Pearl-GEMM is claimed to be about 5% faster than Quack on B200 chips
A former Nvidia researcher says Pearl-GEMM also runs about 42% faster than Flashinfer on B200s, calling Pearl's FP8 mixture-of-experts kernels the fastest for Blackwell.
TLDR
A former Nvidia researcher claims Pearl has the fastest FP8 kernels for mixture-of-experts AI models on Blackwell chips. The post puts Pearl-GEMM about 5% ahead of Quack and about 42% ahead of Flashinfer on B200s. It describes grouped matrix multiplication as one of AI's most optimized operations and says pushing its performance further is difficult.
Combined views
34.8K
2 Sources, first seen ago