Skip to content

coopvec-fp8-with-non-optimal-layout

Status: stub. The full-length analysis is queued for a v1.0.x patch release per ADR 0018, section 5, criterion #6. The companion rule page at docs/rules/coopvec-fp8-with-non-optimal-layout.md contains the canonical detection logic + GPU reasoning.

TL;DR

FP8 (E4M3 and E5M2) is the SM 6.9 cooperative-vector path's lowest-precision data type and the one with the highest throughput on the tensor engines. NVIDIA Ada Lovelace's tensor cores execute FP8 matmul at roughly 2x the FP16 throughput; AMD RDNA 3/4's WMMA-FP8 path matches; Intel Xe-HPG's XMX engines have the same approximate ratio. The throughput advantage requires the engine's native FP8 storage layout — the hardware fetcher assumes the bytes are pre-arranged in the swizzle pattern that lets one fetch deliver one tensor-core operand.

What the rule fires on

A cooperative-vector matrix multiply whose interpretation enum names an FP8 component type (COMPONENT_TYPE_FLOAT_E4M3, COMPONENT_TYPE_FLOAT_E5M2) and whose matrix-layout enum is not one of the optimal layouts (MATRIX_LAYOUT_INFERENCING_OPTIMAL / MATRIX_LAYOUT_TRAINING_OPTIMAL). The SM 6.9 cooperative-vector specification mandates an optimal layout for FP8 matrices: the tensor-engine's FP8 path on every IHV requires the vendor-swizzle layout to function correctly. Generic row-major / column-major FP8 is undefined behaviour.

See the What it detects section of the rule page for the full pattern definition.

Why it matters

The full GPU-mechanism analysis lives in the Why it matters on a GPU section of the companion rule page.

Examples

The bad / good code snippets are kept canonical on the rule page; see coopvec-fp8-with-non-optimal-layout.md -> Examples.

See also


This is a v1.0-ship stub. Full analysis pending; track issue link TBD.

© 2026 NelCit — Apache-2.0 (code), CC-BY-4.0 (docs).