CUDA and openCL Benchmarks

Mumak
Joined: 26 Feb 13
Posts: 335
Credit: 3558684502
RAC: 908859

RE: Thank you for the

Quote:
Thank you for the information. So 7950 is not much slower that 7970 ?

Depends on what you understand with "much" ;)
I'd say it's not that much ;) For some users not worth the price difference and power requirements (input/output, temperatures).
You can still overclock the 7950, I'm running mine at a reduced clock now, because it's quite hot here...

[CSF] Vagelis Giannadakis
[CSF] Vagelis G...
Joined: 18 Mar 05
Posts: 8
Credit: 3760202
RAC: 0

I've set my stock-clocked GTX

I've set my stock-clocked GTX 650Ti crunching WUs for Einstein the past few days. It's been completing PA0007, PA0079 and PA0080 tasks with BRP4cuda32nv270 in ~10000 sec (some in ~15000). I tried running two tasks at a time, setting the relevant Einstein parameter to 0.5, but that pushed completion estimations to more than 8h (29000sec)!!

Reading this thread made me think that my 650Ti would do better with 2 or even three concurrent tasks. Was I wrong, something has changed with Einstein apps, do I need to change something else as well in my BOINC configuration, or something could be wrong with my setup?

I'm on Ubuntu 12.04 x86_64 with Nvidia driver 310.44

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5874
Credit: 118450195019
RAC: 25912923

RE: I've set my

Quote:
I've set my stock-clocked GTX 650Ti crunching WUs for Einstein the past few days. It's been completing PA0007, PA0079 and PA0080 tasks with BRP4cuda32nv270 in ~10000 sec (some in ~15000). I tried running two tasks at a time, setting the relevant Einstein parameter to 0.5, but that pushed completion estimations to more than 8h (29000sec)!!


The Perseus Arm Survey (PAS) run, when it started, had 'bundled tasks' which incorporated 3 actual tasks. Recently, Bernd announced that he was changing the bundle size to 2. That's why you see some taking 10Ksecs and some taking 15ksecs. You will continue to get some of the older, longer running tasks until they are all gone. That could take a week or three.

When you changed the parameter so you could run 2x, did you actually let any tasks complete? The estimate may well have doubled but if you don't run to completion, how will you know what they actually take? I have a 650Ti and it performs significantly better on 2x and even a bit better again on 3x. There are other factors which may effect your experience, but you wont know for sure unless you try (properly) :-).

Cheers,
Gary.

[CSF] Vagelis Giannadakis
[CSF] Vagelis G...
Joined: 18 Mar 05
Posts: 8
Credit: 3760202
RAC: 0

RE: When you changed the

Quote:
When you changed the parameter so you could run 2x, did you actually let any tasks complete? The estimate may well have doubled but if you don't run to completion, how will you know what they actually take? I have a 650Ti and it performs significantly better on 2x and even a bit better again on 3x. There are other factors which may effect your experience, but you wont know for sure unless you try (properly) :-).


Hi, thanks for your response!

I didn't let the tasks complete, but I did let them run for 2-3 hours, so I could make better estimations. In fact, one task (which was close to completion when I enabled two-at-a-time) did complete and another one came in, which I let run for more than an hour. All were estimated to complete in more than 8h, which is worse than running one at a time, taking 2h50m. One-at-a-time in my case means in 8h I will have completed 2.8 WUs, instead of less than two in two-at-a-time.

Bear in mind, I make estimations with completion progress, not taking into account BOINC manager's estimations.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5874
Credit: 118450195019
RAC: 25912923

RE: ... All were estimated

Quote:
... All were estimated to complete in more than 8h, which is worse than running one at a time, taking 2h50m.

If you look carefully at the two tasks that were run for a period as 2x, you will find they were the older (and longer running) 'bundle of 3' tasks whereas those that took 2hr 50min singly were the new, shorter 'bundle of 2' tasks. A bit of an unfair comparison don't you think?

All I can do is repeat myself. If you allow several tasks to run to completion (running x2) you will find that you get an improvement in performance.

In the future, there will be fewer 'older' tasks turning up, so less chance of fooling yourself with an invalid comparison. Any older tasks will be 'resends' of failures, since there are no new 'bundle of 3' tasks anymore. They will be easily spotted in your work cache as they will have a suffix to their name of _2, _3, _4 (or even higher). They will also have a 50% higher time estimate that will be easy to see in BOINC Manager.

Cheers,
Gary.

[CSF] Vagelis Giannadakis
[CSF] Vagelis G...
Joined: 18 Mar 05
Posts: 8
Credit: 3760202
RAC: 0

All three tasks that I tested

All three tasks that I tested in two-at-a-time mode needed just ~3h when run in one-at-a-time mode, they were all of the same type. Thus, at least in my case, two-at-a-time mode didn't work for me. Maybe there's some additional configuration that has to be made in app_config.xml or something.

Sunny129
Sunny129
Joined: 5 Dec 05
Posts: 162
Credit: 160342159
RAC: 0

it appears the last time

it appears the last time dskagcommunity updated the run time spreadsheet was back on April 20, which was well before Bernd's announcement about the reduction in BRP task bundle size. now that the change has been in effect for a few weeks now, a new run time spreadsheet will have to be started since run times are approx. 2/3 what they used to be...

Beyond
Beyond
Joined: 28 Feb 05
Posts: 123
Credit: 2515274311
RAC: 5017373

RE: it appears the last

Quote:
it appears the last time dskagcommunity updated the run time spreadsheet was back on April 20, which was well before Bernd's announcement about the reduction in BRP task bundle size. now that the change has been in effect for a few weeks now, a new run time spreadsheet will have to be started since run times are approx. 2/3 what they used to be...


I think his spreadsheet was for BRP4 tasks, no longer available for GPUs.

k6xt
k6xt
Joined: 11 Sep 05
Posts: 7
Credit: 28650981
RAC: 0

I'm thinking about adding a

I'm thinking about adding a 760GTX card. Are there any updates available beyond the 6xx series?

My performance on a 560ti is unimpressive relative to the comparisons posted here, for example one of mine:
http://einsteinathome.org/task/400146126

Were optimized apps in use? If there are some I'm not using them, but would if I know the link.

I just restarted Einstein. My BSRAC hasn't reflected it yet.

Thanks
Art

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 177
Credit: 12757753172
RAC: 4926766

Hi Art, The times reported in

Hi Art,
The times reported in this thread are for an ealier version of the BRP4 task which is way different than the current BRP4G & BRP5.

Your 560Ti will do just fine here - ~50k credits/day on BRP5 with the right cpu doing the driving. The GTX600/700 series really need PCIe3 to perform at their best and I know your Intel cpu's are only PCIe2, not certain about your AMD machines.

You can get a good idea of card & cpu performance by going thru the host statistics webpage.

Gord

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.