WASM single-stream throughput: raw run data Hardware: Apple M4 Pro (24 GB), 14 cores (10P + 4E), macOS 15.6.1 Build: emcc 5.0.7, -O3 -msimd128 -ffast-math, -sENVIRONMENT=web,node WASM source: wasm/microgpt_inf.c (post RMSNorm-eps fix) Methodology: 100,000 forward passes per run, after 20,000-token warmup, elapsed time from emscripten_get_now() (wraps performance.now) The number depends on the browser and host context. Two environments measured: ------------------------------------------------------------ Environment A: regular Chrome (typical user-facing number) ------------------------------------------------------------ Mean of 5 runs: 1,341,206 tok/sec CV: 0.18% This is what most readers will see when they open the live demo in their own browser. Used as the headline number in README and report. ------------------------------------------------------------ Environment B: Chromium 146 / Electron 41.3.0 inside Claude Code preview ------------------------------------------------------------ Session 1 (5 runs): 1,976,285 / 2,044,990 / 2,040,816 / 2,049,180 / 2,044,990 mean: 2,031,252 CV: 1.36% Session 2 (5 runs): 2,049,180 / 2,044,990 / 2,049,180 / 2,032,520 / 2,049,180 mean: 2,045,010 Combined mean (10 runs): 2,038,131 range: 1,976,285 — 2,049,180 The Electron-embedded Chromium runs noticeably faster than a normal Chrome browser on the same machine — likely a different V8 build / different sandbox / different scheduling. Reported separately; not used as the published headline because most readers won't see it. ------------------------------------------------------------ Numpy reference cross-check (verify_against_numpy.py) ------------------------------------------------------------ Verified live by `python3 wasm/verify_against_numpy.py`: - First-token logits: max |diff| ≈ 1e-6, argmax matches - Autoregressive trace ('emma'): all positions match within fp32 noise ------------------------------------------------------------ To reproduce ------------------------------------------------------------ cd wasm && ./build.sh python3 -m http.server 8765 # open http://localhost:8765 in regular Chrome and click "Benchmark"