Relative Performance of GPU models

[Caprio]

Hello,

Check the link at seti:
http://setiathome.berkeley.edu/gpu_list.php

What strikes me that actually the older 5 generation outperforms the newer 6 generation of Nvidia GPUs for seti work.


14.(0.157) GeForce GTX 670
5.(0.670) GeForce GTX 560 Ti

Now the 560Ti outperforms the GTX670 by a wide margin if this benchmark is correct. Why is the 6-series not doing so well?



1.(1.000) GeForce GTX 580
2.(0.992) GeForce GTX 680
3.(0.753) GeForce GTX 570
4.(0.725) GeForce GTX 480
5.(0.670) GeForce GTX 560 Ti
6.(0.647) GeForce GTX 460
7.(0.366) Quadro 4000
8.(0.359) GeForce GTX 560
9.(0.355) GeForce GTX 550 Ti
10.(0.307) GeForce GTX 275
11.(0.306) GeForce GTX 285
12.(0.284) GeForce 8800 Ultra
13.(0.182) Quadro 1000M
14.(0.157) GeForce GTX 670
15.(0.145) GeForce 9800 GT
16.(0.131) GeForce GTS 250
17.(0.125) GeForce GT 430
18.(0.098) GeForce GTS 450
19.(0.084) GeForce GT 520
20.(0.080) GeForce GT 220
21.(0.069) GeForce 8600 GT
22.(0.066) GeForce 9500 GT

[Jason G]

Hi Caprio,
   The metrics used for that list I haven't looked into in detail, and have noticed it shifts a lot, not necessarily making a lot of sense (to me anyway).  Better to compare with something like Fred's benching tool.  Part of the situation is you have to remember the bulk of any project numbers will come from stock applications, for which (6.10) I believe Kepler GPUs recently have been knobbled with a driver incompatibility with those old apps,.

Both 560ti & 670 are great cards, and I believe with existing apps configured optimally should be capable of producing similar work, with the 670 likely outproducing the 560ti somewhat with apps being released soon.

Next there are significant power & thermal differences between the architectures, making the total cost of ownership of 6 series cards potentially far lower, ( and so arguably better).

When comparing cross generation, as well as maturity of the applications, you'll also need to take into account possible driver & library ongoing refinement.  In general, a pattern of around a year from initial card release sees this stabilise a bit.  560ti took longer to 'sort out' because of some rather radical architecture shifts based on 460s

HTH
Jason

[Caprio]

Hello,

Picking the right card is always a bit tricky. Now one of my 560Ti card had the main fan bearing failure. The dealer proposed to change the Asus 560Ti TOP (900MHz factory clocked) card for a Asus 570 instead. However,  the 570 is a 3 slot card, which I don't like. I could make a counter proposal for a 6xx series card in similar range. But I would still want the possible replacement to have at least better performance on SETI. I do also like the closed card configuration, which blows out the heat on the back, because if you start stockpiling several cards in the same case the internals get very hot....

As said previously the 680 did not increase the performance (with 306.23 drivers) significantly compared to 560Ti (with 285.xx drivers).

[Jason G]

As said previously the 680 did not increase the performance (with 306.23 drivers) significantly compared to 560Ti (with 285.xx drivers).

I'm not so sure about that.  My 680 on old Core2Duo is fairly handily outsripping my Superclocked 560ti (on quad core i5-2400) , with less noise, heat & power.  What are you using to compare performance of these cards & which app ?

[arkayn]

I think my 670 is doing just as well, especially when you consider that I am also gaming on this computer.

[Caprio]

I'm not so sure about that.  My 680 on old Core2Duo is fairly handily outsripping my Superclocked 560ti (on quad core i5-2400) , with less noise, heat & power.  What are you using to compare performance of these cards & which app ?

I am comparing these 2 hosts:
http://setiathome.berkeley.edu/show_host_detail.php?hostid=3806561
http://setiathome.berkeley.edu/show_host_detail.php?hostid=3299266

The latter had 560Ti (900Mhz) replaced with GTX680 on 17th of September 2012.
The RAC has been dropping a bit but no significant rise in last week. Maybe it will take a month for the results to get posted from my wingmates.

http://boincstats.com/en/stats/0/host/list/0/0/35523

Total tasks:
Host 3806561: Core2Duo  E6600 + 560Ti = 4578 RAC 22,497.36
Host 3299266: Core2Quad Q6600 + 680  = 5309 RAC 22,766.58


Now there is some problem with Seti also:

23/09/2012 09:42:34 | SETI@home | Sending scheduler request: Requested by user.
23/09/2012 09:42:34 | SETI@home | Reporting 532 completed tasks, not requesting new tasks
23/09/2012 09:42:34 | SETI@home | [sched_op] CPU work request: 0.00 seconds; 0.00 devices
23/09/2012 09:42:34 | SETI@home | [sched_op] NVIDIA work request: 0.00 seconds; 0.00 devices
23/09/2012 09:43:07 | SETI@home | Temporarily failed upload of 30my12aa.12635.1435.9.10.69_1_0: connect() failed
23/09/2012 09:43:07 | SETI@home | Backing off 3 min 43 sec on upload of 30my12aa.12635.1435.9.10.69_1_0
23/09/2012 09:43:07 |  | [work_fetch] Request work fetch: project finished uploading
23/09/2012 09:43:09 |  | Project communication failed: attempting access to reference site
23/09/2012 09:43:10 |  | Internet access OK - project servers may be temporarily down.
23/09/2012 09:43:24 | SETI@home | Started upload of 30my12aa.12635.1435.9.10.32_1_0
23/09/2012 09:43:28 | SETI@home | Scheduler request completed
23/09/2012 09:43:28 | SETI@home | [sched_op] Server version 701
23/09/2012 09:43:28 | SETI@home | Project requested delay of 303 seconds
23/09/2012 09:43:28 | SETI@home | [sched_op] handle_scheduler_reply(): got ack for task 05jl12aa.30680.16836.11.10.195_1
23/09/2012 09:43:28 | SETI@home | [sched_op] handle_scheduler_reply(): got ack for task 05jl12aa.30680.16836.11.10.180_0
23/09/2012 09:43:28 | SETI@home | [sched_op] handle_scheduler_reply(): got ack for task 05jl12aa.30680.16836.11.10.193_0
23/09/2012 09:43:28 | SETI@home | [sched_op] handle_scheduler_reply(): got ack for task 05jl12aa.30962.16836.12.10.66_1
23/09/2012 09:43:28 | SETI@home | [sched_op] handle_scheduler_reply(): got ack for task 05jl12aa.30680.16836.11.10.181_1
23/09/2012 09:43:28 | SETI@home | [sched_op] handle_scheduler_reply(): got ack for task 05jl12aa.30680.16836.11.10.174_1

I got ack for a whole lot of tasks but not for all....

[Jason G]

Indeed, Server issues are going to see lower than usual RACs at the moment.  Still, both those machines linked are underperforming... In the case of the 560ti it is capable of 30-50% more, if a newer application (x41g)  forced to above normal priority were used.  Also we don;t know if it;s a dedicated cruncher or not.

The 680 won't see full potential until down the road, but is capable of ~37k RAC or a bit more with x41z, excluding CPU... so that linked machine is running at about half capacity (older release app & maybe config issues & used for other things etc)

My own machines:
http://setiathome.berkeley.edu/show_host_detail.php?hostid=5631447
and
http://setiathome.berkeley.edu/show_host_detail.php?hostid=6135654

The evga SC 560ti is running pretty much flat chat (with i5-2400 cpu, Usually ~35k RAC), while on the 680 (with core2duo, usually ~40-42k RAC) I snooze Boinc to play Guild Wars 2 each evening

Jason

[Richard Haselgrove]

Now there is some problem with Seti also:

23/09/2012 09:42:34 | SETI@home | Reporting 532 completed tasks, not requesting new tasks

I got ack for a whole lot of tasks but not for all....

The server will currently ack a maximum of 256 tasks in one report. Users were finding that report counts of 1000 or more were generating 'internal server errors' at the project - the server program that handles the reporting was using too much memory. There was much discussion of it on the main project message board at the time.

If you are using BOINC v6.12 and above, it would be kinder to set

<max_tasks_reported>256</max_tasks_reported>

- no point in reporting more tasks than can be processed, it just adds to the network traffic. Any tasks left over will be kept on your machine and reported at the next server contact.

[Caprio]

Indeed, Server issues are going to see lower than usual RACs at the moment.  Still, both those machines linked are underperforming... In the case of the 560ti it is capable of 30-50% more, if a newer application (x41g)  forced to above normal priority were used.  Also we don;t know if it;s a dedicated cruncher or not.

The 680 won't see full potential until down the road, but is capable of ~37k RAC or a bit more with x41z, excluding CPU... so that linked machine is running at about half capacity (older release app & maybe config issues & used for other things etc)

My own machines:
http://setiathome.berkeley.edu/show_host_detail.php?hostid=5631447
and
http://setiathome.berkeley.edu/show_host_detail.php?hostid=6135654

The evga SC 560ti is running pretty much flat chat (with i5-2400 cpu, Usually ~35k RAC), while on the 680 (with core2duo, usually ~40-42k RAC) I snooze Boinc to play Guild Wars 2 each evening

Jason

These are both dedicated crunchers. Nowadays I wont snooze Seti for gaming.... They just run withouth snoozing.... The application is whatever comes with the Lunatics_Win64_v0.40_setup.exe package (for the Vista64 host). Looks like i have Lunatics_x41g_win32_cuda32.exe in my project folder. Does priority make any difference? How to modify the priority of the process?

[Claggy]

Does priority make any difference? How to modify the priority of the process?

The easiest method of changing priority with x41g is to run eFMer Priority:

http://www.efmer.eu/forum_tt/index.php?topic=198.0

Once you're got hold of x41z that'll no longer be needed, you'll be able to change the priority by editing a config file,

Claggy

[Caprio]

One difference between Jasons clients are the CPUs.

My Core2Duo E6600 contributes merely 2,5k/day and the Core2Quad Q6600 5k/day. The newer intel CPU that Jason is running might contribute more. Any recollection of what Jasons pure CPU contributions are?

[arkayn]

I am also doing more work than you and I only run on the GPU.

[Caprio]

With 2 WU in parallel the 560Ti at 900Mhz is contributing roughly 22,5k/day. -> the Dual Core CPU at 2880 Mhz 2,5k/day = 25

The client with Quad core -> 22,5k/day + 5k/day = 27,5k/day.

Lately performance has been degrading abit for some unknown reason...