CPU and GPU?

[BornAgainPagan]

Hi,
I'm kinda new here, ran Folding@home a few years ago and just started back up again with a new computer. i7 2600k, 8 Gb ram and two HIS IceQ Turbo ATI Radeon HD 6850 video cards in CrossfireX, Win 7 Pro.
My question is can I run the CPU and the GPU client at the same time and does the GPU client support two video cards in CrossfireX. I tried installing the GPU client and got a msg that my GPU was not supported.
So far the CPU client seems to be working good, takes about 7 mins per step of a WU. Would there be any advantage to running both the CPU and the GPU clients at the same time?

Thanks,

*Forgot add that I'm running Folding@home client 6.23

[Zagen30]

Welcome back, BornAgainPagan.

The short answer is yes, you can fold on your CPU and both GPUs, and there is a significant advantage to doing so.

I would recommend downloading and running the v7 client from the main page. It's an all-in-one client that can manage all of your hardware from one location, as opposed to the various v6 clients, where you had to run one per piece of hardware. The default installation should set up one SMP slot and 2 GPU slots automatically, and v7 should acknowledge your GPUs right off the bat. The install guide is here and the FAHControl guide is here; FAHControl is the GUI used to configure and manage the client.

When the installer asks for your username/team information, you'll see a field for a passkey, which can be obtained here. You don't need a passkey, but having a qualified one allows you to earn bonus points for turning in WUs quickly, and with that hardware the bonus points will be substantial assuming you're running near 24/7. As of right now only CPU-based WUs (SMP and uniprocessor) earn bonuses, but news yesterday indicated that GPU WUs will start earning bonuses in the near future (probably on the order of months, but that's just rather uninformed speculation on my part; the Pande Group never gives substantive ETAs). You have to successfully turn in 10 bonus-eligible WUs and maintain a >80% successful completion percentage to earn bonuses. A 2600k should qualify that passkey in a few days, and as long as your system is stable*, you should have no problem maintaining an 80% completion rate.

Regarding Crossfire, FAH does not use it to have multiple cards work on a single WU, as the latency involved in that link (and SLI, for that matter) is too high to be beneficial. But you don't need to disable it, as the client will happily run one GPU slot on each card.

There's one change I'd have you make to the v7 client's configuration. Once it's installed, pause the SMP slot** and change the view mode from Novice to either Advanced or Expert. After that, go to Configure -> Slots, highlight the SMP slot, hit Edit, change the CPUs field in the SMP area from -1 to 6, then hit OK and Save the changes, and restart the slot(s). The reason for this is that SMP work runs as fast as the slowest thread, and the amount of CPU time that an AMD GPU eats up while folding will interfere with the SMP slot and significantly slow it down. Leaving one thread free for each GPU will result in the SMP work getting done much faster since 6 threads running with no interference is much better than 8 running with major interference.

As a heads up, GPU folding does not usually mix well with other GPU-intensive tasks such as gaming. Unlike CPUs, there is no scheduling software for GPUs, so they have a lot of trouble handling concurrent strenuous tasks. If you're going to be gaming or something, it's recommended that you pause the GPU slots while doing that other task and restart them when you're done.

*If you haven't overclocked, you should be fine. If you have, note that being Prime95/<some other stress tester> stable does not mean you're folding stable, as most stress testing apps doesn't really stress the same modules that F@h does. You might want to run StressCPU v2, as that runs the same Gromacs code that F@h uses and is a better indicator of folding stability.

**Either hit the big Pause button on top, which will pause every slot, or right-click the SMP slot in the Folding Slots window and select Pause to only pause that one

[BornAgainPagan]

Thank You for the info, I'll get it all set up in the next couple of days. I don't overclock much, currently running at 4.0 Ghz, temps never go over 52C, with water cooling.
The video cards are factory overclocked, they are stable and never overheat. After the next WU is done, I'll run StressCPU and make sure everything is running stable.
All 3 of my computers run 24/7, I'm going to install FAH on the 2 P4 boxes soon.
Thanks Again, BAP

[bruce]

Welcome back, BornAgainPagan.

I could be wrong, but you may also have a problem with V7's whitelist. I don't see the HD 6850 listed but maybe I just missed it.

If your GPUs are recognized by V7, you'll be fine. If they are not, we'll have to get them added to the whitelist. See viewtopic.php?f=83&t=21205

[Zagen30]

I think it's listed as 0x1002:0x6739:1:4:Barts PRO [ATI Radeon HD 6800 Series]. On the AMD GPU Wikipedia page, the 6850 is the only card with the codename Barts Pro. Several other 6000 series cards appear to be identified in the whitelist by codename rather than exact model number.

[jrweiss]

Another issue with the ATI/AMD GPUs is that CPU Folding is often diminished, despite the GPU not overtly using more than 2-3% of a CPU core. I noticed that with my old 38xx and 46xx series, the offset was so great that I went back to SMP Folding only. You should monitor your production for a while to make sure you're getting what you expect.

[Nathan_P]

Welcome back!

No need to run stressCPU to test your hardware, folding will do a much better job of pushing your hardware to its limits

[BornAgainPagan]

Another issue with the ATI/AMD GPUs is that CPU Folding is often diminished, despite the GPU not overtly using more than 2-3% of a CPU core. I noticed that with my old 38xx and 46xx series, the offset was so great that I went back to SMP Folding only. You should monitor your production for a while to make sure you're getting what you expect.

Hi, Ok a question or two about diminished CPU core, as in, its going to take longer for the CPU WU to finish with the GPU WU going too? Think I have already noticed that, it was taking the CPU WU's about 12 hr's to finish (depending on size) with V6.23. I'm now running the version 7 with only one of the GPU's running a WU and the CPU WU running, the GPU WU is screaming along, but the CPU WU now says it will take 1.9 days to complete. The GPU will complete its WU in about 6 hr's. CPU WU is using up about 70 - 80% of CPU cycles, GPU WU using pretty much 100% of GPU core. All of the Core temps are ok, GPU 56C and CPU 61C. So I"m not sure which would be better, more WU's thru the GPU and CPU WU"S or more WU"s done faster thru the CPU.

Thanks.

[Zagen30]

Well, first off, when you were running 6.23, you were only running in uniprocessor mode, i.e. only one thread was being used to process the WUs it was getting. Comparing run times there to run times when in SMP mode is rather apples-to-oranges unless they happened to be running the same project. Many projects can now be run in either uniprocessor or SMP mode, so that scenario isn't entirely implausible, but I don't know what you were running before and after the switch. Comparing the run times between GPU and CPU projects is even more inaccurate at this point in time.

Second, some projects are just really long. My 3770k can get certain projects' WUs done in an hour and others' take almost 24 hours. 1.9 days does sound too high, though how far into the WU was it when v7 was making this prediction? v7 has prediction accuracy issues at the start of WUs. If you're, say, 20% of the way through and it hasn't been paused recently, however, then it's probably accurate. Did you change the SMP slot to only use 6 threads? What project is the CPU currently running?

Regarding what's the best way to contribute, most people go based on Points Per Day since points are roughly aligned with scientific value. If smp:6 and 2 GPUs is earning less PPD than smp:8 and no GPUs, many would go with the latter option. AMD GPUs are notoriously poor PPD performers, and while I would think that two of them would produce enough points to surpass the PPD loss from only running 6 instead of 8 CPU threads, that's not from experience.

[jrweiss]

I would pause the GPU WU for a while, and see if the Time Per Frame for the SMP WU decreases significantly. If so, you'll have to look at the total PPD change for running the SMP full-bore, or the SMP+GPU. The bonus point allocation significantly increases for SMP WUs for shorter completion times (make sure you have a passkey and have entered it in the FahControl config); I don't know how the GPU points are allocated. You may also decide at least in part based on the significant power usage for the GPU at full load.

[bruce]

AMD GPUs are notoriously poor PPD performers, and while I would think that two of them would produce enough points to surpass the PPD loss from only running 6 instead of 8 CPU threads, that's not from experience.

The choice between SMP:6+2xAMD-GPUs vs. SMP:8 depends mostly on the architecture of your CPU. AMD's bulldozer architecture and Intel's Hyperthreading architecture can be reconfigured from SMP:8 to SMP:6 with a relatively small loss of SMP PPD which can easily be exceeded by the GPUs. The change from 8 to 8 on other CPU architectures which feature on FPU per CPU will have a much greater impact on SMP performance so the choice will be much more difficult to predict without actually testing.