You are viewing the course site for a past offering of this course. The current offering may be found here.
Lecture 23: How GPUs work (50)
ayushsm

Is there an upper limit at which there's no real point to including more threads/fragments? It seems like this is close to what we expect for "high performance" but can we somehow decide at one point there won't be any more benefit (e.g. something like Moore's Law starting to tail off?)

avinashnandakumar

@ayushsm I was thinking the same thing, but I think there maybe different upper limits depending on the purpose/application that we want out out of the GPU. I was watching this really interesting talk from Tesla about how they constructed their neural net accelerator and its very similar to how we talked about what operations we would have to speed up in order to make a GPU really fast and efficient. So I think that if we want to render for a gaming system vs a laptop vs control computer vision system the upper threshold of high performance varies! It's really interesting and cool to think about though.

You must be enrolled in the course to comment