About

An independent reference, maintained by people who got tired of guessing.

Hardware validation lab bench with rack mounted instruments and dark panels — A hardware validation bench.

What we are

This site exists because the useful information about MXFP4 was scattered across a specification document, a handful of vendor pages, and several thousand forum posts.

MXFP4 Community is an independent, vendor-neutral reference for the MXFP4 microscaling format. We are not affiliated with the Open Compute Project, with any hardware vendor, or with any of the quantization libraries mentioned across the site.

The material here is written by practitioners who quantize models and profile kernels, and it is reviewed by other practitioners before it goes up. Where the group disagrees, the page says so rather than picking a side quietly.

Editorial rules

Three rules govern everything on this site. First: separate what the specification mandates from what a particular library happens to do, and label which is which. Second: any number that is not a measurement is labelled illustrative, on the page, in the caption, every time. Third: where the honest answer is that it depends or that nobody has measured it properly, that is the answer we print.

The benchmarks page is the clearest application of this. Every quality and throughput figure on it is invented to show the shape of a tradeoff, and it says so in three places. We would rather host no numbers than numbers people cite as measurements.

What we will not do

We do not maintain a per-SKU hardware support matrix. Firmware, driver and library revisions gate these features, and a stale matrix does more damage than an absent one.

We do not recommend a specific library as the correct choice. The APIs churn faster than the numerics, and a recommendation written today is a liability in six months.

We do not accept sponsored content, vendor-supplied benchmark numbers, or product placement. If a vendor has results worth publishing, they should publish them with a reproducible method and we will link to it.

Corrections

Corrections are the highest-value contribution to a reference site and we treat them that way. A correction with a reproducible test case gets applied and credited. A correction without one gets a conversation first, which is usually productive anyway.

If a page here contradicts the OCP specification, the specification is right and we are wrong; tell us and we will fix it the same week.

Corrections

Corrections welcome. Test cases more so. If something here is wrong, we would rather hear it from you than have it quoted back at us for a year.

Contact · Community

On the record

2024: microscaling spec published by OCP
4.25: bits per element, the number most people get wrong
0: sponsored pages, now or ever
100%: of illustrative figures labelled as such

Site conventions

Code samples are Python-ish pseudocode, optimised for clarity over speed
Every table caption states whether its contents are arithmetic, illustrative or measured
Implementation-specific behaviour is flagged in a warning box, not buried in prose
Hardware claims stay at the level the vendor documentation actually supports
Nothing on this site is generated without a human reviewing the numbers in it