GPGPU is a technique for offloading heavy computation to the graphics card.

GPGPU programs are written as fragment shaders in GLSL, a statically typed javascript-like language with specialized support for vector manipulation.

In webGL, a pixel is a vector with 4 floats. Typically these elements represent RGBA color channels, but we can use them for anything. For example, by thinking of each pixel as a vector containing (positionX, positionY, velocityX, velocityY), we can simulate a particle by integrating velocity into position on every frame.

Drawing a square with dimensions 1000x1000 spawns 1 million threads. The output of these threads can be written to a texture, which is then sampled as input on the next cycle, allowing state to be passed between frames.

A task may benefit from gpgpu if it

1. has a high computation to IO ratio
2. decomposes cleanly to map, reduce, filter, scatter, gather, sort, and search operations
3. has minimal dependency between data elements
4. is expressed mostly through vector arithmetic

Examples include fft, image processing, simulating physical models like weather and cosmology, neural networks, cryptography, and so on.