Introducing NVFP4 for Efficient and Accurate Low-Precision Inference

NVIDIA blog introducing NVFP4. Key facts: NVFP4 = E2M1 (4-bit) + E4M3 FP8 scale per 16-value micro-block + per-tensor FP32 scale. Two-level micro-block scaling. Comparison table: FP4 (E2M1) has no hardware scaling, MXFP4 has 1 shared power-of-two scale per 32-value block, NVFP4 has 1 shared FP8 scale per 16-value block. NVFP4 has hardware-accelerated scaling. Memory: ~3.5x less than FP16, ~1.8x less than FP8. Accuracy: 1% or less degradation vs FP8 on DeepSeek-R1-0528. Energy efficiency: up to 25x (Blackwell) and 50x (Blackwell Ultra) vs H100. Figure 1 shows peak low-precision performance: B200 9/18 PFLOPS dense/sparse, B300 13/18, GB200 10/20, GB300 15/20.