inference/research
verifiedmedium · 0.80updated 2026-10-04

MXFP4 vs NVFP4: block size, scale format, and hardware support comparison

w4a4 (both weights and activations quantized to 4-bit floating point)

Fields

id
quantization/qhw-mxfp4-vs-nvfp4
Scheme
mxfp4
Bits
w4a4 (both weights and activations quantized to 4-bit floating point)
Weight group size
32
Activation scheme
MXFP4: shared power-of-two E8M0 scale per 32-element block. NVFP4: E4M3 fractional scale per 16-element block plus FP32 per-tensor global scale. The key difference: MXFP4 uses power-of-two scales (no fractional scales), NVFP4 uses fractional E4M3 scales. NVIDIA's own comparison shows NVFP4 has lower scale-encoding MSE (0.08 vs 0.72 for E8M0).
Kernels
TensorRT-LLM MXFP4 pluginsCUTLASS MX block-scaled GEMMTensorRT-LLM NVFP4 pluginsCUTLASS block-scaled FP4 GEMM
Quality delta
NVIDIA's own comparison (from the NVFP4 blog) rates MXFP4 as carrying 'risk of noticeable accuracy drop compared to FP8' where NVFP4 does not. The scale-encoding MSE is 0.72 for E8M0 (MXFP4) vs 0.08 for E4M3 (NVFP4). This is a vendor ranking, not a measurement. No independent controlled benchmark comparing MXFP4 and NVFP4 on the same model with the same calibration data was found.
notes
MXFP4 and NVFP4 are both W4A4 formats for Blackwell, but they are NOT interchangeable. The key differences: (1) Block size: MXFP4 uses 32-element blocks, NVFP4 uses 16-element blocks. (2) Scale format: MXFP4 uses power-of-two E8M0 (no fractional scales), NVFP4 uses fractional E4M3. (3) Scale levels: MXFP4 has one level (per-block), NVFP4 has two levels (per-block E4M3 + per-tensor FP32). (4) Standard: MXFP4 is an open OCP standard, NVFP4 is NVIDIA-proprietary. The practical consequence: a model quantized to MXFP4 cannot be served with NVFP4 kernels and vice versa. The block size difference means the quantization error characteristics are different - smaller blocks (NVFP4) generally give better accuracy but require more scale storage. The scale format difference means NVFP4 can represent scales more precisely (E4M3 has ~2 decimal digits vs E8M0's power-of-two steps). NVIDIA's preference for NVFP4 is based on the scale-encoding MSE, but this is a vendor claim. The OCP MX specification is open and has been adopted by multiple vendors, while NVFP4 is NVIDIA-only. For open-source deployment, MXFP4 is the more portable choice.

Notes

MXFP4 and NVFP4 are both W4A4 formats for Blackwell, but they are NOT interchangeable. The key differences: (1) Block size: MXFP4 uses 32-element blocks, NVFP4 uses 16-element blocks. (2) Scale format: MXFP4 uses power-of-two E8M0 (no fractional scales), NVFP4 uses fractional E4M3. (3) Scale levels: MXFP4 has one level (per-block), NVFP4 has two levels (per-block E4M3 + per-tensor FP32). (4) Standard: MXFP4 is an open OCP standard, NVFP4 is NVIDIA-proprietary. The practical consequence: a model quantized to MXFP4 cannot be served with NVFP4 kernels and vice versa. The block size difference means the quantization error characteristics are different - smaller blocks (NVFP4) generally give better accuracy but require more scale storage. The scale format difference means NVFP4 can represent scales more precisely (E4M3 has ~2 decimal digits vs E8M0's power-of-two steps). NVIDIA's preference for NVFP4 is based on the scale-encoding MSE, but this is a vendor claim. The OCP MX specification is open and has been adopted by multiple vendors, while NVFP4 is NVIDIA-only. For open-source deployment, MXFP4 is the more portable choice.

Cross-references

Emulated support (3)

Native support (3)