bruce wrote:The FahCore that works on Fermi/Kappler doesn't use that new feature. Since you used the word "allows" which implies that software is not required to use unified memory. How could that prevent the existing code from working?
From the published GTX 750 Ti benchmark results from Tom's and AnandTech, I think the problem is probably
unrelated to OpenMM or OpenCL, and it's probably related to CPU <-> GPU communication. Specifically, memory
barrier, and memory coherence come to mind.
To quote Dr. Pande, "... because of how our old core 15 and 16 was written, it was in fact easier for us to write
the core (17) from scratch." Perhaps in the new FahCore17 code, certain assumptions are made, about how future
GPU will communicate with CPU, and these assumptions are now no longer valid in Maxwell. However, comparable
assumptions made in FahCore15 are still valid. Hence FahCore15 still runs fine, though with degraded performance
(due to the fact that FahCore15 does not support OpenMM 5.1).