Low / Inconsistent PPD with Radeon 7950 & Core 17
Posted: Thu Jan 08, 2015 4:48 am
I have a couple of new (old) machines to add to my collection of number crunchers, and I am hoping that someone can point me in the right direction to get them cranking out research at as high a rate as possible. Both machines use Opteron 6272 CPUs and Radeon HD 7950 GPUs on stock clocks (for now). One has 8GB of RAM and the other 12GB. Both have slow 5400RPM freebie HDDs (500GB and 1TB). Both were wiped and had a fresh Win8.1x64 install done on them, and based on what I read on here I installed Catalyst 14.9 drivers. Also, both are connected to the same monitor, one via DVI and the other via a DVI=>VGA adapter.
When I first fired up the F@H client and let it run for a couple of hours, I was seeing ~16K PPD from both CPUs, 290K PPD from one of the GPUs (project 13000) and 13K PDD from the other GPU (project 13001). TPF was 2:47 on the one machine with a fast-rising completion percentage to match, and 22:50 on the other. I checked the Catalyst performance panel and both seemed to be getting 100% GPU utilization. I then swapped the monitor cables between the 2 machines (again, both use DVI output but one has a DVI=>VGA adapter since the old monitor has one of each type of input), and I paused/restarted the F@H clients. After that, both were seeing ~13K PPD after an hour. So, I uninstalled the F@H client (I know it is bad practice and I very rarely do it) and fully removed the Catalyst drivers, and reinstalled both (sticking with Catalyst 14.9). After a few hours, both machines were sitting at ~12% completion on 13000 and 13001 WUs, with ~27K PPD and ~22:50 TPF.
SO...was the 290K PPD and 2:47 TPD that I was seeing on project 13000 a fluke of some sort (didn't seem like it since the progress bar was going quite fast)? From looking at some of the GPU PPD databases, it seems like a 7950 should be seeing 70K - 120K PPD, although I am not sure which core version those numbers were with. Anyway, can anyone that has been folding with an HD 7950 chime in about what sort of performance I should be expecting from Core 17 WUs? I will have access to the machines again tomorrow, so if the good folks here need to see logs I can pull them then. Is it still the case that Catalyst 14.9 is preferred for older GPUs rather than the latest version? Are there any secret driver settings hidden somewhere that could be hindering performance? This is probably a totally redundant question at this point too, but is there any dependence on what is connected to the GPU output (I saw mention of dummy plugs being needed on an older folding client)?
Thanks!
When I first fired up the F@H client and let it run for a couple of hours, I was seeing ~16K PPD from both CPUs, 290K PPD from one of the GPUs (project 13000) and 13K PDD from the other GPU (project 13001). TPF was 2:47 on the one machine with a fast-rising completion percentage to match, and 22:50 on the other. I checked the Catalyst performance panel and both seemed to be getting 100% GPU utilization. I then swapped the monitor cables between the 2 machines (again, both use DVI output but one has a DVI=>VGA adapter since the old monitor has one of each type of input), and I paused/restarted the F@H clients. After that, both were seeing ~13K PPD after an hour. So, I uninstalled the F@H client (I know it is bad practice and I very rarely do it) and fully removed the Catalyst drivers, and reinstalled both (sticking with Catalyst 14.9). After a few hours, both machines were sitting at ~12% completion on 13000 and 13001 WUs, with ~27K PPD and ~22:50 TPF.
SO...was the 290K PPD and 2:47 TPD that I was seeing on project 13000 a fluke of some sort (didn't seem like it since the progress bar was going quite fast)? From looking at some of the GPU PPD databases, it seems like a 7950 should be seeing 70K - 120K PPD, although I am not sure which core version those numbers were with. Anyway, can anyone that has been folding with an HD 7950 chime in about what sort of performance I should be expecting from Core 17 WUs? I will have access to the machines again tomorrow, so if the good folks here need to see logs I can pull them then. Is it still the case that Catalyst 14.9 is preferred for older GPUs rather than the latest version? Are there any secret driver settings hidden somewhere that could be hindering performance? This is probably a totally redundant question at this point too, but is there any dependence on what is connected to the GPU output (I saw mention of dummy plugs being needed on an older folding client)?
Thanks!