85% GPU Usage

[{RaW}Eagle1]

Unfortunately project 764x seems to have dried up? Even if I could find the PCIe bandwidth option in the BIOS there would be little point at this time.

Thanks to all those who have posted / private messaged with their assistance and direction.

Eagle1

[GenDrexler]

From about yesterday, after a core15 update my GTX 570 has been folding these seemingly new 7718 credit WUs(P7643 R14, C0, G3 and another P7643 R22, C0, G10) and only seeing 85%-86% load. TPF for the two that have completed so far average around 00:7:21.

On all other WUs usage is usually 99%; whats going on?

I have reverted everything to stock clocks, and ran the client (v6.41) on its own and get the same thing.

[RozSummer]

I'm having the exact same issue on a GTX560 ti 2 win, and on a GTX 550 ti. Each board is in a different system. The GTX 550 ti is running the 285.62 (whql) driver, and the GTX 560 ti 2 win runs the 295.73 (whql) driver (monitor always on). Both cards fluctuate between 81% and 86%. Tried changing priority to give the cards more CPU time (set priority to "high" while SMP on both systems set to "idle"), but the cards still stay in the 81% to 86% range. Any thoughts on what might be causing the under-utilization of the GPUs?

The upside is it's dropped GPU temps by about 10 deg C.

[freeekonaleesh2]

I am using the client control V 7.1.52 and I have come home today and noticed my CPU is at 100%(witch is fine) but my GPU's are only at 70% usage.
Why did the usage go down from 99% to 70 %?
The PRCG are: CPU:8012 (245, 2, 3) GPU's: 7644 (94, 0, 1); 7642 (79, 0, 2); 7643 (121, 0, 0).
So any reason why my GPU's are not being used fully. I have it set to 100% utilization.
Also, if i turn off my smp, the 3 GPU's go up to 80% usage if that helps out; but still not 99% like before.

Thanks in advance!

[P5-133XL]

Same here, all my GPU's switched over to p7641-p7644 today. They are all running at 85-87% and getting 60% their normal PPD (GTX 460's). I'm changing all the client-type's to see if I can get different projects as they finish.

[bruce]

Welcome to foldingforum.org, freeekonaleesh2.

There are many things that happen when you're processing a GPU WU.

1. The CPU must organize some GPU code with a block of data that it needs to process.
2. The PCIe bus must transfer the data and code to the GPU.
3. The GPU cores must complete the computations.
4. The data must transfer back to main RAM through the PCIe bus.
5. Occasionally checkpoints must be written to disk.
6. Repeat as necessary.

(It could be further subdivided, but that's enough to make my point.)

For the most part, steps 1 through 4 can proceed simultaneously, and if any one of them is saturated, that's as fast as the WU can be processed. For many of the relatively small WUs that we've processed in the past, step 3 has been saturated and steps 1, 2, and 4 can be overlapped and end up waiting for 3 to finish, hence the GPU runs at 100%.

Why should we assume that step 3 is the limiting factor? Probably since it has generally been true. Some of the new projects have a lot more atoms and may be using a different type of analysis. Apparently this alters the balance so that 3 is no longer the step that's the limiting factor; something else is at 100%.

Proteins with 21000 atoms have A LOT more data than proteins with 300 or 900 or even 1800.

It should also be noted that the benchmark machine established the appropriate PPD for these projects while it was running at less than 100% GPU since it is also limited to 100% of steps 1, 2, and 4.

Giving your GPU a partial rest while using the PCIe bus at 100% (if that's what is being saturated) isn't all bad. The total processing time of the WU does not establish the PPD. [/quote]

[antropofob]

Not only has it not dried up, it has apparently left the beta status!
I am running project 7641 on V7 (without -advmethods) and folding on GTX560ti installed in PCIX16 slot.
And I am also seeing 86% GPU utilization, and the ETA is record-breaking 17 hours!
On the positive note, the GPU temperature has dropped by 10 degrees Celsius.

[Blacksmith1]

I have similar results to the above poster. I run 3 GPU's in one system. 2 gtx 480, and a single gtx560. the 480's are running at 800/1600/1848 ( overclocking the memory causes problems on my cards for some reason) and getting an average of 70% utilization. even with the much higher points offered by these WU's, the time it takes to finish them has effectively cut my PPD in half on these cards. They are taking approximately 15 hours on these cards.

Specs, M/B ASUS M3N-HT, CPU Phenom 9500 (O/C 2.5 GHz), Mem DDR2 @ 912Mhz

[7im]

http://foldingforum.org/viewtopic.php?p=211806#p211806

[P5-133XL]

For those having PPD issues with these, changing the client-type to advanced will, at the moment, get you a different project series (p762x). While not good PPD-wise for low end GPU's they are still significantly better p7641-p7644. That being said, projects are always moving around so I don't know how long this will last.

[Blacksmith1]

Thanks 7im. I was just giving data. I fold (almost) 24/7/365 no matter what the points work out to.

[{RaW}Eagle1]

. . . If only people would read through existing threads before posting. . . having criticised you earlier for short and sharp responses 7im: I now feel your pain

It is possible to re-gain lost points production by overclocking your graphics card (if there is headroom to do so) as it is believed to be a PCIe latency issue. I can confrim that increasing the clock speed will decrease time per frame, but not load on the card(s) asI have tested this. Please ensure you have ample power supply to run your graphics card at a higher clock speed on other work units before doing so. It's a good way to blow your power supply. . . I speak from experience.

[kencode]

I have notice a throttle-back from 99% activity on my GTX 580 to 75-80% running V7 client on Project 7642 - can't explain any of this, because the temperature is normal. Is there a setting gone awry?

[7im]

Not a setting problem. It's a larger work unit, and the GPU processors are not the limiting factor for these WUs. Feeding the data fast enough to keep the GPU busy seems to be the limiting factor with these WUs.

[Jesse_V]

There is another thread about this strange behavior that you might consider reading:

viewtopic.php?f=74&t=21180