HiF8

HiF8 is Huawei’s 8-bit floating-point format that uses a variable-width exponent prefix and denormal encoding to provide a wider dynamic range than FP8 E4M3 without an extra per-value microscale.

Why It Exists

Eight-bit values reduce tensor storage and movement, but a fixed exponent/mantissa split can lose range or precision on neural-network distributions. HiF8 uses the limited bit budget to adapt exponent width across magnitudes.

How It Works

The prefix field indicates how many exponent bits follow and whether the value is denormal. The white paper reports a combined exponent range of [-22, 15], or 38 powers of two, and describes hidden leading bits plus special encodings for zero, NaN, and infinities.

Tradeoffs

HiF8 is a vendor-specific format, so software, kernels, calibration, accumulation, and interoperability must all support its encoding. The white paper reports format properties and peak hardware support, not a complete accuracy study across models.

Common Confusions

HiF8 vs. FP8 E4M3: Both use eight bits, but HiF8 changes how exponent width is encoded instead of using E4M3’s fixed split.
HiF8 vs. MXFP8: HiF8 is a scalar encoding in the paper; MXFP8 uses a block-level shared scale as part of a microscaling representation.

Where It Appears

Ascend 950 NPU Architecture: Compute, Memory, and Unified Bus — Explains HiF8’s exponent encoding and its role in the third-generation Cube datapath.

FP8 — The broader eight-bit floating-point family.
Microscaling — Shared-scale block formats such as MXFP8 and MXFP4.