Advanced search

Message boards : Graphics cards (GPUs) : GTX 770 GPU work time disparity

Author Message
scotth
Send message
Joined: 26 Feb 13
Posts: 7
Credit: 1,597,754,724
RAC: 496,757
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 40167 - Posted: 14 Feb 2015 | 0:48:07 UTC
Last modified: 14 Feb 2015 | 1:03:54 UTC

I have a few computers crunching for GPUGRID and recently obtained a second GeForce Windforce GTX 770 2GB card. I have one of the cards in my windows based game rig and one card in a Ubuntu (desktop) server (dual socket supermicro). The Ubuntu (14.04 LTS) server rig has been running since about Jan 18th with the GTX 770. I have been tracking their GPUGrid averages through BoincTasks and see the Linux based rig is taking significantly longer, with less average credit, to complete a cuda65 task than my windows rig (1.63 days vs 0.78 days).

Both hosts have up to date drivers and show the cards running at their standard rate (no overclock) of about 1240mhz.

Can anyone provide insight into this discrepancy? Pointers to something else I should check into?

Windows host (GTX 770 (2048MB) driver: 347.9)
http://www.gpugrid.net//show_host_detail.php?hostid=146739

Ubuntu host (GTX 770 (2047MB) driver: 346.35)
http://www.gpugrid.net//show_host_detail.php?hostid=170946
____________

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2058
Credit: 15,019,398,669
RAC: 5,333,772
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 40172 - Posted: 14 Feb 2015 | 19:22:58 UTC - in response to Message 40167.
Last modified: 14 Feb 2015 | 19:23:42 UTC

Both hosts have up to date drivers and show the cards running at their standard rate (no overclock) of about 1240mhz.

According to NVidia's specifications, the GTX770 has 1046/1085MHz core clock, so this card is overclocked (by the manufacturer perhaps).

Can anyone provide insight into this discrepancy? Pointers to something else I should check into?

I think your slower card is downclocked itself, for self-protecting reasons.
I don't know if there's a Linux tool to check this condition, but a restart would resolve this for a while.
To avoid this to happen again, you should (if the first doesn't help try the second etc)
1. increase the card's cooling (increase fan speed, open the PC's case)
2. decrease the GPU's core clock
3. increase the GPU's voltage by 25mV (raising its power target at the same time)

biodoc
Send message
Joined: 26 Aug 08
Posts: 160
Credit: 1,405,920,847
RAC: 14
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 40173 - Posted: 14 Feb 2015 | 21:08:09 UTC

If you'd like to manually control the fan speed and overclock/underclock in linux add

Option "Coolbits" "12"

to the device section of your xorg.conf file and reboot. The file is located here: /etc/X11/xorg.cong

Here's an old link but it's still valid.

http://www.phoronix.com/scan.php?px=MTY1OTM&page=news_item

scotth
Send message
Joined: 26 Feb 13
Posts: 7
Credit: 1,597,754,724
RAC: 496,757
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 40177 - Posted: 14 Feb 2015 | 23:28:24 UTC - in response to Message 40173.
Last modified: 14 Feb 2015 | 23:29:14 UTC

Apologies for not being more specific. Both 770's are running on Gigabyte's default settings. I haven't overclocked them further.

The linux rig 770 is currently clocking at 1254 Mhz with a memory rate of 7010 Mhz. I already have Coolbits 12 enabled from the get go and the fan is set at 70%. As it lives in the basement where it's cooler, the temps hover around 44-45c under load.

My windows rig 770 is clocking at 1241 Mhz with a memory rate of 7012Mhz. The fan is on an auto curve running at around 68% with a temperature of 58c under full load.

Unless there is something else I can check, I may swap the cards and see what happens over a few weeks.
____________

biodoc
Send message
Joined: 26 Aug 08
Posts: 160
Credit: 1,405,920,847
RAC: 14
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 40178 - Posted: 15 Feb 2015 | 11:53:06 UTC
Last modified: 15 Feb 2015 | 11:55:36 UTC

I have a dual boot machine (win 8.1 and mint linux) with a GTX980 and the latest drivers for both OS's. The WU completion times might be a bit faster on Windows. Just a few percent at the most.

Could you post GPU loads and Memory/PCIe loads for both machines? Also, do you have a free core on both machines to feed the GPU? Is it possible you are processor limited on the linux machine? It's a fairly old processor (L5420/socket LGA771) running at 2.5 GHz.

scotth
Send message
Joined: 26 Feb 13
Posts: 7
Credit: 1,597,754,724
RAC: 496,757
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 40180 - Posted: 16 Feb 2015 | 14:50:44 UTC - in response to Message 40178.
Last modified: 16 Feb 2015 | 14:53:06 UTC

Thanks for your suggestion biodoc.

It does look like my Linux rig 770 was CPU starved. I also identified the issue on another Linux rig that has a PNY 650ti (same motherboard type). If I suspend my Rosetta workload on both rigs, and for good measure, restart the boinc client, the current gpugrid task takes a greater CPU percentage. I also saw a 5-6c temp increase on the 770 and the 650ti. Certainly looks like they were hungrier.

I created a gpugrid app_config.xml to specify the GPU and CPU usage. I currently have two of my Linux rigs with a CPU reservation of one full core for the GPU. I can reduce that down to a .5 core after seeing how everything performs after awhile.

Before finding this out, I did get on Ebay and found that prices are not bad for matched sets of Xeon X5460s that run at 3.16Ghz. I may upgrade the CPUs anyway (Tim Taylor grunting in the background "More power!").
____________

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 176,610
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 40219 - Posted: 19 Feb 2015 | 21:24:27 UTC - in response to Message 40180.

I wouldn't bother to buy CPU upgrades for such an old system. You may run into PCIe bandwidth limitations anyway. Unless there's a good reason not to do so
I'd rather replace the machine with a recent quad core. Or even better: the new Skylake, which should arrive around fall. Performane, power consumption and performance per watt will be better. It's a bit expensive for sure, but if you're going to spend money on that old box anyway.. make it count.

MrS
____________
Scanning for our furry friends since Jan 2002

scotth
Send message
Joined: 26 Feb 13
Posts: 7
Credit: 1,597,754,724
RAC: 496,757
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 40382 - Posted: 6 Mar 2015 | 16:28:17 UTC
Last modified: 6 Mar 2015 | 16:31:28 UTC

After creating the app_config.xml, reserving a core for the GPU, GPUGrid is running great on all my Linux hosts. I've seen their complete times drop for the GPUGrid applications. My Linux host with the 770 is now outpacing the Windows 770 by one hour for the GERARD_CXCL12 application, 9h 50m vice 10h 50m. My other two Linux hosts, each with a 650ti, have also seen an increase in their averages.


Windows host (GTX 770 (2048MB) driver: 347.9)
http://www.gpugrid.net//show_host_detail.php?hostid=146739

Ubuntu host (GTX 770 (2047MB) driver: 346.35)
http://www.gpugrid.net//show_host_detail.php?hostid=170946
____________

Post to thread

Message boards : Graphics cards (GPUs) : GTX 770 GPU work time disparity