Thinking of a modern GPU as a “graphics processor” is a bit misleading. GPUs haven’t been purely graphics processors for 15 years or so, they’ve morphed into general-purpose parallel compute processors with a few graphics-specific things implemented in hardware as separate components (e.g. rasterization, fragment blending).

Those hardware stages generally take so little time compared to the rest of the graphics pipeline that it normally makes the most sense to have far more silicon dedicated to general-purpose shader cores than the fixed-function graphics hardware. A single rasterizer unit might be able to produce up to 16 shader threads worth of fragments per cycle, so even if your fragment shader is very simple and only takes 8 cycles per pixel, you can keep 8x16 cores busy with only one rasterizer in this example.

The result is that GPUs are basically just a chip packed full of a staggering number of fully programmable floating-point and integer ALUs, with only a little bit of fixed hardware dedicated to graphics squeezed in between. Any application which doesn’t need the graphics stuff and just wants to run a program on thousands of threads in parallel can simply ignore the graphics hardware and stick to the programmable shader cores, and still be able to leverage nearly all of the chip’s computational power. Heck, a growing number of games are bypassing the fixed-function hardware for some parts of rendering (e.g. compositing with compute shaders instead of drawing screen-sized rectangles, etc.) because it’s faster to simply start a bunch of threads and read+write a bunch of pixels in software.