sswilson wrote:I believe Stanford has typically applied a "30%" possible range on what PPD might be expected, and believe it or not, these come close to being 30% below recent standards. What's making them look soooooo bad is the fact that we just got through a batch of WUs which were providing PPD well over 30% above what has been the standard.
I'd like to see a more consistant PPD being applied, but am willing to cut Stanford some slack on this one.
I'm not aware of any +/- 30% policy, but that doesn't mean it doesn't exist.
What I do know is that they benchmark on a GTX 460 and if you're using anything else, hardware differences will not accelerate all Projects identically, resulting in variations in PPD which Stanford cannot predict nor can they change them.
I've alluded to suspicions that the PCIe bus may be the limiting factor for these projects rather than the compute speed of the GPU. See Subject: TPF is WAY long -- P7641 for one such example.
As has already been pointed out, the speed of the data transfers has two important components, bandwidth and latency.
What I do not see is anybody reporting enough detailed information for us to draw any concrete conclusions.
I was going to play with the bus speed / cpu clock / mem speed to see if I could get any kind of quantifiable results, but unfortunately I've been pulling 8XXX WUs on my GPUs for the last few days.
And as far as the %30 goes.... yeah, nothing official, but it's been the kind of response I remember getting any time I've questioned the PPD output of a particular WU over the years.
bruce wrote:I've alluded to suspicions that the PCIe bus may be the limiting factor for these projects rather than the compute speed of the GPU. See Subject: TPF is WAY long -- P7641 for one such example.
As has already been pointed out, the speed of the data transfers has two important components, bandwidth and latency.
What I do not see is anybody reporting enough detailed information for us to draw any concrete conclusions.
Would not the reference PC-CPU-GPU take care of that in the PPD calculation?
Sorry if I'm being dumb, I just Fold them, & notice the 5k+ drop in PPD
sswilson wrote:I was going to play with the bus speed / cpu clock / mem speed to see if I could get any kind of quantifiable results, but unfortunately I've been pulling 8XXX WUs on my GPUs for the last few days.
And as far as the %30 goes.... yeah, nothing official, but it's been the kind of response I remember getting any time I've questioned the PPD output of a particular WU over the years.
Actually what has been refered to here on this site by most people is 15% + or - which = 30% in reality. Which in my opinion is too much there are some smp projects that currently have a variation of of 100+% 7011 which is the slowest around 20,000 PPD on a i7 970 @ 4.4Ghz and a 7021 which gets 68,000 on the same rig the bulk of the smp projects give me an average of 35,000 PPD. This has been pointed out numerous times but is apparently acceptable so who knows what the real + / - numbers are.
The current 7640 and the 7641 - 7643 were noted as being below normal from the day they first came out but I believe the 7640 was re-benched and adjusted. I do not think the 7641 - 7643 ever were. But apparently they are set at the right value for the science being done.