KWSN Orbiting Fortress Forum Index KWSN Orbiting Fortress
KWSN Distributed Computing Teams forum
 
 FAQFAQ   SearchSearch   MemberlistMemberlist   UsergroupsUsergroups   RegisterRegister 
 ProfileProfile   Log in to check your private messagesLog in to check your private messages   Log inLog in 

Milkyway@Home - GPU & CPU performance stats wanted
Goto page 1, 2, 3, 4  Next
 
Post new topic   Reply to topic    KWSN Orbiting Fortress Forum Index -> KWSN BOINC'ers
View previous topic :: View next topic  
Author Message
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Mon Feb 24, 2014 2:46 pm    Post subject: Milkyway@Home - GPU & CPU performance stats wanted Reply with quote

[update 3] This benchmarking run is now over .
What started off as a singular speed 213.76 credit WU became 2 different speeds by late February, our 'standard' WU was joined by a much quicker one (approx. 1/3 faster). Our standard 'long' 213.76 credit WU became rare 2 months latter & is now extinct, hence the end.

[update 4]Well those WUs have made a sporadic come back! But it's too much of a PITA to be messing with so I've started a new benchmarking run & threads.

***************************************************************************************************

Hello old rivals Smile , hmm actually it might be so old that no one remembers! Lol, well referring to early SETI days, been a while since I posted here I think - yep I see I registered in 2002 & hadn't posted since!

Anyway, seeing as you guys are 1 of the higher ranking teams in MW on RAC & seem to still have an active forum I was hoping your guys could help.

I've been gathering benchmarks for MW from my 'home' team's forum (AnandTech) & the MW forums & got quite a few benchmarks so far (table below & links at the bottom for MW & AT thread), but some h/w is barely represented or not at all! I'm hoping you folk can help out with old or new hardware Smile.

Requirements for benchmark, validated 213.76 credit WUs only, average of 5 WU times, dedicated CPU core for the GPU, please state clock speeds if overclocked (including factory o/cs) or state 'stock'. It would also be handy if you could state your BOINC & driver version & OS, incase it does make any odds.

[update] It seems MW@H have now (20th Feb. ish) released new faster 213.76 credit WUs along with the previously 'standard' speed WUs, so going by the benchmarks gathered upto the 17/2/14, you need to average 5 of the longer 213.76 credit WUs only. Each 'speed' WU should only vary by a few seconds or so where their is nothing else working the GPU.

[update2] As of 21/4/14 the long 213.76 credit WUs are extremely rare now so dig deep to find them! The 2 times I got today I found 5 in 307 & 5 out 1020 results!! Once these WUs are gone it'll be the end of this benchmarking run.

Oh & for the benchmark please only crunch 1 WU at a time per GPU otherwise it will massively increase WU time, naturally! Wink (even if it does increase output), and the WU times seem to fluctuate much more than singly crunched WU so you can't 1/2 the time either. For CPUs you'll want to crunch 1WU per real core.

Current GPU statistics ~ Average Run Time to Complete 1 'long' 213.76 credit WU :-

HD 7970 (GPU 1150 MHz) ....................................... 75s Sunny129
R 280X (GPU 1100 MHz) ......................................... 75s Mumak
R 280X (GPU 1030 MHz) ......................................... 77s Dunx
HD 7970 (GPU 1100 MHz) ....................................... 77s Sunny129
HD 7970 (GPU 1000 MHz, RAM 1400 MHz, Cat 14.1) .. 78s Mesyn191
HD 7950 (GPU 1100 MHz) ....................................... 80s Leonheart
HD 7950 (GPU 1100 MHz) ....................................... 81s Mumak
HD 7970 (GPU 1050 MHz) ....................................... 81s Sunny129
HD 7950 (GPU 900 MHz) ......................................... 82s WES
HD 7970 (stock) .................................................. .. 83s mikey
HD 7970 (GPU 1000 MHz) ....................................... 85s Sunny129
HD 7950 (GPU 1000 MHz, RAM 1500 MHz,Cat 14.1 b) . 87s Leonheart
HD 7970 (GPU 950 MHz) ......................................... 90s Sunny129 (same time when RAM u/c to 800 MHz!)
HD 7950 (GPU 1000 MHz, RAM 1500 MHz) ................ 95s Leonheart
HD 7870 XT (stock) .............................................. 120s Matt
HD 7950 (GPU 850 MHz) ....................................... 121s salvordorhardin
HD 6970 GPU 950 MHz, RAM 900 MHz) ................... 148s Brickhead
HD 6990 (stock) .................................................. 165s Alez [TSBT's Pirate] (dual GPU card)
HD 6970 (stock) .................................................. 165s Mikey
HD 6950 (stock) .................................................. 176s Icecold
HD 5870 (GPU 868 MHz, RAM u/c 900 MHz) ............ 187s Brickhead
HD 5870 (stock) .................................................. 192s mikey
HD 5850 (GPU 850 MHz) ....................................... 210s Assim1
HD 5850 (GPU 850 MHz, RAM u/c 500 MHz) ............ 211s Assim1
HD 5850 (GPU 775 MHz) ....................................... 231s Assim1
HD 5850 (GPU 750 MHz) ....................................... 245s salvordorhardin
HD 5850 (stock) .................................................. 246s Assim1
HD 5850 (GPU 800 MHz) ....................................... 246s petrusbroder
HD 5830 (stock) .................................................. 282s Pheonix
GTX 780 Ti (stock) ............................................... 403s ... Linux - 395s biodoc
R9 270X .................................................. .......... 407s Alez [TSBT's Pirate]
GTX Titan ............................................................ 414s Yankton (GPU load only 18% despite a free CPU core)
HD 4870 512 MB (stock) ....................................... 444s JumpinJohnny
GTX 560 Ti 448c (GPU 880 MHz) ............................ 471s GleeM
HD 4870 1GB (stock) ............................................ 503s Assim1
GTX 570 (stock) ................................................... 520s - Linux biodoc
GTX 480 (GPU 750 MHz) ....................................... 520s Dunx
HD 7970M .......................................................... 525s Spencer
HD 4850 1 GB (stock) .......................................... 553s wayliff
GTX 770 (GPU 1333 MHz) ..................................... 609s Stojag
HD 4830 512 MB (GPU 670 MHz) ........................... 615s Assim1
HD 7770 GHz ed. (GPU 1100 MHz) ........................ 724s Mumak
HD 7770 GHz ed. (stock) ...................................... 807s Deerslayer
GTX 660 Ti (GPU 1046 MHz) ................................. 816s - Linux biodoc
GTX 560 Ti (stock) .............................................. 836s Deerslayer
HD 7750 (GPU 900 MHz, RAM 1300 MHz) .............. 1096s branjo
GTX 460 (GPU 750 MHz) 768 MB .......................... 1127s - Linux Ken g6

Current CPU statistics ~ Average Run Time* to Complete 1 'long' 213.76 credit WU :-

Intel Xeon E3-1230 v3 (Turboing to 3.7 GHz) ....... 9881s Stojag
Intel Core 2 Q9550 (3.6 GHz, 424 MHz FSB) ..... 12,532s Assim1
Intel Core i7-3612QM ..................................... 13,068s mikey
AMD Phenom II X4 965 (3.41 GHz) .................. 13,774s JumpinJohnny
Intel Core 2 Q8300 ........................................ 14,949s mikey
AMD Phenom II X6 1090T ............................... 14,950s mikey
AMD FX 6300 ................................................ 15,521s mikey
AMD FX 6100 ................................................ 17,449s mikey
Intel Core 2 Q8200 ........................................ 19,542s mikey
AMD Phenom 9850 ........................................ 25,844s mikey

OS Windows, unless otherwise stated. CPU at stock clock where no speed shown.

* I latter realised for the CPU tests only that I should of used & asked for the 'CPU time' rather than 'run time' which would include any slow downs due to running other processes. I think that largely due to the use of averaging at least 5 WU times that it will keep the inaccuracy low & that the times are still useful, just bear in mind their will be a little inaccuracy there. Just re-iterate that this refers to the CPU benchmarks only.

Seeing as I can't copy/paste linked names here I'll just do it as plain text, if you want to checkout the benchmarkers original post I suggest you go to the AnandTech thread linked below & click on the linked name there in the table. Unless someone can tell me a quick & easy way to make them work here?

MW forum thread http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=3465&sort_style=6&start=0
AnandTech thread http://forums.anandtech.com/showthread.php?t=2366988
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H


Last edited by Assimilator1 on Thu May 29, 2014 2:06 pm; edited 31 times in total
Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Mon Feb 24, 2014 3:44 pm    Post subject: Reply with quote

[OT]My humble 7750 is shrubbing 159.86-credits WU's only (MilkyWay@Home v1.02 (opencl_amd_ati)). What are 213.76's? Milkyway@Home Separation (Modified Fit)?[/OT]

#ni-2
_________________



Back to top
View user's profile Send private message Visit poster's website
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Mon Feb 24, 2014 3:48 pm    Post subject: Reply with quote

Shrubbing, lol I'd forgotten the KWSN terminology Wink.

Yes your right, the 213.76s are sep mod fit WUs, have you not got that app enabled in the online MW preferences?

No probs on benchmark related problems btw.
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H
Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Mon Feb 24, 2014 3:53 pm    Post subject: Reply with quote

Nope, I unchecked it because it pays less credits per minute than plain MW. Anyway, in the name of the science, I am going to try it #ni-1
_________________



Back to top
View user's profile Send private message Visit poster's website
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Mon Feb 24, 2014 4:00 pm    Post subject: Reply with quote

Interesting, btw they may pay less credit but they also draw much less power.
My rigs power draw jumps from ~267w to ~291w on MW v1.02 WUs! (C2Q @ 3.6 GHz, HD 5850 @ 850 MHz).

And thx for trying it for the benchmark Smile.
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H
Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Mon Feb 24, 2014 4:13 pm    Post subject: Reply with quote

Well, my 7750 has ca. 55-75W TDP, so I am not very worried about its consumption Smile
_________________



Back to top
View user's profile Send private message Visit poster's website
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Mon Feb 24, 2014 4:34 pm    Post subject: Reply with quote

Very Happy, can't remember what mine is for sure, I think it's 160w.
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H
Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Mon Feb 24, 2014 4:45 pm    Post subject: Reply with quote

Meanwhile on one of my wingmen (user bcavnaugh):

- NVIDIA GeForce GTX TITAN (4095MB) driver: 334.89 OpenCL: 1.01
- Microsoft Windows 7, Ultimate x64 Edition, Service Pack 1, (06.01.7601.00)
- 7.2.39
- i7-4960X
- calculation for upper 7 tasks on "Valid Tasks" page

Mean avg: 365.66
Median avg: 375.57

#ni-1
_________________





Last edited by branjo on Tue Feb 25, 2014 4:10 am; edited 1 time in total
Back to top
View user's profile Send private message Visit poster's website
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Mon Feb 24, 2014 6:20 pm    Post subject: Reply with quote

Err so is that for the Titan?
I would think it'd be faster.....

The 7 'tasks' were 213.76 cred WUs right?
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H
Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Tue Feb 25, 2014 4:18 am    Post subject: Reply with quote

Yes. NVIDIA's are bit overrated. And quite bad for DP...

So, my ASUS HD7750 1GB:

- max OC'ed (GPU clock 900MHz, Memory Clock 1 300 MHz)
- Power limit 20%
- Driver (Catalyst) 14.1 Beta
- Microsoft Windows 7 Home Premium x64 Edition, Service Pack 1,
- 8GB RAM
- i7-3770
- BOINC 7.2.39
- GPU utilization: 99%

Analyzed 21 WU's (213,76 credits per):
- mean average: 951,85
- median average: 1 096,11

[OT]For comparison purposes, 60 MilkyWay WU's running 2 concurrent tasks (159,86 credits per):
- mean average: 1 075,98
- median average: 1 080,53
[/OT]

Cheers #ni-1

ETA: other 7 CPU threads running WCG/MCM1

ETA2: the speed varies from 713,08 to 1 099,44 (as indicated in huge difference between mean and median)
_________________





Last edited by branjo on Tue Feb 25, 2014 11:31 am; edited 1 time in total
Back to top
View user's profile Send private message Visit poster's website
Stojag
Grunt
Grunt


Joined: 25 Feb 2014
Posts: 2

PostPosted: Tue Feb 25, 2014 8:54 am    Post subject: Reply with quote

Quote:
Yes. NVIDIA's are bit overrated. And quite bad for DP...


Did you activate native DP support of your TITAN in the NVIDIA Control Panel?

TITAN an 780Ti DP computation should be 1/3rd of SP, instead of 1/24th of normal speed from other NVIDIA cards.

Those times seem to be a bit high for a native DP-support
Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Tue Feb 25, 2014 10:08 am    Post subject: Reply with quote

IDK #ni-2

Quote:
Meanwhile on one of my wingmen (user bcavnaugh):

- NVIDIA GeForce GTX TITAN (4095MB) driver: 334.89 OpenCL: 1.01 ...


Anyway, useful information for people shrubbing on Titan, thanks #ni-1
_________________



Back to top
View user's profile Send private message Visit poster's website
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Tue Feb 25, 2014 12:35 pm    Post subject: Reply with quote

I know most NVidia cards are poor at DP but the Titan is an exception, it is faster than the fastest AMD retail single GPU card (HD 7970 GE) in DP.

So I think Stojag is onto something, you'll want to have a dig around the ctrl panel & see if you can find that option.
Just FYI, looking at the table, a time of 370s for your Titan would make it just ~8% faster than the 780 Ti. The Titans DP is over 7x more powerful than that card! (according to the wiki anyway).

Oh btw, I'm not sure what you mean by 'mean' & 'median' average Embarassed
Anyway the 7750s 1096 average looks about right for your 213.76s Smile, don't know where you get 952s from?

Stojag - exactly where in the ctrl panel is it?
Btw according to the wiki http://en.wikipedia.org/wiki/GeForce_700_Series the 780Ti does not have 1/3 DP.
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H
Back to top
View user's profile Send private message
Plomos
Prince
Prince


Joined: 26 Apr 2011
Posts: 859

PostPosted: Tue Feb 25, 2014 2:12 pm    Post subject: Reply with quote

So i have grabbed a couple of these tasks for my cpu. making sure that it only runs 1 MW WU at a time so as to get a real time estimate, but that means it will take longer to do 5
_________________

Back to top
View user's profile Send private message
branjo
Prince
Prince


Joined: 05 Jan 2006
Posts: 746
Location: Slovakia

PostPosted: Tue Feb 25, 2014 2:57 pm    Post subject: Reply with quote

Assimilator1 wrote:
I know most NVidia cards are poor at DP but the Titan is an exception, it is faster than the fastest AMD retail single GPU card (HD 7970 GE) in DP.

So I think Stojag is onto something, you'll want to have a dig around the ctrl panel & see if you can find that option.
Just FYI, looking at the table, a time of 370s for your Titan would make it just ~8% faster than the 780 Ti. The Titans DP is over 7x more powerful than that card! (according to the wiki anyway).

Oh btw, I'm not sure what you mean by 'mean' & 'median' average Embarassed
Anyway the 7750s 1096 average looks about right for your 213.76s Smile, don't know where you get 952s from?

Stojag - exactly where in the ctrl panel is it?
Btw according to the wiki http://en.wikipedia.org/wiki/GeForce_700_Series the 780Ti does not have 1/3 DP.


OK, I do not have TITAN, the results were from some (for me) unknown person Cool

"Mean" average is the simple (usual) average used by majority of people when calculating averages, e.g. (1+2+13)/3=5.3
"Median" average is the middle value of the list written in order, e.g. 1, 2, 13 = 2.

Median is very useful to exclude extreme values from the list and to get "better" (more precise) average number (like in my case of MilkyWay - median excluded all low times which occurred who-knows-why).

Example: What is the "average" salary of 5 people earning $50 000, $60 000, $65 000, $80 000 and $5 000 000?

If calculating it with usual, "Mean", the average salary of those 5 people is $1 051 000 even though 4 of them are not even close to this "average" Shocked . If we use "Median", the average salary is $65 000 (because it excluded extreme salary of 5 million USD which screwed the calculation). What "average" is more realistic and indicative?

So I agree, "median" average is in my case better indicator than "mean" because there were huge differences between lowest (700 and something) and highest (almost 1 100) times.

In case of MilkyWay application, they are almost the same because all times were more-or-less the same Wink

Cheers #ni-2
_________________



Back to top
View user's profile Send private message Visit poster's website
Plomos
Prince
Prince


Joined: 26 Apr 2011
Posts: 859

PostPosted: Tue Feb 25, 2014 10:11 pm    Post subject: Reply with quote

So far i have finished 2 WUs running them one at a time on the cpu and the average came out to 14,569.78 sec cpu time. I am estimating that the other units will be more or less in the same time area so that number would vary very little.

these are run on an Intel i3 3220 at 3.30 GHz stock
_________________

Back to top
View user's profile Send private message
Stojag
Grunt
Grunt


Joined: 25 Feb 2014
Posts: 2

PostPosted: Wed Feb 26, 2014 3:53 am    Post subject: Reply with quote

Assimilator1 wrote:

Stojag - exactly where in the ctrl panel is it?


As it is described in some reviews, you must look in the 3D-Settings -> Manage 3D-Settings -> Global Setting

There should be an option "CUDA double precission" where you can enable the native DP support.
That may lower the TITANs base clock a bit.
Back to top
View user's profile Send private message
JumpinJohnny
Prince
Prince


Joined: 28 Mar 2013
Posts: 1247
Location: Western New Hamster

PostPosted: Wed Feb 26, 2014 7:55 am    Post subject: AMD / ati 4870 Reply with quote

I have completed your testing using:

GPU Hardware:
AMD / Ati Radeon HD 4870 (double precision)
Memory 512 MB @ GDDR5
Core Clock 750 MHz
Memory Clock 900 MHz
Total Memory Bandwidth 115.2 GByte/s

GPU Software:
Catalyst Version 13.1
2D Driver Version 8.01.01.1248
Direct3D Version 7.14.10.0911
OpenGL Version 6.14.10.11672

Computer using:
AMD Phenom II X4 965 @ 3411 Mhz
4 GB RAM
Windoz 7 Ultimate sp1 (x64)
Boinc 7.2.39 (x64)


Milkyway@Home Separation (Modified Fit) v1.28 (opencl_amd_ati)
1. There are two varieties of this WU that both produce 213.76 credits.
a. fast variety WU mean average = 292.61 seconds
b. slower variety WU median average = 443.77 seconds
2. I am running N-body Simulation 1.40 WU on 3 cores with one CPU core open for GPU use.
_________________

Back to top
View user's profile Send private message
Assimilator1
Knight
Knight


Joined: 23 Aug 2002
Posts: 40
Location: UK, Surrey, Guildford

PostPosted: Wed Feb 26, 2014 3:34 pm    Post subject: Reply with quote

Thx for your results JJ Smile
Your 4870 times make me think that my 4870's GPU load wasn't as near 100% as I thought.....
Btw why did you use mean & median average for the 2 different WUs?
Anyway, 444s looks right to me too.

Are you going to give some CPU times too?

Plomos
Opps, you've mis-read my benchmark requirements slightly, the 1WU rule was meant for GPUs only, for CPUs 1 WU/(real)core is fine.
I'll edit the requirements to make it clearer.

You'll get your results more quickly now Wink.
Btw am I right in thinking that i3s don't have turbo?
TIA

branjo
Re mean & medium, agreed median is more useful generally. But for MW the WUs times don't fluctuate much (bar the long & short WU differences) if GPU load is nr. 100% & the GPU has its own cpu core free, so mean & median average should give nearly identical times.

Btw the big difference in WU times of 213.76 credit WUs I had already pre-warned you about in my benchmark requirements Wink. Since the ~20th Feb MW have released 2 different speed 213.76 credit WUs, as I said I only wanted the times for the 'long' ones Smile.

Btw2, where did you get the Titan times from? I don't see that rig listed in your computers. Which it would be if someone were crunching for you.
I looked at computers beyond 30 days too........

Stojag
Thx for the info, I was hoping it would help branjo, but seeing as it's not his PC it won't Wink nm.
How much does native DP lower the Titans clock btw?
_________________
Team AnandTech - SETI, F@H, Muon1, MW@H, Asteroids@H, LHC@H, Skynet, Rosetta@H
Back to top
View user's profile Send private message
Yankton
Prince
Prince


Joined: 27 Sep 2008
Posts: 1702
Location: California

PostPosted: Wed Feb 26, 2014 4:28 pm    Post subject: Reply with quote

I'll run some titan tests for you in the next day or two. As soon as it's current job ends.
_________________
Some days are worse than others.
Back to top
View user's profile Send private message
Display posts from previous:   
Post new topic   Reply to topic    KWSN Orbiting Fortress Forum Index -> KWSN BOINC'ers All times are GMT - 5 Hours
Goto page 1, 2, 3, 4  Next
Page 1 of 4

 
Jump to:  
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum


Powered by phpBB © 2001, 2005 phpBB Group
Optimized Seti@Home App | BOINC Stats