Measured benchmark: RTX 3060 Laptop, 2026-08-17

This is the only benchmark page in the repository that reports repository-measured numbers. All other performance figures in teaching documents are placeholders unless they link here.

Environment

Item	Value
GPU	NVIDIA GeForce RTX 3060 Laptop GPU, 6 GB, sm_86
Driver / nvidia-smi	591.44
CUDA toolkit	12.0.140 (nvcc)
Host	Ubuntu 24.04, GCC 13.3
Build	Release, --use_fast_math, -lineinfo, -arch=sm_86
WSL note	LD_PRELOAD=/usr/lib/wsl/lib/libcuda.so.1 was required on this WSL2 host

01-sgemm-tutorial ladder

Command:

cd 01-sgemm-tutorial
make benchmark
./build/sgemm_benchmark -a

Warmup: 5 runs. Timed runs: 20 per kernel, averaged with CUDA events.

Kernel	512^3 GFLOPS	1024^3 GFLOPS	2048^3 GFLOPS	4096^3 GFLOPS
cuBLAS	4242	5581	6390	5695
Naive	597	583	674	666
Tiled 32x32	665	925	933	910
Bank Conflict Free	505	657	667	655
Double Buffer	496	678	683	667
Tensor Core WMMA	1626	1093	2768	4099

Key finding on this GPU: the simple tiled kernel is faster than both bank-conflict-free and double-buffer variants at 1024^3. Micro-optimizations must be profiled on the target hardware; they do not always improve.

WMMA correctness uses an FP16-input quantization-aware tolerance (atol = 0.0015 * sqrt(K)), because fixed small tolerances fail at K=2048+ even though the kernel is working as designed.

02-tensorcraft-core GEMM variants

Command:

cmake --preset default
cmake --build --preset default --target gemm_benchmark
./build/default/bin/gemm_benchmark --benchmark_min_time=0.1s --benchmark_repetitions=3

Median wall time, 3 repetitions:

Shape	Naive GFLOPS	Tiled GFLOPS	Double Buffer GFLOPS
256^3	257	329	319
512^3	634	767	861
1024^3	641	913	952
2048^3	688	968	1057

For this teaching library, Double Buffer starts to help at larger sizes, but all custom kernels remain far below cuBLAS. The benchmark target exists to teach measurement, not to claim cuBLAS parity.

Known limitation

ncu profiling was attempted on this WSL2 host and failed with ERR_NVGPUCTRPERM (GPU performance counters not exposed). A reproducible runbook is provided in Performance Tuning; real traces must be captured on a machine where counters are enabled.