Has anyone figured out the low GPU usage issue yet?

[Un4Given]

Hey folks,

I've been bouncing around the threads and noticed a great deal of the low GPU usage issues going on. I too am seeing that issue and have tried the fixes that others have suggested. I even ran into an issue along the way where F@H would go Ready > Running > Ready > Running > FAIL. I've uninstalled and used other cleaners to make sure all previous driver files were gone, and even upgraded to 7.3.6 even though I like the 7.2.9 interface better. Finally got F@H to stop crashing, but I'm back to mid 60% on GPU usage. I've tried several driver versions from 12.4 through the latest 13.2 Beta.

Has anyone found anything that consistently resolves this issue? This is with a 6970, but my 7950 doesn't seem to have this issue, and constantly runs in the 91-93% range.

[bollix47]

Hi Un4Given and welcome to the folding support forum.

As you've already seen this is not an uncommon problem. However, a new GPU core is currently being tested which should correct most, if not all, of the low usage problems. Until it's released to the public we can only be patient or apply to join the beta team.

Also of note, there are some projects that make better use of your GPU than others. Hopefully, in time, all projects will use your GPU to it's maximum potential.

[mmonnin]

12.8 was the last good driver version. Others have modified drivers or applied the good SDK to drivers past 12.8.

[Un4Given]

Thanks bollix. Since I also do BOINC tasks I'm not really at a point where I want to be a beta tester, but thanks.

I'll wait for the newest release and see how it goes.

[bruce]

At any time, one of the steps in processing the WU is blocking higher throughput. Frequently it's the cores in the GPU that are the speed limitation and then the GPU is reported at working at 100%. Obviously that's not the only step that might be working at 100%. The PCIe bus has to transfer the data to/from main RAM and the GPU and it might be a limiting factor. The FahCore running on the CPU side has to assemble a block of work to submit to the drivers and the drivers have to decide when to have them transferred to the GPU. Either of those can be the limiting factor. There probably are other steps along the way that could be limiting throughput. With a complex series of steps, why should we assume that only the GPU's compute step should be the step that's saturated?

Assume, for a moment, that we can instantly replace our GPU with faster hardware (more cores/faster clock rates/etc.) while keeping everything else the same. Every time we boost the GPU performace, it puts more stress on the other steps and at some point the GPU is no longer the limiting factor. Something else is. Now assume you also have the ability to change the performance of the hardware involved in that particular step. Speeding it up will shift the limiting factor to something else. The real trick is figuring out which factor is limiting throughput and what can be done about it -- and that's not easy to do.

[Un4Given]

Hi bruce. The system in question has no function other than crunching. As a 20 year system admin, I stripped out all unnecessary Windows programs and services, as well as disabled all legacy hardware in the BIOS. This isn't a resource issue, but I do very much appreciate you taking the time to respond. Community support is what makes dealing with these issues bearable.

[djay86]

All i can say id that core_16 is really dependant on the cpu, when with my 285GTX core_11 used a little of the CPU like 2~3 percent.

Edit: 13.2 beta works fine from my side 90% and above gpu load, but now it took 20% of my cpu usage limiting my CPU folding.

[mmonnin]

Try 13.1 if you are on 5xxx/6xxx series AMD.

[djay86]

Try 13.1 if you are on 5xxx/6xxx series AMD.

I Forgot to indicate that I have a 7950 and a 3570K as CPU

[tictoc]

All i can say id that core_16 is really dependant on the cpu, when with my 285GTX core_11 used a little of the CPU like 2~3 percent.

Edit: 13.2 beta works fine from my side 90% and above gpu load, but now it took 20% of my cpu usage limiting my CPU folding.

You might want to check your TPF on the 13.2 beta drivers.

With the 13.2 beta drivers GPU load will be at 90+%, but TPF and PPD numbers correspond to something closer to 60-70% GPU usage.

[mmonnin]

If you're running production WUs then you'll want 12.8 or modified newer drivers using SDK 2.7.

[djay86]

All i can say id that core_16 is really dependant on the cpu, when with my 285GTX core_11 used a little of the CPU like 2~3 percent.

Edit: 13.2 beta works fine from my side 90% and above gpu load, but now it took 20% of my cpu usage limiting my CPU folding.

You might want to check your TPF on the 13.2 beta drivers.

With the 13.2 beta drivers GPU load will be at 90+%, but TPF and PPD numbers correspond to something closer to 60-70% GPU usage.

Sorry, I don't now what TPF mean.

[Jesse_V]

Sorry, I don't now what TPF mean.

Not a problem. TPF stands for "time per frame". A frame is a fraction of a Work Unit, usually 1%. You can check the TPF value using FAHControl. Its listed on the right-hand side.

The TPF number is useful because it's a raw measurement of how quickly your computer is competing that Work Unit. The TPF number can be very different depending on the project, but it can still be useful for direct comparisons. The Points Per Day value is calculated from that, and the resulting number tries to even the variation out. So both numbers can be handy, but for different reasons.

[djay86]

Thanks for the explanation.
I checked the TPF and PPD and i have 6k PPD, i guess it's alright for an HD7950, coupled with the CPU it's a total of 10k PPD it's not that bad.

[tictoc]

A 7950 on the 12.8 drivers, or a modified newer driver, will get 8-12k PPD. With more of the "good" 11292 WUs being sent out recently you would probably get close to 12k PPD.

Even if you are not looking for more points, your GPU will get more work done on the proper driver. With 90% GPU usage and 6k PPD your GPU is currently only running at ±50% efficency.