Results 1 to 4 of 4
Thread: ~interesting~ GPU behavior
04-30-12, 12:07 AM #1
~interesting~ GPU behavior
So in another post I was playing around with various GPU configurations trying to find a balance in PPD gain on the GPU's and PPD loss in overhead on the CPU.
550 TI about 11,200 PPD
I noticed when I added my 8800GT (5800 PPD) the 550 would drop to about 7000 PPD no matter the WU. Remove the 8800GT and we are back to 11,000. Swap it with the 9600GT (3800 PPD) the 550 stays over 11,0000.
Now the only possible explanation I can come up with is the 8800GT is PCI-E 1.0 and the other two are PCI-E 2.0 cards. I know the bus is backward compatible but I didn't think it would drop everything on the PCI bus to 1.0 speeds. Also I didn't think we were hitting the PCI-E 1.0 speed limits yet even on current high end cards. Does folding use up this much bus bandwidth?
04-30-12, 05:02 AM #2
The size of the memory in the card, it's speed, and the number of shaders it has, determines that.
For a card with a small amount of memory, it will have to be given a new packet, very frequently.
On your system at least, I'd say you have shown that yes, it does knock down the other card to the lower PCI speed, and yes, it does use up a substantial amount of the bandwidth, at that speed.
I believe the bandwidth on the bus is the reason that the second card always slows down the first one's folding, when you have a multi-core cpu. Haven't seen any actual figures for it, however.
04-30-12, 10:20 AM #3
Is it possible that on the 8800GT install that you actually wound up with two instances running on the 560? That would result in 1/2 the ppd. THe other possibility is that with the 8800GT installed that OS power saving kicks in and drops the clock on the 560ti, also resulting in 1/2 the ppd. It IS NOT a bandwidth issue.
04-30-12, 07:23 PM #4
I didn't think we were reaching PCI-E limits (4GB/s 1.0 16x) but that was the only major part that I found different between the 8800GT and 9600GT.
This is all in Linux by the way.
I had checked the Nvidia Settings app which shows the 550Ti at "performance" setting while folding and believe me I was watching that screen closely for the temps. I swapped these three cards around a lot and I did notice when running memtestG80 if I set it to run on GPU 3 (starts at 0) then it would error out and tell me which GPU's are at 0, 1, and 2. Which one's were where changed each time I swapped a card but stayed persistent after a reboot if the config did not change. I did accidentally run a G80 client on the Fermi card but this was quickly caught and corrected. So yes the correct clients were running on the correct cards after I caught my mistake.
I could see this being a driver bug but not an installation issue. In Linux the driver dynamically detects who is where and presents the features for what it detects in each bus. I had no issues with Nvidia settings showing the correct current config.
I will have my water cooling parts in later this week and I have two GPU water blocks in the shipment. I'm really on the fence about my GPU configuration. I would love to get the 8800GT in as my second card and put and end to the 9600GT but I think the G80 based folding cores are hogging more CPU time than the OPENMM core the Fermi uses and between the 4000PPD drop (driver or bus issue) on the 8800 and the CPU PPD drop from using the 9600 I think I'm better off replacing them with another 550 Ti. I know the 560 is a lot better but after the water parts I've ordered I just can't justify $200 on another card when the 550's are about $100. This would also finally give me an SLI option.
Even if I go with a second 550 I will still have an open slot for the 8800 if I can figure out what is going wrong. But this might be time to pass the torch. This card was free thanks to this team and I know some one who currently has a zero PPD but has contributed a lot to team 32. I should ping him and see if he can use this.