A public laboratory for GPU algorithm research. Each lab below is a working WebGPU implementation: hand-written compute shaders, measured against reference implementations, with the optimized variants sealed in the vault. Built to be run, raced, and inspected.
A quantized transformer inference engine hand-written in WGSL compute: packed 4-bit GEMM, RoPE, causal attention with a persistent KV cache, and top-p sampling. Tokens stream in your browser; every kernel's latency is on screen.
in development · milestone 1 of 4Computer vision kernels racing side by side. A webcam feed or a rendered 3D scene flows through WebGPU compute passes: reference implementations vs optimized variants, with GPU timestamp latency and a pixel-identical correctness badge.
in developmentSolver visualizations from the research literature: multi-scale variational optical flow, material point method simulation, and a ReSTIR-style path tracer. Each entry ships with its base implementation open.
in developmentWebXR research with GPU-computed vision passes. Meta Quest passthrough experiments first. A clear-eyed note on smart glasses: consumer AR glasses expose no browser capture API, and this lab says so rather than pretend.
research