Counterintuitive NVFP4 Kernel Optimization: More Bits, Up to 10% Faster

The Humansand team optimized a highly tuned NVFP4 decode kernel using more bits for higher precision, yet achieved up to 10% faster generation—defying the usual speed-accuracy tradeoff of low-bit inference.

2026-10-24 ~ 2026-10-24 · 3 related posts

• Humansand Speeds Up NVFP4 Decode Kernel by Up to 10% While Improving Accuracy — noahdgoodman · 2026-10-24
• NVFP4 decode kernel gets more bits and 10% faster generation at once — gharik · 2026-10-24

1 near-duplicate retellings: gharik