GPU alone 3x faster than with CPU?

[AveryAndrews]

I am finding that my Phenom II X6 1100T + GTX 750, on an oldish mobo (GA-78LMT-USB3), under W10, gets an estimated 200000 PPD with only the GPU slot enabled, 80,000 with both, apparently regardless of the setting for number of CPU threads (client choose, 3, 4, 5 all seem to produce pretty much the same results). Looking at the # of steps per unit time in the logs seems to corroborate the claim that the GPU does about 3x more work then the CPU is not folding. Is this a known effect or something unusual about the setup? It would presumably be nice to get the full efffect of the GPU with some additional work by the CPU.

Update/Answer: I think the key to the puzzle is given by bruce far below, which is that jobs are given to a range of cores that can get them done in time, so my conclusion is that the GTX 750 is near the top of one range but the bottom of a higher one, since the jobs it gets the Hi PPD on always have a rather small base credit, but it collects lots of bonus points, while the normal PPD ones have much larger base credits, so not so many bonus points.

[JimboPalmer]

Welcome to Folding@Home!

Without seeing the first 200 lines of your log, It is hard to state facts.

viewtopic.php?f=24&t=26036

If you just want guesses, I would choose 4 CPU treads.

The GTX 750 has about the same folding performance as the more modern GT 1030, It should be about 80,000. Does the 200,000 PPD last an entire Work Unit?

If so, to heck with the CPU, take the 200,000! (I get about that with a GTX 1050 ti)

[AveryAndrews]

It tends to float down somewhere in the 170000's. Sometimes I've seen wildly exaggerated PPD estimates even when the job has being going for some time, but this appears to be consistent over several trials. The performance of the GTX with 4 cores running has been 60k-ish. So I suspect some kind of reporting bug rather than a magic GTX that is mysteriously crippled when a CPU is put on the job.

Simply having or putting the CPU on pause does not produce the apparent boost, the slot has to be absent. I haven't tried removing it when a CPU job is done but a GPU job is still running.

Here is the log for the beginning of the currently running job .. this one started at c. 207000PPD and was at c. 176000 when I copied, decaying a bit faster than usual, back up to 183773 a few minutes later.

Code: Select all

*********************** Log Started 2020-08-14T06:51:53Z ***********************
06:51:53:************************* Folding@home Client *************************
06:51:53:        Website: https://foldingathome.org/
06:51:53:      Copyright: (c) 2009-2018 foldingathome.org
06:51:53:         Author: Joseph Coffland <joseph@cauldrondevelopment.com>
06:51:53:           Args: 
06:51:53:         Config: C:\Users\Avery\AppData\Roaming\FAHClient\config.xml
06:51:53:******************************** Build ********************************
06:51:53:        Version: 7.5.1
06:51:53:           Date: May 11 2018
06:51:53:           Time: 13:06:32
06:51:53:     Repository: Git
06:51:53:       Revision: 4705bf53c635f88b8fe85af7675557e15d491ff0
06:51:53:         Branch: master
06:51:53:       Compiler: Visual C++ 2008
06:51:53:        Options: /TP /nologo /EHa /wd4297 /wd4103 /Ox /MT
06:51:53:       Platform: win32 10
06:51:53:           Bits: 32
06:51:53:           Mode: Release
06:51:53:******************************* System ********************************
06:51:53:            CPU: AMD Phenom(tm) II X6 1100T Processor
06:51:53:         CPU ID: AuthenticAMD Family 16 Model 10 Stepping 0
06:51:53:           CPUs: 6
06:51:53:         Memory: 16.00GiB
06:51:53:    Free Memory: 13.66GiB
06:51:53:        Threads: WINDOWS_THREADS
06:51:53:     OS Version: 6.2
06:51:53:    Has Battery: false
06:51:53:     On Battery: false
06:51:53:     UTC Offset: 10
06:51:53:            PID: 11072
06:51:53:            CWD: C:\Users\Avery\AppData\Roaming\FAHClient
06:51:53:             OS: Windows 10 Home
06:51:53:        OS Arch: AMD64
06:51:53:           GPUs: 1
06:51:53:          GPU 0: Bus:1 Slot:0 Func:0 NVIDIA:4 GM107 [GeForce GTX 750] 1111
06:51:53:  CUDA Device 0: Platform:0 Device:0 Bus:1 Slot:0 Compute:5.0 Driver:11.0
06:51:53:OpenCL Device 0: Platform:0 Device:0 Bus:1 Slot:0 Compute:1.2 Driver:451.67
06:51:53:  Win32 Service: false
06:51:53:***********************************************************************
06:51:53:<config>
06:51:53:  <!-- Network -->
06:51:53:  <proxy v=':8080'/>
06:51:53:
06:51:53:  <!-- Slot Control -->
06:51:53:  <power v='full'/>
06:51:53:
06:51:53:  <!-- User Information -->
06:51:53:  <passkey v='********************************'/>
06:51:53:  <user v='Horza49'/>
06:51:53:
06:51:53:  <!-- Folding Slots -->
06:51:53:  <slot id='1' type='GPU'/>
06:51:53:</config>
06:51:53:Trying to access database...
06:51:53:Successfully acquired database lock
06:51:53:Enabled folding slot 01: READY gpu:0:GM107 [GeForce GTX 750] 1111
06:51:53:WU00:FS01:Connecting to 65.254.110.245:8080
06:51:54:WU00:FS01:Assigned to work server 18.188.125.154
06:51:54:WU00:FS01:Requesting new work unit for slot 01: READY gpu:0:GM107 [GeForce GTX 750] 1111 from 18.188.125.154
06:51:54:WU00:FS01:Connecting to 18.188.125.154:8080
06:51:55:WU00:FS01:Downloading 358.50KiB
06:51:56:WU00:FS01:Download complete
06:51:56:WU00:FS01:Received Unit: id:00 state:DOWNLOAD error:NO_ERROR project:13421 run:8462 clone:75 gen:1 core:0x22 unit:0x0000000312bc7d9a5f349dcc0125dafc
06:51:56:WU00:FS01:Starting
06:51:56:WU00:FS01:Running FahCore: "C:\Program Files (x86)\FAHClient/FAHCoreWrapper.exe" C:\Users\Avery\AppData\Roaming\FAHClient\cores/cores.foldingathome.org/win/64bit/22-0.0.11/Core_22.fah/FahCore_22.exe -dir 00 -suffix 01 -version 705 -lifeline 11072 -checkpoint 15 -gpu-vendor nvidia -opencl-platform 0 -opencl-device 0 -cuda-device 0 -gpu 0
06:51:57:WU00:FS01:Started FahCore on PID 10576
06:51:58:WU00:FS01:Core PID:10620
06:51:58:WU00:FS01:FahCore 0x22 started
06:51:58:WU00:FS01:0x22:*********************** Log Started 2020-08-14T06:51:58Z ***********************
06:51:58:WU00:FS01:0x22:*************************** Core22 Folding@home Core ***************************
06:51:58:WU00:FS01:0x22:       Core: Core22
06:51:58:WU00:FS01:0x22:       Type: 0x22
06:51:58:WU00:FS01:0x22:    Version: 0.0.11
06:51:58:WU00:FS01:0x22:     Author: Joseph Coffland <joseph@cauldrondevelopment.com>
06:51:58:WU00:FS01:0x22:  Copyright: 2020 foldingathome.org
06:51:58:WU00:FS01:0x22:   Homepage: https://foldingathome.org/
06:51:58:WU00:FS01:0x22:       Date: Jun 26 2020
06:51:58:WU00:FS01:0x22:       Time: 19:49:16
06:51:58:WU00:FS01:0x22:   Revision: 22010df8a4db48db1b35d33e666b64d8ce48689d
06:51:58:WU00:FS01:0x22:     Branch: core22-0.0.11
06:51:58:WU00:FS01:0x22:   Compiler: Visual C++ 2015
06:51:58:WU00:FS01:0x22:    Options: /TP /nologo /EHa /wd4297 /wd4103 /O2 /Ob3 /Zc:throwingNew /MT
06:51:58:WU00:FS01:0x22:   Platform: win32 10
06:51:58:WU00:FS01:0x22:       Bits: 64
06:51:58:WU00:FS01:0x22:       Mode: Release
06:51:58:WU00:FS01:0x22:Maintainers: John Chodera <john.chodera@choderalab.org> and Peter Eastman
06:51:58:WU00:FS01:0x22:             <peastman@stanford.edu>
06:51:58:WU00:FS01:0x22:       Args: -dir 00 -suffix 01 -version 705 -lifeline 10576 -checkpoint 15
06:51:58:WU00:FS01:0x22:             -gpu-vendor nvidia -opencl-platform 0 -opencl-device 0 -cuda-device
06:51:58:WU00:FS01:0x22:             0 -gpu 0
06:51:58:WU00:FS01:0x22:************************************ libFAH ************************************
06:51:58:WU00:FS01:0x22:       Date: Jun 26 2020
06:51:58:WU00:FS01:0x22:       Time: 19:47:12
06:51:58:WU00:FS01:0x22:   Revision: 2b383f4f04f38511dff592885d7c0400e72bdf43
06:51:58:WU00:FS01:0x22:     Branch: HEAD
06:51:58:WU00:FS01:0x22:   Compiler: Visual C++ 2015
06:51:58:WU00:FS01:0x22:    Options: /TP /nologo /EHa /wd4297 /wd4103 /O2 /Ob3 /Zc:throwingNew /MT
06:51:58:WU00:FS01:0x22:   Platform: win32 10
06:51:58:WU00:FS01:0x22:       Bits: 64
06:51:58:WU00:FS01:0x22:       Mode: Release
06:51:58:WU00:FS01:0x22:************************************ CBang *************************************
06:51:58:WU00:FS01:0x22:       Date: Jun 26 2020
06:51:58:WU00:FS01:0x22:       Time: 19:46:11
06:51:58:WU00:FS01:0x22:   Revision: f8529962055b0e7bde23e429f5072ff758089dee
06:51:58:WU00:FS01:0x22:     Branch: master
06:51:58:WU00:FS01:0x22:   Compiler: Visual C++ 2015
06:51:58:WU00:FS01:0x22:    Options: /TP /nologo /EHa /wd4297 /wd4103 /O2 /Ob3 /Zc:throwingNew /MT
06:51:58:WU00:FS01:0x22:   Platform: win32 10
06:51:58:WU00:FS01:0x22:       Bits: 64
06:51:58:WU00:FS01:0x22:       Mode: Release
06:51:58:WU00:FS01:0x22:************************************ System ************************************
06:51:58:WU00:FS01:0x22:        CPU: AMD Phenom(tm) II X6 1100T Processor
06:51:58:WU00:FS01:0x22:     CPU ID: AuthenticAMD Family 16 Model 10 Stepping 0
06:51:59:WU00:FS01:0x22:       CPUs: 6
06:51:59:WU00:FS01:0x22:     Memory: 16.00GiB
06:51:59:WU00:FS01:0x22:Free Memory: 13.79GiB
06:51:59:WU00:FS01:0x22:    Threads: WINDOWS_THREADS
06:51:59:WU00:FS01:0x22: OS Version: 6.2
06:51:59:WU00:FS01:0x22:Has Battery: false
06:51:59:WU00:FS01:0x22: On Battery: false
06:51:59:WU00:FS01:0x22: UTC Offset: 10
06:51:59:WU00:FS01:0x22:        PID: 10620
06:51:59:WU00:FS01:0x22:        CWD: C:\Users\Avery\AppData\Roaming\FAHClient\work
06:51:59:WU00:FS01:0x22:********************************************************************************
06:51:59:WU00:FS01:0x22:Project: 13421 (Run 8462, Clone 75, Gen 1)
06:51:59:WU00:FS01:0x22:Unit: 0x0000000312bc7d9a5f349dcc0125dafc
06:51:59:WU00:FS01:0x22:Reading tar file core.xml
06:51:59:WU00:FS01:0x22:Reading tar file integrator.xml
06:51:59:WU00:FS01:0x22:Reading tar file state.xml.bz2
06:51:59:WU00:FS01:0x22:Reading tar file system.xml.bz2
06:51:59:WU00:FS01:0x22:Digital signatures verified
06:51:59:WU00:FS01:0x22:Folding@home GPU Core22 Folding@home Core
06:51:59:WU00:FS01:0x22:Version 0.0.11
06:51:59:WU00:FS01:0x22:  Checkpoint write interval: 50000 steps (5%) [20 total]
06:51:59:WU00:FS01:0x22:  JSON viewer frame write interval: 10000 steps (1%) [100 total]
06:51:59:WU00:FS01:0x22:  XTC frame write interval: 250000 steps (25%) [4 total]
06:51:59:WU00:FS01:0x22:  Global context and integrator variables write interval: 25000 steps (2.5%) [40 total]
06:52:08:WU00:FS01:0x22:Completed 0 out of 1000000 steps (0%)
06:53:11:WU00:FS01:0x22:Completed 10000 out of 1000000 steps (1%)
06:54:12:WU00:FS01:0x22:Completed 20000 out of 1000000 steps (2%)
06:55:14:WU00:FS01:0x22:Completed 30000 out of 1000000 steps (3%)
06:56:15:WU00:FS01:0x22:Completed 40000 out of 1000000 steps (4%)
06:57:16:WU00:FS01:0x22:Completed 50000 out of 1000000 steps (5%)
06:58:19:WU00:FS01:0x22:Completed 60000 out of 1000000 steps (6%)
06:59:20:WU00:FS01:0x22:Completed 70000 out of 1000000 steps (7%)
07:00:23:WU00:FS01:0x22:Completed 80000 out of 1000000 steps (8%)
07:01:24:WU00:FS01:0x22:Completed 90000 out of 1000000 steps (9%)
07:02:26:WU00:FS01:0x22:Completed 100000 out of 1000000 steps (10%)
07:03:28:WU00:FS01:0x22:Completed 110000 out of 1000000 steps (11%)
07:04:29:WU00:FS01:0x22:Completed 120000 out of 1000000 steps (12%)
07:05:30:WU00:FS01:0x22:Completed 130000 out of 1000000 steps (13%)
07:06:31:WU00:FS01:0x22:Completed 140000 out of 1000000 steps (14%)
07:07:32:WU00:FS01:0x22:Completed 150000 out of 1000000 steps (15%)
07:08:35:WU00:FS01:0x22:Completed 160000 out of 1000000 steps (16%)
07:09:37:WU00:FS01:0x22:Completed 170000 out of 1000000 steps (17%)

[JimboPalmer]

So far, I see no reason for a slowdown with CPU folding, but I can't argue with results.

Your Nvidia Driver is up to date
06:51:53:OpenCL Device 0: Platform:0 Device:0 Bus:1 Slot:0 Compute:1.2 Driver:451.67

Your Client software is a couple releases back. Current is 7.6.13, i think.
06:51:53: Version: 7.5.1

But mostly, everything i know to look for, looks good.

Studying power draw, your GPU pulls a modest 55 Watts
https://www.techpowerup.com/gpu-specs/g ... -750.c1986

Your motherboard mentions supporting 125 watts.
https://www.gigabyte.com/Motherboard/GA ... -rev-60#ov

Your CPU mentions pulling 125 watts.
https://www.cpubenchmark.net/cpu.php?cp ... 00T&id=394

How big is your Power Supply Unit?

I use GPU-Z to diagnose why my GPUs are limited, their is a line in the Sensors that gives hints, called Perf/Cap Reason.

https://www.techpowerup.com/gpuz/

[Neil-B]

Have no clue what is going on ... but I have a faint recollection there has been another such thread on the forums within the last six months on the same type of issue - I might try finding it later.

Looking at your log the %age time per fold steps are 61-63 secs stable so it isn't that your client/core is folding with variable times - looks like something odd with the estimation - are you using the web or advanced control to look at the estimated times.

Would be work looking at the final actual points accrued for a few WUs - remembering there can be a fair bit of variation between projects (and even within projects for 13421) - which I think will show little differrence between having CPU slot running or not (but I may be wrong).

[MeeLee]

Something doesn't make sense,
The 750 doesn't make 200k PPD. The 750 Ti makes anywhere from 35-80k PPD, from what I read online.
I guess you saw a estimated PPD count, that changed as the counter gets more accurate with every passing % of the WU.
Also, if you pause a WU, the PPD count lowers.

[AveryAndrews]

Thanks for all the suggestions!

After several jobs with 3x realistic estimates, the one it's doing now has a sensible one (c. 53000PPD). I've been looking at the logs, but they are a bit chaotic, so the next time I get a wild PPD estimate I'll manage them carefully. The high estimates can persist to the very end of a job, eg 187423 with 4 minutes to go. I'm aware that PPD can very with the job, but by a factor of 3 seems excessive. In terns of the PPD/Watts ratio, it might be better to keep the CPU slot off anyway, but I think that would partly depend on how urgent you think CV-19 is (pretty damn urgent in my view, which is why I have an [older] computer on it as well is my better one, whose performance seems typical). [older inserted later]

[bruce]

For "traditional" projects (i.e. anything pre-COVID and some early COVID projects) a factor of 3 would be excessive. Because of the nature of many COVID projects, the word "excessive" has lost its meaning.

[BobWilliams757]

Thanks for all the suggestions!

After several jobs with 3x realistic estimates, the one it's doing now has a sensible one (c. 53000PPD). I've been looking at the logs, but they are a bit chaotic, so the next time I get a wild PPD estimate I'll manage them carefully. The high estimates can persist to the very end of a job, eg 187423 with 4 minutes to go. I'm aware that PPD can very with the job, but by a factor of 3 seems excessive. In terns of the PPD/Watts ratio, it might be better to keep the CPU slot off anyway, but I think that would partly depend on how urgent you think CV-19 is (pretty damn urgent in my view, which is why I have an computer on it as well is my better one, whose performance seems typical).

Some of the current work units give big returns on lesser cards, but that still doesn't seem to explain why adding the CPU would slow things down. But in terms of of points(or science) done per watt, GPU is almost always going to be more efficient. In my case with onboard graphics, adding CPU folding often doesn't make sense, as it adds little to throughput, since CPU and GPU use the same memory and resources. But in your case, that CPU should be worth 20k or so more points probably.


I'm over a factor of 4 with GPU folding on project 13421. Here and there freaky fast or slow work units pop up.

[bruce]

...adding CPU folding often doesn't make sense, as it adds little to throughput, since CPU and GPU use the same memory and resources. But in your case, that CPU should be worth 20k or so more points probably.

I'm over a factor of 4 with GPU folding on project 13421.

Project 1342* is not a single project, it's a combination of different projects combined under a single project number. Each Run should have a differnent baseline points setting, but that's not possible. If you process a WU from each run, the average performance should work out about right. When you complain about a single run, nothing can be done except to increase/decrease the points on every other run in the project.

I may be able to explain how adding a CPU slot can reduce your total PPD. Yes, that's countger-intuitive.


The GPU slot uses a CPU thread (or two) for two purposes. First, for NV GPUs, one thread continually polls the GPU to see if the GPU will let it move data to/from main RAM and GPU RAM. On a "traditional" FAH project, the CPU also does a "sanity check" which calculates the free energy to compare with the GPUs free energy total. If they're essentially equal, the GPU is most likey doing a good job. If they're decidedly unequal, the GPU is assumed to be unstable and an error flag is raised and the WU is aborted.

Project 134xx are doing an unusual percentage of free energy calculations so the additional processing has to compete for resources with the WU being processed by the CPU slot.

[gunnarre]

The Phenom II X6 1100T can only sustain boost clocks on 3 of the 6 cores, and due to the reasons bruce outlined above, Nvidia GPUs are very sensitive to CPU frequency - especially on the 134xx projects. If the CPU has a lot more cores, and AVX support, then it might pay off to still fold on the CPU because the GPU performance decrease is made up by the CPU folding. But for a Phenom II X6 feeding an NVidia GPU, you might be limited to 1 or maybe 2 threads of CPU folding before you're downclocked.

What is happening with your CPU clocks and temperature while you fold on both CPU and GPU?

[BobWilliams757]

...adding CPU folding often doesn't make sense, as it adds little to throughput, since CPU and GPU use the same memory and resources. But in your case, that CPU should be worth 20k or so more points probably.

I'm over a factor of 4 with GPU folding on project 13421.

Project 1342* is not a single project, it's a combination of different projects combined under a single project number. Each Run should have a differnent baseline points setting, but that's not possible. If you process a WU from each run, the average performance should work out about right. When you complain about a single run, nothing can be done except to increase/decrease the points on every other run in the project.

I may be able to explain how adding a CPU slot can reduce your total PPD. Yes, that's countger-intuitive.


The GPU slot uses a CPU thread (or two) for two purposes. First, for NV GPUs, one thread continually polls the GPU to see if the GPU will let it move data to/from main RAM and GPU RAM. On a "traditional" FAH project, the CPU also does a "sanity check" which calculates the free energy to compare with the GPUs free energy total. If they're essentially equal, the GPU is most likey doing a good job. If they're decidedly unequal, the GPU is assumed to be unstable and an error flag is raised and the WU is aborted.

Project 134xx are doing an unusual percentage of free energy calculations so the additional processing has to compete for resources with the WU being processed by the CPU slot.

Just for the record, I don't complain about points regardless. They are what they are. If I feel adding a comment might help another folder or researcher, I might comment with my results on said project.... but either way I still fold them.

In my case with onboard graphics it's possible I might not "see" the conflicts others might have. In my case adding CPU is of little value regardless of the project running, even those GPU projects that take very few CPU resources to fold them.

This 13421 has generated a lot of concerns and inconsistencies for many, but really for me it's been fairly consistent. Of 50-55 WU's I've had four or five each either freaky high or slow in PPD return. Other than that they fall within a fairly tight return range. But that seems to be the case for me with some other projects that many report issues with as well. This little rig might be usually slow, but it seems to be really stable and handle just about any WU's that come this way.

[AveryAndrews]

So I think I have gotten some idea as to where the very hi PPD for the GTX 750 are coming from, which is jobs with an unusually high bonus factor for a reasonably prompt result, e.g base credit 3190, estimated credit 133887, job done in a bit under 2 hrs giving a 188000-ish estimated PPD. Whereas a more mormal PPD estimation came from a job with base credit 29000 and estimamed credit 70034.

I had been puzzled as to why I had never seen the former sort of job with CPU slots also folding, but then one came up yesterday (the card had been running in an FX-8350 for several months, but that machine died so I shifted it to the Phenom). So that would appear to have been chance.

So the phenomenon is probably somewhat capricious assignment bonus points, perhaps currently more than previously usual.

[bruce]

So the phenomenon is probably somewhat capricious assignment bonus points, perhaps currently more than previously usual.

Possibly, but it's more likely a capricious assignment of deadlines. The deadlines need to be consistent with the expected completion time given the complexity of the analysis and the speed of the hardware to which the WU is assigned.

All WUs are assigned to a range of hardware. The slowest member of that range should be able to meet the deadline but with some difficulty if it doesn't fold 24x7. The faster memebers of the range should get increasing bonuses. If the deadline is improperly set or the benchmark suggests too many baseline points, the PPD can become excessive.

[AveryAndrews]

I see; perhaps the GTX 750 is at the upper end of a slower equipment range, which would explain why it gets excessive points to a considerably greater extent than all the other cores I have run (assorted feeblsh-ish CPUs, Ryzen 7 3700X, Radeon RX-580). The Radeon sometimes gets excessive points, but not to the same extent.