MXFP4
Also known as microscaling FP4, OCP MX FP4
In plain English
MXFP4 is a 4-bit format that gives small groups of numbers their own scale, helping very compact values keep enough useful range.
Technical definition
MXFP4 is a microscaling four-bit floating-point format that shares a scale across small blocks of values.
Engineering details
Block-level scaling gives four-bit values a useful local dynamic range while keeping storage and movement compact. Hardware and software must agree on the block layout, scale representation, and supported matrix kernels.
Why it matters
MXFP4 is used in AMD and cross-vendor low-precision inference paths. Checkpoint preparation and kernel coverage determine the practical result; bit width alone does not capture it.
How to read it in InferenceX
InferenceX records MXFP4 as part of a complete engine and hardware recipe. Comparisons with NVFP4 or FP8 should use the same model, sequence length, quality requirements, and interactivity target.
Source material
See the concept in real benchmarks
AMD MI355X Kimi K2.5 Inference: 7.7x Throughput, Up To 15x Interactivity in 25 Days on vLLM
vLLM PR #35850 Fixed AITER MLA Dispatch on MI355X CDNA4, Unlocking Kimi K2.5 Inference Performance at TP=8, Shipped in vLLM 0.18
InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX
GB300 NVL72, MI355X, B200, H100, Disaggregated Serving, Wide Expert Parallelism, Large Mixture of Experts, SGLang, vLLM, TRTLLM