Swarm, racing Rust, JavaScript, Python and WebGPU

Swarm runs three particle simulations, each written several times over, and races the versions against each other live in your browser. The simulations are a spiral galaxy, fireworks, and a flock of birds.

What is being raced

Each lane runs in its own Web Worker and draws to its own canvas, so they don’t slow each other down or block the page. The views are side by side (Rust and JavaScript), a three-way race (adding NumPy), CPU against GPU, and full-screen single lanes. A link such as #race/flock or #gpu/galaxy opens a particular view.

The simulations

• Galaxy: gravity from a central mass, and clicking drops black holes.
• Fireworks: bursts with gravity, air drag and fading sparks. Black holes swallow sparks.
• Flock: each bird matches the heading of the birds near it. A grid of neighbourhood cells means each bird only looks at nearby cells, which is what lets it scale to millions of birds. Click to send in a hawk.

Keeping the race fair

A benchmark is only worth anything if every lane does the same work. For instance, the flock’s gentle breeze needs a sine function, and each language’s built-in one differs in speed and accuracy. So all of the lanes use the same fast polynomial approximation of sine (accurate to about 0.1%), keeping the comparison like for like.

The GPU lane

The WebGPU lane is the most involved. All three simulations run as compute shaders. After moving, each particle adds itself to its own pixel’s count with an atomic operation, and a final pass turns those counts into colours directly on the canvas. Each frame goes to the GPU as a single batch, and the GPU timestamps its own passes, so the Physics figure it reports includes the pixel counting. It goes up to 16 million particles, while the CPU lanes stop at 4 million. It needs a browser with WebGPU in workers, such as a recent Chrome or Edge.

How it’s served

There is no server logic at all. Docker runs nginx serving static files: the page, the JavaScript, and the compiled swarm.wasm. The Rust source builds inside the Docker image. The NumPy lane downloads Pyodide and NumPy (about 10 MB) from a CDN the first time and browsers cache it afterwards.

Notes

• Results depend heavily on your machine, browser and whether WebGPU is available. The point is the comparison on one machine, not an absolute score.
• Needs a recent browser: Chrome, Edge, Firefox, or Safari 16.4 or later.