nVidia GPU errors on Gravitational Wave search O2
4 Mar 2020 20:39:19 UTC
Topic 220898
(moderation:
I built a new computer from cheap parts and have been testing it. https://einsteinathome.org/host/12809649 I'm getting errors from the Gravitational Wave search O2 Multi-Directional GPU v2.07 () windows_x86_64. From reading some posts I'm guessing my problem is only having 1GB for GPU memory. Do people agree that is the likely problem?
It took awhile but I finally found the preferences where I can set what applications to allow to run. Are there any other applications I should avoid?
Thank you
PS. I've mainly been running SETI @ Home for the past 20+ years but I'll probably be switching to Einstein or another project soon.
deadmann wrote:From reading
)
My RX 570 (Win7 64-bit) is using only 287 MB for the GW, so I think it should work.
But if all else fails, put your GTX 750 on GPUGrid. My GTX 750 Ti (on Ubuntu 18.04) is doing their work units in around three hours. Of course, it is not the fastest card, but is still one of the most efficient ones around. I am comparing it to a GTX 1060 now, and it is doing very well in output/watt.
deadmann wrote:I built a new
)
Hi Deadmann,
Welcome to Einstein.
Unfortunately the link you supplied is not working properly. Here is one that works.
You should be able to use a GTX 750 with 1GB. It probably won't work with the GW GPU search but it certainly will with the FGRP (gamma-ray pulsar) search. Unfortunately it will be quite slow but it will work. Both searches will work with something more recent. If you have access to 2nd hand stuff, something like an AMD RX 570 will really fly. Those are quite cheap, particularly used and would give you about 7 times the output of a GTX 750. Since Einstein only uses OpenCL for GPU apps and not CUDA, AMD has a bit of an advantage.
I took a look at the stderr output returned by one of your failed GW GPU tasks. It seems like there was some sort of memory error transferring stuff from host memory to GPU memory. Just go with the FGRP GPU app if you want to keep using the GTX 750.
You should choose just one GPU search. The gamma-ray pulsar search will work best for you. By default, all searches may be enabled. Until you get some experience, just choose one CPU search and one GPU search. You could start with just the GW CPU search for your CPU threads and the FGRPB1G search (Fermi gamma-ray pulsar) for your 750.
Cheers,
Gary.
Jim1348: It's possible that
)
Jim1348:
It's possible that the memory differences between nVidia and AMD could be in differences in OpenCL. Or in how the code is optimized. I had CPU-Z running and the memory usage kept going up till it errored out. The Gamma-ray pulsar binary search #1 running now is about 965MB now.
Gary
I've been a bit put off about AMD products. I have 2 computers (I don't need both) that are socket FM2 and have have APUs. It's been difficult to find the correct drivers for them on AMD's web site (at least the last few times I've looked. you have to look by product line. not chip model number). I also find it annoying that if the GPU portion of the APU gets busy then it down clocks the CPU portion. I guess this is a feature to keep the APU within the 65W total power consumption. (In hindsight I should have bought a 90W APU.)
I'll keep a look out for a AMD RX 570. I see the "buy it now" ones on ebay are about $90 shipped.
I'm trying to decide if this computer is reliable enough for me to use as a main computer / server. The eithernet and USB 3.0 have given me some problems. It's a new cheap Chinese X79 board. (check youtube videos on X79 boards) All the other parts are used. Installing the realtek drivers seems to have fixed the net.
Thank you both for your help / advice.
PS. In case people have missed the announcement: SETI @ Home will stop distributing work and will go into hibernation March 31.
deadmann wrote:Jim1348: It's
)
Try here for older AMD/ATI drivers: http://www.hal6000.com/seti/boinc_ati_gpu_cheat_sheet.htm
You have to remember that
)
You have to remember that Nvidia cards can only access to OpenCL is only 27% of what the card list. AMD is more around 67%. So if you're card is already starting with a low RAM count, then it's very probable that you are running out of RAM.
I'm quite baffled by those
)
I'm quite baffled by those recent computational errors. About two weeks ago I've been able to successfully process GW02 searches with my GTX 780 and now almost every single search fails. I've tried all sorts of fixes: doing single searches only, resetting the entire project and finally removing it completely from BOINC Manager and adding again. I suspect the recent Nvidia drivers for my card may have something to do with this situation.
drencrome wrote:I'm quite
)
Sometimes cards fail. I have had several go out on me, including a couple of GTX 750 Ti (out of a total of six).
Might be of interest:Just
)
Might be of interest:
Just switched my NVIDIA GTX 750Ti (2GB) small air cooled desktop rig (WIN7 Intel i7 6700) from doing
GPU Gamma-Ray Pulsary Binary Search #1 WUs
to
GPU Gravitational Wave Search O2 Multi-Directional WUs.
Here are my GPU GTX 750Ti (no OC) approximate values:
Gamma-Ray Gravtational Wave GW Vela
run-time <70 Min <25 Min <40 Min
power <30 Watt <30 Watt <35 Watt
mem <770 MB <900 MB <1790 MB <------ !!
load <100% Load <90% Load <99% Load
temp <55°C Temp <65°C Temp <66°C Temp
credit 3464 Points 1000 Points 2000 Points
Room temperatur 27°C.
The GTX 750 has only 1GB.
GR example:
LATeah1063L09_412.0_0_0.0_2381260_1
GW example:
h1_1261.30_O2C02Cl4In0__O2MDFG2f_G34731_1261.60Hz_109_4
GW Vela example:
h1_1261.30_O2C02Cl4In0__O2MDFV2f_VelaJr1_1262.00Hz_248_1
Never had any problems with EatH using my "low end" GTX 750Ti and/or GTX 960 GPUs!
AND I really don't care if the developers user CUDA or OpenCL - they surely have/had there reasons for devloping the WUs as they are.
As long as all is running fine, let's just happily crunch!