Benchmark methodology
Sources, scope, and reproducible measurements.
01Published report#
The 2026-09-29 tables live in benchmarks/RESULTS.md; the browser does not run live C++ benchmarks. The top of that file records the evening of 2026-10-21 on this Ryzen 7 5800H: GEMM kept kernel avx2-4x16, N=2048 Nexus 330/362/369 GFLOPS, TensorFlow 389. CPU Conv2d was behind PyTorch at 31.6 GFLOPS; Nexus was about 4.9 to 6.8 GFLOPS. Depthwise AVX was ahead of PyTorch 4.93 and TensorFlow 0.688 on that column. AVX-512 is not selected. There is no world ranking. A lowest median in 7 of 9 September workloads describes only that single-thread table. GPU figures are a historical archive. MSVC 19.44, /O2, Release. GPU RTX 3070 Laptop, CUDA 13.4, sm_86.
02CPU method#
Module tests report medians from 20 warmups and 30×5 calls. ReLU is forward-only; the others include forward+backward. Cross-library and revision kernel comparisons use three rounds and 90 samples; raw samples are downloadable.
03GPU and precision#
cudaEvent, resident device data; H2D/D2H excluded. The source reports default TF32 tensor cores for Ampere cuBLAS SGEMM and maximum absolute deviation 1.8e−5 at size 4096. Do not generalize to other versions or hardware.
04Reproduce#
Build Release and record hardware and power profile. More warmups/repeats and median/p95 reporting strengthen comparisons. Do not directly compare CPU forward+backward pairs with GPU forward kernels.
.\build\Release\nexus_model_bench.exe.\build-cuda\Release\nexus_model_bench_gpu.exe