Thanks for the write-up SKY! A thorough assessment. If the project maturation for OpenCL units turns out to be anything like that for CUDA, then later, larger projects will likely yield even greater dividends for cards with more stream processors as opposed to those with higher clock frequencies.
Given the resource requirements for AMD-based projects, it behooves Stanford to figure out a way to park the data in VRAM rather than rely on swaps with the system memory. User-added environment variables could partially do this with Brook. I really don't see any reason why Stanford can't do it completely with OpenCL. Certainly, in the current bigadv era, this advancement is too-little-too-late for dedicated folders and should be understood as an improvement for regular users. And regular users don't mess around with environment variables. As you said - a good first step. But there remains much to improve for AMD folders.
|