The browser-training speedup was a curve, not one headline number

The WebGPU path did not have one honest speedup number.

Against multithreaded WASM SIMD, the measured improvement was 2.6 times on the Small preset, 6.8 times on Medium, 9.3 times on Large, and 12.1 times on XL. Loss drift ranged from 1.1% to 2.5% across the same presets.

The improvement grew with model size because larger matmuls made better use of workgroup tiling and register blocking. Dispatch overhead mattered more on the smaller preset and became less important as the matrix work grew.

This made the old single-number headline misleading even though that number had been measured. It described one point on a curve and invited readers to apply it to every model size.

The better reporting shape keeps the preset, baseline, step definition, and loss-parity result beside each speed measurement. That is enough context to reproduce the comparison and to see where the gain comes from.

The full table is at https://posttrainllm.com/devlog.