Originally Posted by u3b3rg33k
Originally Posted by axipher
PM him and ask for a screenshot, but 8500 PPD on a brand new architecture running an old core doesn't actually sound too far off to me. Even the 5XXX series which was out then the newest AMD GPU core came out do a better job then the 6XXX if I'm not mistaken.
Stanford now faces the problem that the best core for the AMD GPU's (the x16 cores) are still in "beta" and even they don't take advantage of the newest cards so the likelihood of Stanford releasing new core anytime soon for the 6XXX series let alone the 7XXX is pretty low.
If I'm wrong or anyone has other info, feel free to correct me.
Keep in mind that while the "old core, new arch" argument has some validity, it's running on openCL, not Directcompute.
I was hoping for an x86 like IPC improvement, (GCN is, after all "designed with GPU computing in mind") like how a sandy bridge runs the same code that was optimized for a P4, better than the P4. Keeping in mind this is all still in beta...
I don;t know much about OpenCL, but I'm wondering if the language includes features and algorithms for 5XXX, 6XXX, and 7XXX that are more efficient than the last, but the FAH cores are designed using features for only the 5XXX.
Maybe someone with OpenCL knowledge can chyme in here.
So GCN may be a new architecture, as some of 6XXX was, but new architecture won't necessarily mean better performance on the older features if the code running on them doesn't make use of them. Far all you know, the FAH cores could be using just the basics of OpenCL so they can run on AMD GPU's and aren't actually using any optimized features.