OCP microscaling formats

The one-byte element types E2M1, E2M3, and E3M2 used in OCP MX blocks, together with the E8M0 shared scale and contextual block semantics. Standard supplies the six concrete 32-element MX profiles, exact decoding, and destination quantization with proved nearest-even rounding at the selected shared scale. Its dot products accumulate exact rational products across any number of blocks, then round once to binary32.

E8M0

Ordinary binary formats, including IEEE, finite-max-NaN, FNUZ, and fully finite encodings, live in their standard or family packages under FloatLib.Floats.Formats.

This import collects the OCP microscaling concepts:

• E8M0 exponent-only scales, which have no sign bit, zero, or fraction;
• block decoding with one shared E8M0 scale;
• concrete 32-element profiles, including fixed-point INT8 with six fractional bits;
• max-binade scale selection and the SAT/OVF element conversion policies;
• mixed-profile dot products with exact accumulation and one binary32 projection;
• their exact numerical-system contracts and numerics automation; and
• optional #float_info inspection metadata.