Folding Forum

I've had the same problem with these units ever since I began receiving them. GPU use never reaches higher than 80%-81%. Both times I ran one on an old GT 430 and then I upgraded to the GTX 560 Ti. Both cards never even got close to 99% Frame times decreased and far higher PPD because of the more powerful GTX. Several other posts have the same thing happening. It is most likely the PCI-E bandwidth speed not being able to keep up with the amount of data created by the GPU. Overclocking will make this even worse, although the frame times can be decreased with more MHz. What sounds silly is what works. Actually down-clock to the speeds of a lower end card. Keep in mind this is my own experiment and even I'm not going to continue this method after this post. I'm running at 500mhz gpu, 525mhz memory, 1000 MHz shader clock, and 1.087 voltage. With a GTX 560 Ti... I'm getting up to 88% usage out of the GPU as apposed to 80% or less. The frame time and the PPD both suffer from this so I don' really think it is worth it just to get the extra use percentage increase. Basically the data is being created slower so the bus can move it. Like a funnel or hourglass effect.

I put in the original post identifying this as an issue, however I can confirm that overclocking the cards has no effect on the GPU usage. My cards stayed at 86% load regardless of their clock settings, however PPD saw a proportional increase. . . This led me to believe the limiting factor was the latency / speed of the data transfer on the PCI bus rather than the bandwidth.

viewtopic.php?f=74&t=21177

That's exactly what I was thinking too. I don't think that I used the correct terms, It was a long day for me and I was falling asleep while typing! Checking my folding rig is always the last thing I do at night. It would be cool if they made a project core that could analyze the cure/remedy for micro naps, insomnia, and sleep paralysis.

In theory there should be a "tipping point" on a down-clocked graphics card at which point the supply of data via the PCI bus matches that of the clock settings. At that point one might expect to see a proportional increase in load due to the graphics card no longer having to wait for data from the PCI bus.

If I happen to pick up work units that behave in a similar way again then I'll downclock my cards, bit by bit and publish a table of the results. Instead of underclocking proportional to the lost load (14%), I overclocked by 14% to attempt to regain lost PPD. My theory is that if I had underclocked by 14% rather than overclocking I would see load levels of much closer to 100%. In theory a card runnnig a shader clock of 700MHz achieving 80% load should have the same work rate as the same card with a 540MHz shader clock running at 100%. . . Although I'm willing to bet I've over simplified the issue.

There may be a "tipping point" of sorts, but how much time does it take to transfer the data? Is it always possible to move 100% of the data during the same time period that the shaders process 100% of the data or does some of each always have to be processed while the other one is waiting?