low GPU utilization

[doune]

I've got a 4930k (6 physical cores, 12 with Hyperthreading turned on) with an estimated PPD of about 55k+ and it regularly sits at 11/12 CPU cores on folding (fahcore_a7) with 1/12 left to feed the GPU folding task (fahcore_a21.) This means my CPU is pretty much at 100% while dedicated to folding. That's great.

However, my GPU seems to hover around 5-8%. That seems low. The estimated PPD is about the same 55k as my CPU and, I would assume, the GPU number would normally be a lot higher.

So, is this low utilization normal? If not, what's the most obvious thing to check to make it do more science?

[toTOW]

We need more details about your setup. Here's a good place to start : How to provide enough info to get helpful support..

[JimboPalmer]

"what's the most obvious thing to check to make it do more science?"
Try to determine the manufacturer of your GPU (AMD or Nvidia) and then download the latest driver directly from that manufacturer.

They may bundle a lot of software with the driver, but all you need is the driver unless the OpenCL software is separate, then you need that too.

[doune]

It's a GTX 660. Drivers are relatively current, I usually get motivated to update them a few times a year since I refuse to install GeForce Experience now that it requires being logged in.

Code: Select all

*********************** Log Started 2018-07-11T23:37:51Z ***********************
23:37:51:************************* Folding@home Client *************************
23:37:51:      Website: http://folding.stanford.edu/
23:37:51:    Copyright: (c) 2009-2014 Stanford University
23:37:51:       Author: Joseph Coffland <joseph@cauldrondevelopment.com>
23:37:51:         Args: 
23:37:51:       Config: F:/FAHClient/config.xml
23:37:51:******************************** Build ********************************
23:37:51:      Version: 7.4.4
23:37:51:         Date: Mar 4 2014
23:37:51:         Time: 20:26:54
23:37:51:      SVN Rev: 4130
23:37:51:       Branch: fah/trunk/client
23:37:51:     Compiler: Intel(R) C++ MSVC 1500 mode 1200
23:37:51:      Options: /TP /nologo /EHa /Qdiag-disable:4297,4103,1786,279 /Ox -arch:SSE
23:37:51:               /QaxSSE2,SSE3,SSSE3,SSE4.1,SSE4.2 /Qopenmp /Qrestrict /MT /Qmkl
23:37:51:     Platform: win32 XP
23:37:51:         Bits: 32
23:37:51:         Mode: Release
23:37:51:******************************* System ********************************
23:37:51:          CPU: Intel(R) Core(TM) i7-4930K CPU @ 3.40GHz
23:37:51:       CPU ID: GenuineIntel Family 6 Model 62 Stepping 4
23:37:51:         CPUs: 12
23:37:51:       Memory: 15.94GiB
23:37:51:  Free Memory: 14.04GiB
23:37:51:      Threads: WINDOWS_THREADS
23:37:51:   OS Version: 6.2
23:37:51:  Has Battery: false
23:37:51:   On Battery: false
23:37:51:   UTC Offset: -5
23:37:51:          PID: 10584
23:37:51:          CWD: F:/FAHClient
23:37:51:           OS: Windows 10 Enterprise
23:37:51:      OS Arch: AMD64
23:37:51:         GPUs: 1
23:37:51:        GPU 0: NVIDIA:3 GK106 [GeForce GTX 660]
23:37:51:         CUDA: 3.0
23:37:51:  CUDA Driver: 9020
23:37:51:Win32 Service: false
23:37:51:***********************************************************************
23:37:51:<config>
23:37:51:  <!-- Slot Control -->
23:37:51:  <power v='FULL'/>
23:37:51:
23:37:51:  <!-- User Information -->
23:37:51:  <passkey v='********************************'/>
23:37:51:  <team v='233892'/>
23:37:51:  <user v='morphal'/>
23:37:51:
23:37:51:  <!-- Folding Slots -->
23:37:51:  <slot id='0' type='CPU'/>
23:37:51:  <slot id='1' type='GPU'/>
23:37:51:</config>
23:37:51:Trying to access database...
23:37:51:Successfully acquired database lock
23:37:51:Enabled folding slot 00: READY cpu:11
23:37:51:Enabled folding slot 01: READY gpu:0:GK106 [GeForce GTX 660]
23:37:51:WU02:FS01:Starting
23:37:51:WU02:FS01:Running FahCore: "F:\Program Files (x86)\FAHClient/FAHCoreWrapper.exe" F:/FAHClient/cores/cores.foldingathome.org/Win32/AMD64/NVIDIA/Fermi/Core_21.fah/FahCore_21.exe -dir 02 -suffix 01 -version 704 -lifeline 10584 -checkpoint 15 -gpu 0 -gpu-vendor nvidia
23:37:51:WU02:FS01:Started FahCore on PID 10780
23:37:51:WU02:FS01:Core PID:10824
23:37:51:WU02:FS01:FahCore 0x21 started
23:37:51:WU00:FS00:Starting
23:37:51:WU00:FS00:Running FahCore: "F:\Program Files (x86)\FAHClient/FAHCoreWrapper.exe" F:/FAHClient/cores/cores.foldingathome.org/Win32/AMD64/AVX/Core_a7.fah/FahCore_a7.exe -dir 00 -suffix 01 -version 704 -lifeline 10584 -checkpoint 15 -np 11
23:37:51:WU00:FS00:Started FahCore on PID 10844
23:37:51:WU00:FS00:Core PID:10868
23:37:51:WU00:FS00:FahCore 0xa7 started
23:37:52:WU02:FS01:0x21:*********************** Log Started 2018-07-11T23:37:51Z ***********************
23:37:52:WU02:FS01:0x21:Project: 11713 (Run 11, Clone 483, Gen 100)
23:37:52:WU02:FS01:0x21:Unit: 0x000000848ca304e75adf776e8755791e
23:37:52:WU02:FS01:0x21:CPU: 0x00000000000000000000000000000000
23:37:52:WU02:FS01:0x21:Machine: 1
23:37:52:WU02:FS01:0x21:Digital signatures verified
23:37:52:WU02:FS01:0x21:Folding@home GPU Core21 Folding@home Core
23:37:52:WU02:FS01:0x21:Version 0.0.18
23:37:52:WU02:FS01:0x21:  Found a checkpoint file
23:37:52:WU00:FS00:0xa7:*********************** Log Started 2018-07-11T23:37:51Z ***********************
23:37:52:WU00:FS00:0xa7:************************** Gromacs Folding@home Core ***************************
23:37:52:WU00:FS00:0xa7:       Type: 0xa7
23:37:52:WU00:FS00:0xa7:       Core: Gromacs
23:37:52:WU00:FS00:0xa7:    Website: https://foldingathome.org/
23:37:52:WU00:FS00:0xa7:  Copyright: (c) 2009-2018 foldingathome.org
23:37:52:WU00:FS00:0xa7:     Author: Joseph Coffland <joseph@cauldrondevelopment.com>
23:37:52:WU00:FS00:0xa7:       Args: -dir 00 -suffix 01 -version 704 -lifeline 10844 -checkpoint 15 -np
23:37:52:WU00:FS00:0xa7:             11
23:37:52:WU00:FS00:0xa7:     Config: <none>
23:37:52:WU00:FS00:0xa7:************************************ Build *************************************
23:37:52:WU00:FS00:0xa7:    Version: 0.0.17
23:37:52:WU00:FS00:0xa7:       Date: Apr 27 2018
23:37:52:WU00:FS00:0xa7:       Time: 16:19:36
23:37:52:WU00:FS00:0xa7: Repository: Git
23:37:52:WU00:FS00:0xa7:   Revision: 21359963583d09ec2063ef946399441c4df4ccd7
23:37:52:WU00:FS00:0xa7:     Branch: master
23:37:52:WU00:FS00:0xa7:   Compiler: Visual C++ 2008
23:37:52:WU00:FS00:0xa7:    Options: /TP /nologo /EHa /wd4297 /wd4103 /Ox /MT
23:37:52:WU00:FS00:0xa7:   Platform: win32 10
23:37:52:WU00:FS00:0xa7:       Bits: 64
23:37:52:WU00:FS00:0xa7:       Mode: Release
23:37:52:WU00:FS00:0xa7:       SIMD: avx_256
23:37:52:WU00:FS00:0xa7:************************************ System ************************************
23:37:52:WU00:FS00:0xa7:        CPU: Unknown
23:37:52:WU00:FS00:0xa7:     CPU ID: 
23:37:52:WU00:FS00:0xa7:       CPUs: 12
23:37:52:WU00:FS00:0xa7:     Memory: 15.94GiB
23:37:52:WU00:FS00:0xa7:Free Memory: 13.95GiB
23:37:52:WU00:FS00:0xa7:    Threads: WINDOWS_THREADS
23:37:52:WU00:FS00:0xa7: OS Version: 6.2
23:37:52:WU00:FS00:0xa7:Has Battery: false
23:37:52:WU00:FS00:0xa7: On Battery: false
23:37:52:WU00:FS00:0xa7: UTC Offset: -5
23:37:52:WU00:FS00:0xa7:        PID: 10868
23:37:52:WU00:FS00:0xa7:        CWD: F:\FAHClient\work
23:37:52:WU00:FS00:0xa7:         OS: Windows 10 Pro
23:37:52:WU00:FS00:0xa7:    OS Arch: AMD64
23:37:52:WU00:FS00:0xa7:********************************************************************************
23:37:52:WU00:FS00:0xa7:Project: 13811 (Run 0, Clone 1660, Gen 153)
23:37:52:WU00:FS00:0xa7:Unit: 0x000000ae80fccb025ac12c98b168d479
23:37:52:WU00:FS00:0xa7:Digital signatures verified
23:37:52:WU00:FS00:0xa7:Reducing thread count from 11 to 10 to avoid domain decomposition by a prime number > 3
23:37:52:WU00:FS00:0xa7:Calling: mdrun -s frame153.tpr -o frame153.trr -x frame153.xtc -cpt 15 -nt 10
23:37:52:WU00:FS00:0xa7:Steps: first=38250000 total=250000
23:37:54:WU00:FS00:0xa7:Completed 1 out of 250000 steps (0%)
23:37:57:WU02:FS01:0x21:Completed 5750000 out of 7500000 steps (76%)
23:37:57:WU02:FS01:0x21:Temperature control disabled. Requirements: single Nvidia GPU, tmax must be < 110 and twait >= 900

[toTOW]

How do yo measure GPU load ? If you're using Windows Task Manager it's not showing FAH load by default. Use GPUZ instead or select Compute instead of 3D in Windows Task Manager.

[bruce]

I've got a 4930k (6 physical cores, 12 with Hyperthreading turned on) with an estimated PPD of about 55k+ and it regularly sits at 11/12 CPU cores on folding (fahcore_a7) with 1/12 left to feed the GPU folding task (fahcore_a21.) This means my CPU is pretty much at 100% while dedicated to folding. That's great.

FAHCore_a7 will not use 11 CPUs. Although your configuration may be set for 11, it will only use 10.

23:37:52:WU00:FS00:0xa7:Reducing thread count from 11 to 10 to avoid domain decomposition by a prime number > 3

I suspect 1 CPU is busy moving data to/from the GPU running FAHCore_21 and the other is idle or doing something else.

FAHCore_a7 cannot process with what are called "large primes" and 11 falls into that category so it is automatically reduce to 10. If your 12-thread CPU is reporting 100%, then temporarily pause that WU and CPU processing will be reduced by 10 threads. Then see what other processes are still running.

[JimboPalmer]

Random comments, only one is actual help.

There is a newer client 7.5.1 it will not help this
https://foldingathome.org/start-folding/
There are newer Nvidia drivers. Also won't make a difference. (all you need is the video driver, de-select the other software)
https://www.geforce.com/drivers/results/135676
Your CPU is a generation called Haswell-E, that runs the newest F@H CPU code in the fastest way, as it supports AVX2.
https://en.wikipedia.org/wiki/Advanced_ ... tensions_2
You have 12 CPU threads available and one is reserved to support the GPU. Since 11 is Prime and F@H hates large Primes and multiples of Primes, 11 is reduced to 10 used by core_a7
I do not see a Passkey. Getting a passkey will not do more science, but after 10 WUs, you will get more points.
https://apps.foldingathome.org/getpasskey.py
Your GPU is a generation called Kepler. That is the oldest generation still supported by Nvidia and by F@H. You may wish to plan for a newer GPU down the road.
https://en.wikipedia.org/wiki/Kepler_(m ... hitecture)

[Joe_H]

Your GPU is a generation called Kepler. That is the oldest generation still supported by Nvidia and by F@H. You may wish to plan for a newer GPU down the road.

The folding client still supports one older generation of nVidia GPU's. Those are based on the Fermi architecture. However low end Fermi and Kepler cards will no longer work with GPU WU's currently available as the cards do not support double precision calculations. A small portion of the calculations require the extra precision for the WU processing.

The GTX 660 does support use of double precision.

[doune]

@toTOW nailed it.

I was using Task Manager and I didn't even realize you could change the view. However, looking at those other views raises the next question.

Compute_0 is at 100%, woohoo! Any idea why Compute_1 is idle?

[doune]

Eh, I'm going to chalk it up to Task Manager having superfluous display options.

GPU-Z does show a flat GPU load of nearly 100% so I'm going to assume everything is running as fast as it can. It still seems weird that the GPU and CPU have the same PPD but the universe is a weird place.

[JimboPalmer]

It still seems weird that the GPU and CPU have the same PPD but the universe is a weird place.

1) You have a fairly recent, powerful CPU with 10 threads.
2) you have a fairly ancient, midrange at the time, GPU.
3) you do not use a PassKey.

[doune]

It still seems weird that the GPU and CPU have the same PPD but the universe is a weird place.

1) You have a fairly recent, powerful CPU with 10 threads.
2) you have a fairly ancient, midrange at the time, GPU.
3) you do not use a PassKey.

The GPU came out in 2012 and the CPU came out in 2013; they're both ancient.
I do use a PassKey. I can see it in the UI and in the log.

Code: Select all

23:37:51:  <!-- User Information -->
23:37:51:  <passkey v='********************************'/>
23:37:51:  <team v='233892'/>
23:37:51:  <user v='morphal'/>

A new GPU will happen, probably within the next few months; it's been coming for a while now. I just hear the mantra of "CPU sucks for folding, GPU is amazing" and I was surprised that their actual output was so similar. When I saw the "8%" utilization, I thought I had discovered the reason why. Now that I see it's actually running at full load, I can rest easy until I get motivated for the upgrade.

[JimboPalmer]

Time is passing slowly for the CPUs, everything newer than a Pentium III can run a core_a4 WU, core_a7 as well. but it will be much faster on a CPU with AVX2 which was introduced with Haswell.
A recent update of Windows 7 finally will not run on a Pentium III.

On the video card side, your Kepler was replaced by Maxwell, which was replaced by Pascal, which was replaced by Volta. Any day now Turing will be out.
Nvidia has ended support for 32 bit OSes, and the card just before yours, Fermi. Time is passing faster for video cards.

(My old fleet of Pentiums and Core Duos produces 60,000 PPD over 17 CPUs, no AVX2. The one GTX 1050ti does 150,000 PPD, and it is a low profile card with no power connectors, low to midrange Pascal)

[Nert]

My old fleet of Pentiums and Core Duos produces 60,000 PPD over 17 CPUs, no AVX2. The one GTX 1050ti does 150,000 PPD, and it is a low profile card with no power connectors, low to midrange Pasca

One thing that influenced my decision to re-enable CPU folding on my two systems was that not all diseases are being researched on GPU's. I might be wrong, but I think that Alzheimers is only being researched on CPU A4 type units (bryostatin compound ?). Total ppd are about the same for me with or without CPU folding (~4 million). Only 25,000 or so are coming from the CPU, but I get satisfaction in knowing that I'm helping research a disease that got me involved with all of this in the first place. When I first turned on CPU folding, I used -1 for the number of CPU threads. Total ppd actually dropped a small percentage from what I was getting with GPU folding alone. Limiting the he number of CPU threads to 2 seems to have fixed that.

[foldy]

@JimboPalmer: Just a technical detail: The core_a7 can use AVX not AVX2. AVX is supported since Intel Sandy Bridge or AMD Bulldozer CPUs. If AVX is not available then SSE is used like FAHCore_a4 for Intel Core2xxx CPUs.