WebGPU

What it replaces

WebGPU is the successor to WebGL, and the difference is not mainly speed. WebGL was a graphics API that people abused to do maths: to multiply two matrices you encoded them as textures, drew a triangle and read the pixels back. WebGPU exposes the GPU the way native APIs like Vulkan, Metal and Direct3D 12 do, with explicit compute shaders, buffers and command queues, so numerical work is a first-class operation rather than a trick.

Shaders are written in WGSL, a language defined by the spec rather than borrowed from OpenGL. In practice most people never see it: they use a library that ships its own kernels.

Where you actually meet it

The feature detection is one line, and everything else depends on it:

if (!navigator.gpu) {
  // No WebGPU. There is no software fallback worth having.
  return;
}

The visible application is on-device machine learning. Runtimes such as web-llm, ONNX Runtime Web and transformers.js compile models to WebGPU kernels so inference happens on the visitor's own hardware. NorthernGo uses this for its local generation mode: a quantized model is downloaded into the browser once and then runs on your graphics card, with nothing sent off the device.

What to watch out for

• It is not a compatibility layer for CUDA. The mature Python GPU ecosystem does not come along. You get what your runtime has ported.
• Browser support is uneven, not universal. Desktop Chromium has had it longest; other engines and platforms have arrived at different times and behind different flags. Always feature-detect, and give a real fallback rather than a broken page.
• Available video memory is the real ceiling. A model that needs more VRAM than the machine has does not run slowly, it fails — often with an error that says nothing useful.
• It is slower than native. Expect a fraction of what the same GPU does through a native runtime. That is usually an acceptable price for requiring no install at all.

Mobile support exists on paper. Downloading several gigabytes of weights over a phone connection, then holding them in memory, is a different question from whether the API is present.