NVIDIA NVFP4

NVIDIA NVFP4 is NVIDIA's 4-bit floating-point datatype and low-precision recipe for AI model training and inference, introduced with NVIDIA Blackwell and documented in NVIDIA Transformer Engine. It uses E2M1 4-bit values with hierarchical scaling, including FP8 block scaling and an FP32 global tensor scale, to improve low-precision accuracy and efficiency.