Attempt to Rank Nvidia Cards based on crunching ability

Message boards : Number crunching : Attempt to Rank Nvidia Cards based on crunching ability
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · Next

AuthorMessage
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028534 - Posted: 26 Aug 2010, 3:49:02 UTC - in response to Message 1028530.  

@Sutaru, that 8245 for the 260-216 was a result from Tim Norton...It seems fairly low. Perhaps he hit a chunk of VHARs or something.


@Sutaru and @razamatraz

have redone some results on the same card and average RAC has increased from the 8245 to 10222 using the calc Razamatraz suggested - increase is probably due to increasing speed to 650 from 630 but this only affects part of the results

the new result is over 194 wu rather than 57 for previous result

there are some reasons why @Sutaru has a much high estimated rac for his GTX260-216
its more over-clocked than mine at 680 vs 650
i have turned down the memory to 700 and Sutaru has his at 1500 - done to save power
the inconsistency of the credit system - looking at the two examples Sutaru lists there is a 25% difference
he is using winxp and i am using win7 with different drivers

looking at Sutaru's run times i.e. GPU secs per wu for a normal wu i get approx 1000 to 1100 sec Sutaru gets 500 to 550 secs - logically if the credit system was consistent my rac should be half Sutaru and so his estimate of rac of 15k vs my original of 8k is about correct?

but whats also interesting is that VW Bobier is getting run times of 14-15 minutes (GTX295) compared to mine of 17-18 which is more what you would expect - 500 to 550 sec is more GTX460+ territory?

There are a lot of "types" of GTX 260-216 out on the market - some are highly over clocked with from the stock 576mhz up to 700+ again this will add variations to the scores

after the outage i will change the overclock and memory speed on this card to see if it changes the results significantly - originally thought that the memory speed for SETI on the card did not matter that much but that appears not to be the case - i used to run Milkyway and Collatz on the card and i think thats where the idea of turning down the memory speed came from also i may then try changing my drivers to the old ones Sutaru is using to see how much difference that may make if they work on Win7

if the results are consistent then i should be able to get to within say 10-15% of the high rac Sutaru is getting as my card will not go beyond 650mhz it will take a couple of weeks to get enough results with

a) higher memory speeds
b) older drivers 190.38 that Sutaru is using

the machine which the results are from is a dedicated cruncher so not affected by other factors

Yer lucky, I'm holding back as My card is the 1st of 3, I plan to buy 2 more and water cool all 3, plus the mosfets and the south bridge chip and the cpu, So I need to get 2-420 Radiators, 3-gpu water blocks(Bitspower or Danger Den), an EK cpu water block and a kit from EK for the motherboard stuff, then I'd need a pump, a reservoir, a koolance external mount for a 420 radiator, 12-140mm fans w/120mm screw holes(Don't worry I know why I'm saying this as It has to do with the particular radiator type I want to buy), tubing, clamps, distilled water, a silver coil(or something to that effect that kills algae, I know what It, I just don't remember the name currently) and a few other parts. Once water cooled I may see if all 6 gpus will do 700MHz or so.

Oh and remember the ram speed of 1008MHz on My 295? It relates to memory bandwidth, as in 112.9 GB/s.
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028534 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028575 - Posted: 26 Aug 2010, 9:49:30 UTC - in response to Message 1028530.  
Last modified: 26 Aug 2010, 10:38:22 UTC

AFAIK, MilkyWay@home don't use (much) the graphic card RAM.
Because of no memory controller load (0 %) in GPU-Z.
If my GPU crunch SETI@home, it's loaded up to ~ 70 %.
So I guess SETI@home profit from high graphic card RAM speed.

AFAIK, if you lower the RAM, you could reach higher core/shader speed.

So you could test this and maybe this.

It's well that my GPUs are manufacturer OCed, so I don't need to think about this.. ;-)
And I have 3 (GIGABYTE) and 4 (EVGA, 2 years + extension. AFAIK, lifelong not in Europe/Germany) years warranty on the GPUs. X-D

The GIGABYTE GTX260(-216) SOC have not the 'stock' RAM. It's a 'specially' RAM. So they could adjust the graphic cards RAM speed from 1,000 to 1,250 MHz. (@ 680/1500/1250)
EVGA could OC the 'stock' RAM to 1,151 MHz at the SSC OC edition. (@ 675/1453/1151)

I tested manually OC, but the nVIDIA software was buggy. But I think because of the old 190.38 driver. But it's the fastest driver on my systems.
So why I should take a newer nVIDIA driver (which will work maybe better with the OC tool) and then I could maybe OC higher, but the result with old driver and not additonal OC is faster.. ;-)

If you have WinXP and a GTX2xx GPU (and it's a pure crunching machine), I would take the 190.x and stock MB 6.09 cuda23 app. And BOINC Rescheduler, because of no VLAR kill.. ;-)
I have a test in my profile about which (currently available) app is the fastest on my systems. And I don't think a new CUDA app (of the opt. crew) will change this picture on my systems. On GTX4xx (Fermi) GPUs - maybe yes.

You have Win7, maybe the lowest would be 197.x . I don't know if 190.x will work well on your OS. But you could test it.
ID: 1028575 · Report as offensive
Profile Tim Norton
Volunteer tester
Avatar

Send message
Joined: 2 Jun 99
Posts: 835
Credit: 33,540,164
RAC: 0
United Kingdom
Message 1028581 - Posted: 26 Aug 2010, 10:42:52 UTC - in response to Message 1028575.  
Last modified: 26 Aug 2010, 10:44:08 UTC

Hi Sutaru

yep my gpu utilisation is approx 70% as well

just checked my cruncher and Nvidia drivers had dropped the speed back to 576 after yesterdays reboot after MS updates applied - but the good news is i have just spent 5 minutes tweeking the settings and have now got the gpu at 670 and memory back up to 1000/1300 :)

its taken a minute to a minute and a half off the runtimes for cuda so down to just under 13 minutes now - small sample set of 5 wu

if you have special ram in your cards that may account for some of the speed difference and i will try out older drivers at the weekend when i will have more time. if necessary i can change the OS back to xp but will do the drivers first

all the testing i have done shows the opp app to be faster in win7, at least, interesting that the stock cuda app is faster for you - your task list shows anonymous GPU though - have you referenced the stock app in your app info then?
Tim

ID: 1028581 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028600 - Posted: 26 Aug 2010, 12:38:59 UTC - in response to Message 1028581.  

Which tool you use for OC your GPU?
Normally you can make a 'OC profile' which will be loaded automatically always after reboot of your system.

AFAIK, WinXP is still the fastest Windows crunching OS for SETI@home.
But it's maybe only ~ 5 % or something. But for hardcore cruncher is this a lot! ;-)
If you are happy with Win7 and it's a daily use machine, maybe it would be well to sticky with Win7.

Yes I have 'anonymous' in the PC overviews.
But I have app_info.xml with stock MB 6.09 cuda23 for GPU and AK v8b for CPU. On both machines.
And I need to use the BOINC Rescheduler for to protect my GPUs for VLAR WUs.

My GIGABYTE have avg. ~ 45 % and peak ~ 70 % memory controller load if SETI@home CUDA.
It's the secondary GPU in this PC and only for CUDA. Display only to onboard GPU. But both with WinDesktop - without, no CUDA possible.

The 940 BE machine, only to the first GPU display connected. The other 3 GPUs no WinDesktop (which will maybe increase the performance).

ID: 1028600 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028629 - Posted: 26 Aug 2010, 14:13:32 UTC - in response to Message 1028600.  

Which tool you use for OC your GPU?
Normally you can make a 'OC profile' which will be loaded automatically always after reboot of your system.

AFAIK, WinXP is still the fastest Windows crunching OS for SETI@home.
But it's maybe only ~ 5 % or something. But for hardcore cruncher is this a lot! ;-)
If you are happy with Win7 and it's a daily use machine, maybe it would be well to sticky with Win7.

Yes I have 'anonymous' in the PC overviews.
But I have app_info.xml with stock MB 6.09 cuda23 for GPU and AK v8b for CPU. On both machines.
And I need to use the BOINC Rescheduler for to protect my GPUs for VLAR WUs.

My GIGABYTE have avg. ~ 45 % and peak ~ 70 % memory controller load if SETI@home CUDA.
It's the secondary GPU in this PC and only for CUDA. Display only to onboard GPU. But both with WinDesktop - without, no CUDA possible.

The 940 BE machine, only to the first GPU display connected. The other 3 GPUs no WinDesktop (which will maybe increase the performance).

I've found I don't need to use any type of Rescheduler anymore Sutaru, But to each their own. But then I'm only seeing this now, As I haven't seen any new vlars sent to the gpu in over a week or so. Just vlars to the cpu.

The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028629 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028644 - Posted: 26 Aug 2010, 14:46:15 UTC - in response to Message 1028629.  
Last modified: 26 Aug 2010, 14:52:33 UTC

I still recommend to use the BOINC Rescheduler of Fred.
1st, VLAR WUs. 2nd, to eliminate/protect the -177 errors.

Last time (~ one week) after I switched on again the PC after a small break, I got ~ 600 VLAR WUs for the GPUs on my 940 BE with 4x GTX260 machine (~ 30 %). All re-sent without .vlar_x ending.
[EDIT: The 940 BE need ~ 19 days for to crunch them.]

'Worst case' up to two years we could see old re-sent VLAR WUs on GPUs, IIRC.

I let run the BOINC Rescheduler one time/day. And will continue at least the next two years with this. ;-)
ID: 1028644 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028646 - Posted: 26 Aug 2010, 14:52:15 UTC - in response to Message 1028644.  
Last modified: 26 Aug 2010, 14:52:31 UTC

I still recommend to use the BOINC Rescheduler of Fred.
1st, VLAR WUs. 2nd, to eliminate/protect the -177 errors.

Last time (~ one week) after I switched on again the PC after a small break, I got ~ 600 VLAR WUs for the GPUs on my 940 BE with 4x GTX260 machine (~ 30 %). All re-sent without .vlar_x ending.

'Worst case' up to two years we could see old re-sent VLAR WUs on GPUs, IIRC.

I let run the BOINC Rescheduler one time/day. And will continue at least the next two years with this. ;-)

I got the -177 errors on My cpu and on My gpus when I used the flop entry in My app info file, Without the flop entry I don't get -177 errors, anywhere.
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028646 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028650 - Posted: 26 Aug 2010, 14:59:15 UTC - in response to Message 1028646.  

But I thought it's still recommended to use flops entries in app_info.xml, no?

How are your estimate times of the WUs in BOINC. They are correct?

ID: 1028650 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028653 - Posted: 26 Aug 2010, 15:07:07 UTC - in response to Message 1028650.  
Last modified: 26 Aug 2010, 15:07:37 UTC

But I thought it's still recommended to use flops entries in app_info.xml, no?

How are your estimate times of the WUs in BOINC. They are correct?

If I use the flop entries, I get -177 errors almost all the time as I said, So why should I? I'm using the anonymous platform, So something at Seti is still broken and so I won't use something that does not work right...
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028653 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028656 - Posted: 26 Aug 2010, 15:13:27 UTC - in response to Message 1028653.  

IIRC, the flops entries would help also if you use the Rescheduler tool.
Estimate times are better of renamed WUs.

And Fred's Rescheduler have an option for to check that the -177 are not happen.

Also I have set 0.5 min and 1.5 max DCF.

And the Rescheduler tool adjust all well.

ID: 1028656 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028658 - Posted: 26 Aug 2010, 15:29:23 UTC - in response to Message 1028656.  
Last modified: 26 Aug 2010, 15:32:28 UTC

IIRC, the flops entries would help also if you use the Rescheduler tool.
Estimate times are better of renamed WUs.

And Fred's Rescheduler have an option for to check that the -177 are not happen.

Also I have set 0.5 min and 1.5 max DCF.

And the Rescheduler tool adjust all well.

I'll pass, My system works as I've designed It and lets see I use the following two files:

AK_v8_win_x64_SSSE3x.exe(cpu)
Lunatics_x32f_win32_cuda30_preview.exe(gpus)

So like I said I don't need either the rescheduler or the flops as flops causes Me problems, So I will pass...


I also use BoincTasks 0.72 instead of Boinc Manager and My dcf stays near 1.25 so far and It been around that for about a week. So why do I need a rescheduler program made by Fred when I use BT 0.72 which is made by Fred???

Ok?
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028658 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028668 - Posted: 26 Aug 2010, 16:13:51 UTC - in response to Message 1028658.  
Last modified: 26 Aug 2010, 16:18:38 UTC

Why you start negative vibrations?

O.K., it's your choice.

Instead of to crunch the VLAR WUs of your GPU on your CPU you let kill them because of the 'optimized' CUDA app.

Sure, then you don't need to use all 'the stuff', for to protect the SETI@home internet pipe for unneeded traffic..


EDIT: BTW. Maybe it would be well if we would go back to topic..
ID: 1028668 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028669 - Posted: 26 Aug 2010, 16:22:50 UTC
Last modified: 26 Aug 2010, 16:28:25 UTC

Look Saturu, My gpu app does no killing ok? As the Lunatics did not add such a feature, So please go edit Your comment before time runs out...

You started It in this post Here. And You wanted to know how I could crunch so fast, Yet I don't kill any WU as there is no need to as I do not use apps that do that, Ok?????
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028669 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028674 - Posted: 26 Aug 2010, 16:41:22 UTC - in response to Message 1028669.  

You say that your 'Lunatics_x32f_win32_cuda30_preview.exe' will not delete/kill VLAR WUs if they are on your PC for your GPUs?

ID: 1028674 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028676 - Posted: 26 Aug 2010, 16:43:49 UTC - in response to Message 1028674.  

OTOH, what you mean with fast?

Your GPU make a 0.44x AR WU in 500 - 515 secs.?

ID: 1028676 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028679 - Posted: 26 Aug 2010, 16:53:05 UTC - in response to Message 1028674.  

You say that your 'Lunatics_x32f_win32_cuda30_preview.exe' will not delete/kill VLAR WUs if they are on your PC for your GPUs?

Not as far as I know and rescheduler 1.9 would have said something, Otherwise It would not have had any WU's to reschedule, Only the 12b seti app did the killing and I don't have that old app.
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028679 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028680 - Posted: 26 Aug 2010, 16:55:06 UTC - in response to Message 1028676.  

OTOH, what you mean with fast?

Your GPU make a 0.44x AR WU in 500 - 515 secs.?

Look a killed WU will generate an error of It's own and I have not seen any such error, Ok? So drop It... The subject of killed WU's is dead, I will not discuss It as I do not do that.
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028680 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1028685 - Posted: 26 Aug 2010, 17:19:55 UTC - in response to Message 1028680.  

Just a hint..

You should look to the error overview of your PC.
A lot of -177 errors.
You should use Fred's BOINC Rescheduler.


And if or not this new CUDA build kill/delete VLAR WUs, could answer someone of the opt. crew.
Because of this is a BETA app, I have no access to the BETA area of the Lunatics forum, so I can't read about.


I mean/t nothing as offensive.
I want only to help.

ID: 1028685 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 65709
Credit: 55,293,173
RAC: 49
United States
Message 1028702 - Posted: 26 Aug 2010, 18:34:56 UTC - in response to Message 1028685.  

Just a hint..

You should look to the error overview of your PC.
A lot of -177 errors.
You should use Fred's BOINC Rescheduler.


And if or not this new CUDA build kill/delete VLAR WUs, could answer someone of the opt. crew.
Because of this is a BETA app, I have no access to the BETA area of the Lunatics forum, so I can't read about.


I mean/t nothing as offensive.
I want only to help.

BoincTasks has not had a -177 error show up since the 23rd of this month, I need no help, My PC works perfectly.

This new build will not kill/delete, As that would be a -6 error in BoincTasks and I have not seen that in the History/Logs at all.

None taken.
The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's
ID: 1028702 · Report as offensive
JohnDK Crowdfunding Project Donor*Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 28 May 00
Posts: 1222
Credit: 451,243,443
RAC: 1,127
Denmark
Message 1028709 - Posted: 26 Aug 2010, 19:01:29 UTC - in response to Message 1028702.  

[BoincTasks has not had a -177 error show up since the 23rd of this month, I need no help, My PC works perfectly.

Not for nothing, but there's tons of -177 errors in you list, including a few from Aug 24, the last day before the outage. But if your PC runs perfect what do I know :))
ID: 1028709 · Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · Next

Message boards : Number crunching : Attempt to Rank Nvidia Cards based on crunching ability


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.