A graphics performance wishlist: What we would like hardware to be good at

Algorithms in computer graphics and numerical simulation often possess high intrinsic parallelism, and should be well suited to hardware acceleration on GPUs or similar devices. To explore how well current hardware achieves this goal, I will discuss several representative problems taken from these two fields. For each one, I'll present the structure and characteristics of state of the art massively parallel algorithms for the problem, the limiting factors controlling performance on current GPUs, and what ideal hardware might look like for each problem. In terms of GPU performance in particular, the conclusion is that most parallel algorithms map across fairly well (at least given the recent dynamic parallelism extension) given sufficient programmer time and expertise, saturating either bandwidth or compute depending on the problem. Therefore, the hard challenge (unsurprisingly) is to lower the effect required to perform this mapping.

Speaker: Geoffrey Irving

Woznaik Lounge