Repository navigation

perf(ArcKernels): blockwise-FP8 GEMM on tensor cores, scale promoted per 128-K - #200

Merged

heydryft merged 4 commits into

perf/fp8-tensorcore-gemm

Aug 21, 2026

Merged

perf(ArcKernels): blockwise-FP8 GEMM on tensor cores, scale promoted per 128-K#200
heydryft merged 4 commits into
masterfrom
perf/fp8-tensorcore-gemm

Commits on Aug 21, 2026
- perf(ArcKernels): blockwise-FP8 GEMM on tensor cores, scale promoted per 128-K
  
  heydryft
  and
  claude
  committed
  1b937eb
- fix(ArcKernels): WMMA ldm must be a multiple of 32 B, not 16 -- the C++ guide understates the ISA
  
  heydryft
  and
  claude
  committed
  776eafa
- docs(ArcKernels): flag that lowering ARC_FP8_CUBLAS_MIN_M makes this kernel dead code
  
  heydryft
  and
  claude
  committed
  197c847
- docs(ArcKernels): name the occupancy limit to check first, before anyone measures
  
  heydryft
  and
  claude
  committed
  b0ef81d