Back to AIBriefs
How-ToDevelopersAI Models

NVIDIA NVFP4 recipe boosts pretraining speed on Blackwell

NVIDIA NVFP4 recipe boosts pretraining speed on BlackwellNVIDIA Developer Blog

NVFP4 4-bit mixed-precision training delivers 7x GEMM throughput vs FP8 on Hopper with no accuracy loss. The recipe in MaxText and TransformerEngine enables faster LLM pretraining on Blackwell GPUs.

·
9 days ago
NVIDIA NVFP4 recipe boosts pretraining speed on Blackwell — AIBriefs