Thinking Machines details end-to-end 4-bit training: NVFP4 up to 6x faster than BF16 on GB300
rohanpaul_ai · x · 2026-11-14
Thinking Machines Lab published a technical post, "End-to-End 4-Bit Training" by Jack Cook and collaborators, on managing the speed-accuracy tradeoff of low-precision training.
Key points
- Fewer bits cut memory and speed up GPUs, but rounding errors accumulate across layers and optimizer steps and can impair learning.
- Training combines three formats: BF16, MXFP8, and NVFP4.
Hardware gains
- Runs on NVIDIA's Grace Blackwell Ultra (GB300), built to optimize NVFP4 matmuls.
- NVFP4 stores 3.6x fewer bits than BF16 and runs up to 6x faster: 15 PFLOPS for NVFP4 vs 5.5 for MXFP8 and 2.5 for BF16 on GB300.
A rare public look at end-to-end 4-bit training practice, with direct implications for training cost and scaling.
More from Infra
- Nationwide comms failure hits Walmart, Target, Chipotle — gamers just worry about GTA VI — mallow610 · 2026-11-14
- LLM-built regfile compiler achieves much higher logic density — IgorCarron · 2026-11-14
- Compacting at 500k tokens costs a third less than 1M: real usage data on context compression — EricBuess · 2026-11-14
- Musk says SpaceX is already testing space data centers, plans hourly launches — 4KTV · 2026-11-14
- Atreides sends letter to Nvidia board outlining minimum governance standards — GavinSBaker · 2026-11-14
- OSINT deep-dive: FAA nationwide grounding caused by long-haul ATC comms, radar and GPS fine — eigenrobot · 2026-11-14