Benchmark Results

Representative performance notes, not a universal promise

Reference snapshot

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

Kernel	GFLOPS	vs cuBLAS
cuBLAS	5727	100.0%
Tensor Core (WMMA compute-only)	2300	40.2%
Tiled	753	13.1%
Double Buffer	701	12.2%
Bank-Free	673	11.8%
Naive	604	10.6%

Tensor Core note

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

Related references

• Getting Started
• Learning Path
• Kernel progression in README