GEMM on AMD EPYC and Tenstorrent Wormhole
BF16 GEMM kernel hitting 2,879 GFLOPS (93% of theoretical peak) with AVX-512BF16 intrinsics — 44% faster than AMD's AOCL library and 1st on the course leaderboard. Plus a multi-core port to Tenstorrent Wormhole in TT-Metal.
Read more →