Connect Four: one model, two runtimes (Speed)

Can the same 7.4 M parameter model as in the main demo run faster? Yes: with custom kernels in onepass-webgpu, a small WebGPU runtime, from the fp32 file or the main demo's int8 file, compared with onnxruntime-web on WebAssembly. A perfect Connect Four solver is timed alongside: faster on most moves, but far less consistent. The story behind it: One millisecond to make a move.

Loading…

Who starts?

Time per move

Every AI move is scored by every engine, one after the other.

engine	this move	median	slowest	moves

Benchmark: 200 boards

Each engine scores the same 200 boards from 30 complete games, one at a time.

What gets loaded

engine	runtime code	model file	start-up

Start-up is loading the engine in your browser and its first move, not the download.

Comparison

Lower is better; the best in each column is blue, the worst red. Times fill in after the benchmark.

engine	size (runtime + model)	start-up	median	mean	p95

Technical checks

Before the WebGPU runtime is trusted, it must agree with onnxruntime-web (wasm, same fp32 file) on reference positions. The encoder is also checked against the Python tooling's bytes.