Previous | Next --- Slide 57 of 63
Back to Lecture Thumbnails
Kuragama

Is it possible for non-graphics applications of GPUs (e.g. mathematical / scientific applications and cryptocurrency mining) be designed to take advantage of the fixed-function blocks of the GPU, or are they largely limited to the generic cores?

keenan

@Kuragama Yes, in fact, before there were ever languages like CUDA and OpenCL which let you write non-graphics code for the GPU, there was a whole research trend called "GPGPU" or "general purpose programming on the GPU," where the idea was to figure out how to express general purpose computations in terms of graphics operations. E.g., want to take the average of two very large vectors? Ok, blend them together using 50% alpha; now you're using the Zbuffer/blend silicon (also known as the "raster operations" or ROP) to do your calculations, rather than the shader. This was kind of a dark time, and people now just write everything using the languages mentioned above. But yes, if you really wanted to squeeze max performance out of the card, you could still play these kinds of tricks...