gpukernel

Real-time GPU compute in the browser via WebGPU compute shaders (WGSL). gpukernel runs a parallel N-body particle simulation on the GPU and benchmarks the same kernel against a CPU baseline — with an automatic CPU fallback when WebGPU isn't available, so it works in every browser.

What it demonstrates

• Real GPGPU. A WGSL compute shader with storage buffers, bind groups, and workgroup dispatch — one thread per particle. (ADR 0001)
• Honest capability handling. Runtime WebGPU detection with a CPU fallback; the app degrades gracefully instead of failing. (ADR 0002)
• A fair benchmark. Identical N-body math on GPU and CPU, compared by throughput (particle-updates/second). (ADR 0003)

Layout

Path	What it is
web/	Next.js frontend + all WebGPU/CPU compute
docs/	Architecture + ADRs

There is no backend — all computation runs client-side in the browser.

Develop

Prerequisites: Node 20+. (A .devcontainer ships it.)

cd web
npm install
npm run dev      # http://localhost:3000

Open in a WebGPU-capable browser (recent Chrome/Edge/Safari) for the GPU path; other browsers automatically use the CPU fallback.

Deploy

Static frontend on Vercel — Root Directory = web. No backend, no env vars.

Status

• WGSL compute kernel + GPU pipeline
• CPU fallback + runtime capability detection
• GPU vs CPU benchmark
• Deployed on Vercel