Benchmark Results

Representative performance notes, not a universal promise.

This page is now only the results snapshot surface. For trust policy and interpretation rules, use Validation. For experiment design, use Methodology.

Reference snapshot

End-to-end snapshot

Sample numbers from an RTX 3060 Laptop at 1024 x 1024 x 1024:

Kernel	GFLOPS	vs cuBLAS
cuBLAS	5727	100.0%
Tiled	753	13.1%
Double Buffer	701	12.2%
Bank-Free	673	11.8%
Naive	604	10.6%

Compute-only WMMA snapshot

The repository also reports a narrower fast-path measurement for Tensor Core-friendly shapes:

Kernel	GFLOPS	vs cuBLAS
Tensor Core (WMMA compute-only)	2300	40.2%

The benchmark harness also emits WMMA end-to-end results, but this condensed page does not publish a single headline number for that path because FP32→FP16 conversion and fallback behavior depend strongly on the exact local execution path. Treat the compute-only row as an upper-bound reference, then use Benchmark Scope and local runs to interpret the full end-to-end gap.

How to read this page

• These numbers are a representative local snapshot, not a promise for every GPU.
• Compare them only after reading Benchmark Scope.
• Read the end-to-end table first; treat WMMA compute-only as a narrow fast-path label, not a replacement for end-to-end behavior.
• Assume hosted CI did not prove these numbers. Only local GPU runs can.

Tensor Core note

The benchmark suite reports:

• WMMA end-to-end: the safe FP32 wrapper, including conversion and fallback handling
• WMMA compute-only: the pure pre-converted FP16 path, shown only when M, K, and N are multiples of 16

When dimensions are not Tensor Core friendly, the implementation falls back to a safer FP32 path instead of forcing WMMA.

Read together with

• Validation Overview
• Correctness Policy
• Benchmark Scope
• Reproducibility
• Methodology