GPU compute (general-purpose)
What
GPU compute is using the graphics card for non-graphics work — particle systems, physics, ML inference, image processing, sorting, simulation — by writing programs (compute shaders / kernels) that run massively in parallel across thousands of GPU cores. CUDA was the first commercially serious surface (2007), OpenCL tried and stalled, Vulkan/Metal/D3D12 made it standard in modern graphics APIs, and WebGPU finally brought it to the browser.
Why It's In The Canon
"GPU compute on the web" is what makes the browser viable as a creative computation surface, not just a display surface. Three concrete reasons it matters for design and creative tooling:
-
Particle systems and generative motion — you can simulate 100k–1M particles at 60fps in a browser tab. The aesthetic of "alive interfaces" — Apple's WWDC scenes, Linear's particle backgrounds, every credit sequence with feedback simulations — is GPU compute.
-
Real-time ML in the browser — transformers.js, web-llm, ONNX-on-WebGPU, image segmentation, depth estimation, style transfer. The design surface gains "intelligence" without a server roundtrip.
-
Image + video processing without a backend — color grading, blur cascades, tone mapping, optical flow. Tools that used to require Premiere can ship as web apps.
For practice: get over the "compute shaders are scary" framing. The mental model is: a function that runs on a 3D grid of threads, can read/write buffers, and is dispatched in batches. That's it. The complexity is in deciding workgroup sizes and avoiding race conditions, not in the language.
