swap components (epilogue, tile size).
Advertisement
How it works end to end
Epilogues: apply activation or fuse operations at end of matmul.
Precision: FP8, FP16, BF16, INT8, INT4 supported.
Used by: FlashAttention, TensorRT-LLM, custom kernels.