Analyzing NVIDIA's NVFP4 Pre-training Approach
nrehiew_ · x · 2026-07-13
The post analyzes NVIDIA's NVFP4 pre-training approach, noting that it heavily borrows from previous Nemotron work:
- Hadamard transform: Applied to weight gradient calculations to reduce the impact of outliers.
- Selective high precision: Certain layers (like the final layer) maintain high precision because they require a larger dynamic range and mantissa than FP4.
- Stochastic rounding: Uses stochastic rather than deterministic rounding in gradient calculations to prevent bias.
To validate the approach, the team trained smaller models on up to 16T tokens, showing only about a 0.4% relative training loss gap compared to the BF16 baseline.
Related event: Deep Dive into NVIDIA's NVFP4 Quantization and Pretraining(2 posts)→
More from Infra
- MiniMax H3 runs fully on a Galaxy S26: 5-second video with audio in ~7 minutes — BidFunny4862 · 2026-10-27
- Unconventional AI acquires inference chip firm NeuReality in 1,000x efficiency bid — NaveenGRao · 2026-10-27
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27