Benchmark Results

Representative performance notes, not a universal promise

Read this first

Performance depends on GPU model, CUDA version, clocks, thermals, and matrix shape. Treat the numbers below as a reference snapshot that helps explain the optimization ladder, not as a guarantee for every machine.

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	2300	40.2%
Tiled	753	13.1%
Double Buffer	701	12.2%
Bank-Free	673	11.8%
Naive	604	10.6%

What matters more than the exact number

Transition	Main lesson
Naive -> Tiled	Shared-memory reuse matters immediately
Tiled -> Bank-Free	Memory layout details can remove hidden bottlenecks
Bank-Free -> Double Buffer	Overlap and staging help when memory stalls dominate
Double Buffer -> Tensor Core	Specialized hardware changes the ceiling dramatically

Tensor Core note

The benchmark reports:

WMMA end-to-end: includes conversion and fallback handling
WMMA compute-only: 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.

Reproduce on your machine

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

./build/bin/sgemm_benchmark -a
./build/bin/sgemm_benchmark --dims 256 384 640

If you want longer measurements:

./build/bin/sgemm_benchmark -a --warmup 10 --benchmark 50