New NVFP4 trick: compressing block scales saves 45-48% per expert matrix, bit-exact
TheZachMueller · x · 2026-10-16
A new NVFP4 technique (shown on Qwen): block scales are 1 byte each, but 99%+ of scales within a row sit within a few E4M3 code steps of each other. Encoding each 16-row slab with a 1-byte row base plus 4-bit offsets—keeping 0.3% exceptions at full quality for 4 extra bits—saves 45-48% storage per expert matrix while remaining bit-exact, touching neither the packed E2M1 weights nor the FP32 global scale.
Related event: NVIDIA's NVFP4 lossless compression trick saves ~6GB VRAM(2 posts)→
More from Infra
- Samsung cuts Q4 phone output as Nittobo shifts E-glass to AI-chip T-glass — zephyr_z9 · 2026-10-16
- White House AI chief discusses Empire AI, New York's shared compute infrastructure — mkratsios47 · 2026-10-16
- llama.cpp merges Muse Spark Open (816B-A42B) text, vision and audio support in 4 hours — jacek2023 · 2026-10-16
- FT warns of 'Darwin moment' for AI borrowers as CoreWeave bonds sink to distress levels — firstadopter · 2026-10-16
- Phone as interface, home servers as the lab: a two-workflow local AI setup — EAccelerate_42 · 2026-10-16
- "A datacenter is a bank": CoreWeave drama revives calls for算力 supervision — arthurcolle · 2026-10-16