Measured on your GPU · 20 / 20 verified on both APIs

CUDA ↔ WebGPU

RTX 5080 · CUDA 13.3 · NVIDIA driver 616.64 · 2026-09-11 UTC
Ten identical kernel sources, two sizes, real GPU execution. Median microseconds per operation. Small text shows the observed sample range.

WebGPU / CUDA above 1 means CUDA graph replay was faster. CUDA graph timing and WebGPU compute-pass timing both batch GPU-resident operations. Ordinary CUDA launches expose submission gaps. WebGPU wall time also includes encoding and completion waiting. Tiny differences can be noise.

Workload size

Kernel	WebGPU GPU µs	CUDA graph GPU µs	WebGPU / CUDA	CUDA ordinary GPU µs	WebGPU wall µs
saxpysmall	1.631.63–1.79	1.491.49–1.80	1.09×	6.66	3.31
saxpy_vec4small	1.441.41–1.60	1.401.27–1.72	1.03×	6.46	2.96
matmul_naivesmall	19.9719.33–20.35	6.626.62–6.96	3.01×	10.25	24.87
matmul_tiledsmall	3.203.20–3.33	2.882.88–3.01	1.11×	7.00	5.52
matmul_registersmall	5.895.76–6.53	3.063.06–3.18	1.92×	8.00	11.11
reduce_sumsmall	3.683.68–3.71	3.553.39–3.59	1.04×	18.63	7.78
convolutionsmall	2.052.05–2.11	1.811.67–1.98	1.13×	6.34	3.10
histogramsmall	4.104.10–4.29	3.823.82–4.06	1.07×	13.90	7.10
transposesmall	1.761.73–1.95	1.211.21–1.38	1.46×	6.73	3.07
particlessmall	1.441.41–1.54	1.741.74–1.82	0.83×	6.74	3.31
saxpylarge	16.2616.16–16.29	16.1616.06–16.33	1.01×	18.88	17.51
saxpy_vec4large	10.029.95–10.05	10.079.97–10.32	0.99×	12.67	11.25
matmul_naivelarge	231.42230.53–232.06	107.18106.95–108.53	2.16×	109.64	235.83
matmul_tiledlarge	64.2663.62–64.77	63.0862.21–63.34	1.02×	65.20	66.86
matmul_registerlarge	46.2145.95–47.49	43.0842.92–44.10	1.07×	45.82	49.00
reduce_sumlarge	20.6120.48–20.77	18.7118.64–18.86	1.10×	25.36	25.57
convolutionlarge	21.0920.90–21.25	16.3916.34–16.55	1.29×	18.99	22.23
histogramlarge	24.7024.64–24.99	19.8619.73–19.95	1.24×	23.36	27.85
transposelarge	13.4113.31–13.60	9.169.07–9.20	1.46×	11.16	15.04
particleslarge	4.514.51–4.74	4.934.72–5.03	0.91×	7.68	6.22

Nine samples; 2,048 repeats per sample, 512 for matrices. Full reduction chains and histogram clearing are included. Particle compute excludes rendering. Inputs reset outside each sample; buffers are reused inside the sample. Float32 reference checks pass; integer histograms match exactly. This measures these kernels, not optimized CUDA libraries or cold transfers.

Full methodology and reproduction commands · Combined raw results · Open live particle app