MXINT8 (OCP MX) | Floating Point Format Guide

Bit Layout

Sign

S

1 bit

Integer

V

1 bit

Fraction

F

6 bits

MXINT8 stores the same eight two's complement bits as INT8, but the binary point sits after the second bit rather than at the end. The result is a fixed-point format covering ±1.984375 in steps of 1⁄64.

Overview

MXINT8 is one of the element types of the OCP Microscaling (MX) Specification v1.0. Thirty-two MXINT8 elements plus one E8M0 shared scale make an MXINT8 block. The elements carry precision near 1.0; the block scale carries the dynamic range.

Because every element in a block sits in a fixed relative range around the shared scale, an element type does not need an exponent field of its own. MXINT8 spends all eight bits on significand, giving it uniform absolute precision across its range — unlike FP8 E4M3, whose precision degrades as magnitude grows.

The scale is implicit, not stored MX Table 6 defines MXINT8 as INT8 bits with an implicit scale of 2^-6. Nothing in the encoding records it: the byte 0x40 is the integer 64 read as INT8 and the value 1.0 read as MXINT8.

Encoding Rules

Every Value

value = twos_complement(bits) / 64

The step between adjacent encodings — the ULP — is exactly 1⁄64 = 0.015625 everywhere, which is what makes MXINT8 a fixed-point rather than a floating-point type.

Special Values (MX Table 6)

Zero: 0 0.000000 = 0. There is exactly one zero, with no signed variant.
Max: 0 1.111111 = +1 63⁄64 = +1.984375, and 1 0.000001 = −1.984375.
Min magnitude: 0 0.000001 = +1⁄64, and 1 1.111111 = −1⁄64.
No Infinity or NaN: there is no encoding to spare, so conversion always saturates.

Symmetric range: −2 is left unused MX §5.3.4 allows the most-negative encoding 0x80 (−2) to go unused so the positive and negative extremes match and no negative bias is introduced. This converter takes that option: encoding never produces 0x80, and −2 saturates to −1.984375. Decoding 0x80 still reports −2, since decoding is total over the whole bit space.

Interactive Value Visualizer

Click any bit to flip it, or enter a decimal or hex value. Note that the signed column reads the bits as MXINT8 while the unsigned column reads them at the same 1⁄64 scale without a sign bit.

Decimal:

Hex:

Dynamic Range & Precision

Property	MXINT8	Unsigned (8-bit, 2^-6 scale)
Total Bits	8	8
Fraction Bits	6	6
Implicit Scale	2^-6	2^-6
Minimum Value	-1.984375	0
Maximum Value	1.984375	3.984375
Representable Values	256	256
Step (ULP)	0.015625	0.015625
Symmetric Range	Yes	N/A

Special Values & Bit Patterns

Value	Bit Pattern	Signed	Unsigned
Zero	00000000	0	0
One (64 raw)	01000000	1	1
Max Unsigned	11111111	-0.015625	3.984375
Max Signed	01111111	1.984375	1.984375
All Ones (-0.015625)	11111111	-0.015625	3.984375
Min Signed (-1.984375)	10000001	-1.984375	2.015625
Sign Bit Only (unused)	10000000	-2	2

Format Comparison

Property	MXINT8	INT8	INT4	INT16
Total Bits	8	8	4	16
Signed Min	-1.984375	-128	-8	-32,768
Signed Max	1.984375	127	7	32,767
Unsigned Max	3.984375	255	15	65,535
Step (ULP)	0.015625	1	1	1
Values	256	256	16	65,536

Where MXINT8 Is Used

OCP Microscaling: the OCP MX Spec v1.0 defines MXINT8 as a concrete MX format: 32-element blocks with an E8M0 shared scale.
AMD Quark: Quark provides OCP_MXINT8Spec for MXINT8 quantization on ROCm.
Microscaling research: the Microscaling Data Formats for Deep Learning paper evaluates MXINT8 against MXFP8 for inference and training.
Plain INT8 comparison: the same bits without an implicit scale are ordinary INT8, which ML pipelines pair with an explicit per-tensor or per-channel scale factor instead of a per-block one.