Comments by "Anders Juel Jensen" (@andersjjensen) on "The Rise and Fall of the Cray Supercomputer" video.

@magnemoe1 Actually it's modern GPUs that work a lot like Cray's geometry vectors. Old GPUs were a hardware implementation of fixed function geometry pipelines like OpenGL up to version 2.x and DirectX up to version 10. They were basically "a program made in hardware". Modern GPUs are SIMD devices. (SIMD = Single Instruction, Multiple Data). Or pretty much what Jon described as vectors. But they're also massively parallel with thousands of cores. So if you want to project a 3D scene onto a 2D plane (the screen) you're running the same tiny little program, with enormous amounts of geometry data as arguments, split up over thousands of cores. In the next frame the angle has changed only slightly, so the tiny little program (called a shader program) only have a few cos()/sin() values updated to reflect that, and the whole thing goes again. In addition to being able to work on the vertices of a 3D mesh modern GPUs can also work directly on the pixels in the 2D projection to adjust brightness, transparency, etc. This is why running computer graphics becomes more and more demanding as resolution increases. But it basically works the same: a very small program iterates over all the pixels, but to do that effectively long rows are loaded into wide registers in many cores. (Disclaimer to fellow nerds: This was purposefully kept incredibly superficial to the point of introducing slight inaccuracies)
1