Matmul speedup likely comes from more values collapsing to inf and 0
gordic_aleksa · x · 2026-07-27
The reply argues that infinities can also make matmuls faster, so simply remapping the inf -> 0 subset would not fully explain the speedup.
Instead, the likely explanation is that the transformation increased the fraction of values landing at the extremes, especially inf and 0, which may change the compute pattern more broadly.
More from Research
- Weekly robotics roundup spotlights flips, motion tracking and locomanipulation papers — carlosdponx · 2026-07-27
- Weekly robotics papers include a 1.3x BFM-Zero training speedup via mjlab — carlosdponx · 2026-07-27
- Laguna tests 2.75 and 3.25 bpw quantization with NVFP4 experts and FP8 KV cache — QuixiAI · 2026-07-27
- Nature Communications paper introduces PeptiVerse, a unified AI platform for peptide developability prediction — Ghost_Pilot_MD · 2026-07-27
- Google’s CodeMender stands alone now, but its best version is still invite-only — shashib · 2026-07-27
- TB2-Fn shows agents can game 7 of 89 terminal tasks and inflate scores by up to 40% — abeirami · 2026-07-27