0x17 Core lower PPD

[Smith6612]

Hey all, first time posting here.

I've been folding for around 6 years now, and to start off, I appreciate the fact that distributed computing projects like this exist. I've been donating power wherever I can, on newer multi-core systems I own and on capable graphics cards. Recently, one system has come to my attention and I wanted to ask abut it.

I have a Lenovo laptop which is used to fold, which happens to include an Intel i5 Ivybridge dual core CPU @3Ghz (Turbo), along with an NVIDIA NVS 5400M. While CPU folding is going all fine and dandy, the new OpenCL-based Core 17 core is causing both a 4000PPD drop in comparison to the CUDA-based Core 15 core, in addition to a long Time per Fold (difference of 5-10 minutes vs 45 minutes).

Overall it seems the card, due to it's nature, is not only both a low end card (as to be expected), but is very, very inefficient at OpenCL folding versus CUDA folding. Unless there's a driver fix that I could implement, is there a way to restrict the folding client to only fold CUDA work units? Normally, this card with CUDA will do around 6,000 points per day, and with OpenCL not trying to consume an entire thread on the CPU, the CPU Client can also do around 6,000 Points per Day. If the laptop had an AMD card in it, I wouldn't mind OpenCL since those cards do OpenCL nicely, and are stuck to it anyways.

Any help is appreciated

[Zagen30]

Welcome to the forums.

There are definite pros and cons to core 17. On Kepler cards, especially high-end ones, the PPD is fantastic, largely due to the QRB. On Fermi cards and low-end cards in general, that QRB ends up working against you, as projects are benchmarked against a mid-range card.

The only way I know of to guarantee core 15 WUs is to run the v6 GPU client (found here). Core 17 requires v7, so v6 cannot get it. There are no drivers that will help the situation, though it's been widely reported that the more recent Nvidia drivers (331.xx, 332.xx, and 334.xx) folding much worse than some older ones (326.xx or 327.xx) on everything except GK110 cards (780, 780 Ti, Titan). This is usually is most noticeable with mid-range Keplers like the 770, 660 Ti, etc., but it may also be true for Fermis.

[Smith6612]

Thanks for the reply! I can give the v6 GPU client a try. I used to run that back in the Vista days.

[Zagen30]

By the way the driver issues only apply to core 17 as far as I know.

[n_w95482]

There are no drivers that will help the situation, though it's been widely reported that the more recent Nvidia drivers (331.xx, 332.xx, and 334.xx) folding much worse than some older ones (326.xx or 327.xx) on everything except GK110 cards (780, 780 Ti, Titan). This is usually is most noticeable with mid-range Keplers like the 770, 660 Ti, etc., but it may also be true for Fermis.

I haven't tested it with my GTX 570 yet, but my GTX 460 1 GB gets the same PPD with 331.82 as it does with 327.23. That was definitely not the case with my GTX 680 - PPD fell off of a cliff with anything newer than 327.23. So, at least with my small test sample, Fermi seems to be unaffected by the latest drivers.