Attempt to Rank Nvidia Cards based on crunching ability


log in

Advanced search

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

1 · 2 · 3 · 4 . . . 5 · Next
Author Message
Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026031 - Posted: 18 Aug 2010, 7:01:27 UTC

So I was thinking it'd be nice to have a ranking system for performance on currently available Nvidia cards for SETI. I figure it should go something like this.

8400GS 512 meg only
9400 GT
210/310 (very small improvement over 9400 just faster clock speed)
9500 GT
GT 220/315
9600 GT

GT 320 (OEM only, I've never seen one.)
GT 240/GT 340


9800 GT (formerly known as 8800 GT)
9800 GTX

GTS 250 (formerly known as 9800 GTX+)
GTX 260 (there were 2 versions of this, the 192 and 216 versions)
GTX 280
GTX 275
GTX 285

GTX 460
GTX 470

GTX 295

GTX 480


The GTX 280 fits in around 275 level, but is unavailable. The 460 is potentially faster than the 285, but FERMI code is not at that point yet so it is about even. The GTX 480 should be faster than a 295, but until FERMI code matures the 295 is actually faster.....correct me if I am wrong.

This is in theory, with no overclocking. Keep in mind there are many versions of each card, some of the lower end cards have versions with DDR2 and DDR3. memory size can matter, but not much as long as there is enough. Also we recently found that Frizz' 240 is faster than my 9800 GT, but I have been running FERMI on it so that may account for the difference.


Yes I know some cards are missing. They are not available at stores currently so I see no point in including them.

IF you think I am mistaken feel free to chime in.

It'd be nice to put RACs beside them...anyone want to help? I figure we need 2 RACs that are within 10% to validate or we get inconclusive and need a third opinion.

R

Profile -= Vyper =-
Volunteer tester
Avatar
Send message
Joined: 5 Sep 99
Posts: 1039
Credit: 300,848,558
RAC: 154,990
Sweden
Message 1026032 - Posted: 18 Aug 2010, 7:06:51 UTC - in response to Message 1026031.

You can move the bunch from GTX260 to GTX285 past GTX460.

I can vouch for that :)

Kind regards Vyper
____________

_________________________________________________________________________
Addicted to SETI crunching!
Founder of GPU Users Group

Profile Frizz
Volunteer tester
Avatar
Send message
Joined: 17 May 99
Posts: 271
Credit: 5,852,934
RAC: 0
New Zealand
Message 1026033 - Posted: 18 Aug 2010, 7:17:07 UTC - in response to Message 1026031.

Great idea!

One question and one suggestion:

- Could you please define how to calculate RAC for this survey? So we really get comparable numbers. Things to take into account: stock clocks vs. overclocked, stock application vs. optimized apps, how many days for RAC calculation, etc.

- Can we please include wattage, e.g. points/watt/day or something like.
____________
Petition against 1366x768 glare displays: http://www.facebook.com/home.php?sk=group_153240404724993

Profile [seti.international] Dirk Sadowski
Volunteer tester
Avatar
Send message
Joined: 6 Apr 07
Posts: 7018
Credit: 59,131,626
RAC: 20,248
Germany
Message 1026063 - Posted: 18 Aug 2010, 10:46:38 UTC
Last modified: 18 Aug 2010, 10:47:53 UTC

I don't know if the new Cr.-System is finished balanced..

In past for my manufacturer OCed GTX260-216 @ 680/1500/1250 MHz -> ~ 15,000 SETI@home-RAC.

With WinXP 32bit, stock MB 6.09 CUDA, nVIDIA driver 190.38 . Max performance on my systems.

GPU: ~ 40 W idle, ~ 140 W if SETI@home CUDA.
____________
BR



>Das Deutsche Cafe. The German Cafe.<

Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026150 - Posted: 18 Aug 2010, 16:16:05 UTC
Last modified: 18 Aug 2010, 17:00:47 UTC

@Vyper, I was attempting to rank them based on theoretical performance, not current performance, although perhaps I should change that. I also find that my 460 is faster than my 280 was, by a hair. I will put a note in there though.


Frizz, yes, I like your RAC method. makes sense, the one thing is that there is a 30 second or so load time on each unit, so what I would propose for methodology is (using excel or other spreadsheet likely) take your last 20 - 50 WU's crunched on the card sum up the credit call it A, sum up the gpu crunch time in seconds, call it B, add 30 * Number of units and call that C.

Now take A/(B+C) to get credit per second. Multiply by 86400 to get Credits per day which should be theoretical maximum RAC.


Yes adding power usage would be nice. It's tough to do though unless you have a motherboard with onboard video that you can compare with. Just measuring at idle and then load is poor methodology. We can try to estimate though. So far GTX 460 running full load is about 120 Watts...keeping in mind FERMI apps, my GTX 280 ran at a ridiculous 210+ or so (OC model), the 9800 seems to take about 90 meaning it is pulling some through the MB.


I'd like to try and go for base stats (reference boards) but that is impossible since no one runs them so I'll go for close. Obviously different OSes, different manufacturers and different driver versions, CUDA versions etc will have differences (especially CUDA versions). Lets go for the fastest RAC on the stock clocked board, so optimized apps, but no huge overclocks...if you're OCed we can compensate a bit.

It'd be really nice if I could edit the old post with updates instead of making them down the thread.

Profile [seti.international] Dirk Sadowski
Volunteer tester
Avatar
Send message
Joined: 6 Apr 07
Posts: 7018
Credit: 59,131,626
RAC: 20,248
Germany
Message 1026161 - Posted: 18 Aug 2010, 16:58:22 UTC - in response to Message 1026150.
Last modified: 18 Aug 2010, 17:06:50 UTC

(...)
Yes adding power usage would be nice. It's tough to do though unless you have a motherboard with onboard video that you can compare with. Just measuring at idle and then load is poor methodology. We can try to estimate though. So far GTX 460 running full load is about 130 Watts...keeping in mind FERMI apps, my GTX 280 ran at a ridiculous 210+ or so (OC model), the 9800 seems to take about 90 meaning it is pulling some through the MB.
(...)


My E7600 with onboard GPU had idle: ~ 52 W.
Added the GTX260-216, the whole system at idle: ~ 95 W.
Full loaded only with SETI@home CUDA: ~ 215 W. Idle CPU.

(Screen connected to onboard, GTX260-216 only for CUDA)

This means on my/this system: ~ 40 W idle GPU, additional ~ 120 W if SETI@home CUDA.

So ~ 160 W additional for this system after adding the GTX260-216 (manufacturer OCed, GIGABYTE GTX260 SOC).

You see also different W usage at different projects.
SETI@home use the GPU and GPU RAM -> more W and higher temps.
MilkyWay@home use the GPU different, because in GPU-Z no memory controller load compared to SETI@home -> less W and lower temps.

IIRC, the whole system took ~ 190 W if only MW@h CUDA. Idle CPU.

So ~ 25 W difference between S@h and MW@h CUDA on my graphic card.
____________
BR



>Das Deutsche Cafe. The German Cafe.<

Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026163 - Posted: 18 Aug 2010, 17:08:14 UTC - in response to Message 1026161.

Thanks Sutaru

You've got a good setup for getting power readings.

I think we should go for SETI readings only, not other projects.

For MilkyWay Only a small number of the cores are double precision capable on Nvidia GTX 2XX, so it may read as 100% use but it can't actually use all the SPs in CUDA. This is why ATI cards crush Nvidia cards at MW. All of their SPs are double precision, my 3870 (3 years old) can do 50,000 RAC on MW if I get WUs.

Lets try for power readings in SETI only though for this little project.

Profile Tim Norton
Volunteer tester
Avatar
Send message
Joined: 2 Jun 99
Posts: 835
Credit: 33,540,164
RAC: 0
United Kingdom
Message 1026169 - Posted: 18 Aug 2010, 17:52:16 UTC - in response to Message 1026150.

Frizz, yes, I like your RAC method. makes sense, the one thing is that there is a 30 second or so load time on each unit, so what I would propose for methodology is (using excel or other spreadsheet likely) take your last 20 - 50 WU's crunched on the card sum up the credit call it A, sum up the gpu crunch time in seconds, call it B, add 30 * Number of units and call that C.

Now take A/(B+C) to get credit per second. Multiply by 86400 to get Credits per day which should be theoretical maximum RAC.

razamatraz

Great idea this could have used the list when i was trying to decide if my 8600GTS was worth the hassle of putting into an old machine for a bit more RAC :)

questions on you calculation - how do we get the credit scores easily as with all the huge pending lists we have it could take a while unless we fiddle around with a validated list of wu? But the bigger problem i see is that the amount of credit you get from SETI is very variable i.e. in part its dependant on your wingmen and their run times etc so will give you a false impression if your set of chosen wu are biased one way or the other. I haven't looked but is there an easy to follow calc based on the work done cobblestones(?) i think which we could use our run times to calc a credit score or is this some mystic art know only to the SETI gods?

your proposed calc is fine its just getting "A" that we need to refine :)

yes the resulting numbers would be a theoretical maximum (subject to variables of OS, driver versions, OC settings etc) but i do not think that matters as we a trying to get a qualitative measure of what throughput to expect to do the comparisons etc

If we also publish the range of possibilities per GPU type as an average would that be sufficient for people to gauge the likely speed a card type would give them or do we need to be more specific than that as different types of the "same" GPU design can have moderately different speed settings enough to change the calc enough to be significant - maybe we have to wait till the results are in

Other things we may have to note is what the gpu is doing i.e. is it a dedicated cruncher or running a monitor or two? I have two GTX260's and one is basically a dedicated cruncher the other has two 24" monitors at high res hanging off it as well which gives it a bit more work to do and you can see the average RAC being lower etc

Anyway have waffled on too long - over to others to comment


____________
Tim

Profile Dave Cummings
Volunteer tester
Send message
Joined: 16 May 09
Posts: 204
Credit: 919,706
RAC: 12
United Kingdom
Message 1026177 - Posted: 18 Aug 2010, 18:27:38 UTC

I am looking in the next couple of months at one of these:

http://www.scan.co.uk/Products/768MB-Palit-GTX-460-3600MHz-GDDR5-GPU-675MHz-Shader-Clock-1350MHz-336-Cores-DVI-D-Sub-HDMI

how does that rate? I currently have a 89400gt 1gb

Profile Dirk Villarreal Wittich
Avatar
Send message
Joined: 25 Apr 00
Posts: 2098
Credit: 365,578
RAC: 0
Holy See (Vatican City)
Message 1026179 - Posted: 18 Aug 2010, 18:38:34 UTC - in response to Message 1026177.


I am looking in the next couple of months at one of these:

http://www.scan.co.uk/Products/768MB-Palit-GTX-460-3600MHz-GDDR5-GPU-675MHz-Shader-Clock-1350MHz-336-Cores-DVI-D-Sub-HDMI

how does that rate? I currently have a 89400gt 1gb


Just making the link clicable--->scan.co.uk
____________

Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026182 - Posted: 18 Aug 2010, 18:53:09 UTC - in response to Message 1026177.
Last modified: 18 Aug 2010, 19:04:26 UTC

@Dave I have a 460, not that model, but mine is capable of about 14,000 RAC right now...software updates may improve that...puts it in the category of the 260 - 285 but where in that mix is still to be determined, Nvidia is still on the first driver version that supports it as well so there is room to get better.

I'd fully expect a 15 - 20 times improvement over a 9400 GT only considering GPUs. your Athlon X4 is likely doing a good part of your crunching though.

@Tim....I would suggest that we take the last 20 - 50 validated work units. The larger the number of WUs you use the better the calculation, it should help smooth out things like wingmen crunch times, VLARs, VHARs etc. Lately I've been finding more consistent crunch time to credit ratios than before too, hopefully that holds.

As for GPU use, I generally don't use crunch on my GPUs when the computer is in use and I don't think having idle monitors hooked up should slow it down much. If a lot of your units are with other graphically intense things going on though they could be significantly slowed.

There is no way to be precise, but having them in order with power consumption values and approximate RAC would be nice even if it is +/- 15%

Sten-Arne
Volunteer tester
Send message
Joined: 1 Nov 08
Posts: 3327
Credit: 18,961,665
RAC: 25,549
Sweden
Message 1026185 - Posted: 18 Aug 2010, 18:56:35 UTC - in response to Message 1026031.

So I was thinking it'd be nice to have a ranking system for performance on currently available Nvidia cards for SETI. I figure it should go something like this.

8400GS 512 meg only
9400 GT
210/310 (very small improvement over 9400 just faster clock speed)
9500 GT
GT 220/315
9600 GT

GT 320 (OEM only, I've never seen one.)
GT 240/GT 340


9800 GT (formerly known as 8800 GT)
9800 GTX

GTS 250 (formerly known as 9800 GTX+)
GTX 260 (there were 2 versions of this, the 192 and 216 versions)
GTX 280
GTX 275
GTX 285

GTX 460
GTX 470

GTX 295

GTX 480


The GTX 280 fits in around 275 level, but is unavailable. The 460 is potentially faster than the 285, but FERMI code is not at that point yet so it is about even. The GTX 480 should be faster than a 295, but until FERMI code matures the 295 is actually faster.....correct me if I am wrong.

This is in theory, with no overclocking. Keep in mind there are many versions of each card, some of the lower end cards have versions with DDR2 and DDR3. memory size can matter, but not much as long as there is enough. Also we recently found that Frizz' 240 is faster than my 9800 GT, but I have been running FERMI on it so that may account for the difference.


Yes I know some cards are missing. They are not available at stores currently so I see no point in including them.

IF you think I am mistaken feel free to chime in.

It'd be nice to put RACs beside them...anyone want to help? I figure we need 2 RACs that are within 10% to validate or we get inconclusive and need a third opinion.

R


Do not forget the awesom, extremely powerful, CUDA capable, Nvidia ION GPU. Measured at mind boggling 35 GFLOPS.

:-)


____________

Profile John Clark
Volunteer tester
Avatar
Send message
Joined: 29 Sep 99
Posts: 16515
Credit: 4,418,829
RAC: 0
United Kingdom
Message 1026190 - Posted: 18 Aug 2010, 19:16:39 UTC

I hope the ATI cards will figure when SETI can run on these GPUs?
____________
It's good to be back amongst friends and colleagues



Profile Tim Norton
Volunteer tester
Avatar
Send message
Joined: 2 Jun 99
Posts: 835
Credit: 33,540,164
RAC: 0
United Kingdom
Message 1026211 - Posted: 18 Aug 2010, 20:00:41 UTC
Last modified: 18 Aug 2010, 20:01:10 UTC

Ok i'll go first with some results (limited at the moment)

Host http://setiathome.berkeley.edu/show_host_detail.php?hostid=5430117 - so you can check my maths :)

BFG GPU 8600GTS 256 meg, 710/1500/1008 (stock), win server 2003 x86 sp2, driver 197.45

dedicated cruncher (monitor off)

Using the valid results i have (13 No) and using razmatraz equation i get the following

RAC := (1301.84/(34310.01+390))*86400 = 3241.47

Results range from 25.49 to 152.91 credit with times varying from 588 to 3541 respectively - so quite a range

on average the card does a wu every 50 to 62 minutes so say 24 per day which would equate to 2400 if the average RAC per unit is 100.

our proposed calc is giving an average of about 135

Power no idea - is there a utility that will tell me?
____________
Tim

Profile Frizz
Volunteer tester
Avatar
Send message
Joined: 17 May 99
Posts: 271
Credit: 5,852,934
RAC: 0
New Zealand
Message 1026212 - Posted: 18 Aug 2010, 20:12:36 UTC - in response to Message 1026211.

@razamatraz: Could you please Edit/Update your initial posting as the results are comming in?
____________
Petition against 1366x768 glare displays: http://www.facebook.com/home.php?sk=group_153240404724993

Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026213 - Posted: 18 Aug 2010, 20:20:22 UTC - in response to Message 1026212.
Last modified: 18 Aug 2010, 20:51:45 UTC

It locks out edit after an hour....I assumed I could do that to, but it appears I can't.

I might have to host the list somewhere and link to it.

http://www.rjsharkey.com/SETI/GPU.htm

Try that for now. I'm at work and MS word is the only quick html editor...I'll do better from home.

Profile Tim Norton
Volunteer tester
Avatar
Send message
Joined: 2 Jun 99
Posts: 835
Credit: 33,540,164
RAC: 0
United Kingdom
Message 1026219 - Posted: 18 Aug 2010, 20:45:25 UTC - in response to Message 1026211.
Last modified: 18 Aug 2010, 20:56:29 UTC

Second one

Host http://setiathome.berkeley.edu/show_host_detail.php?hostid=5462467 - so you can check my maths :)

Zotac GPU GTX260-216 896 meg, 630/700/1300 (not stock - memory turned down to aid cooling), win7 64 bit, driver 257.21

dedicated cruncher (monitor off)

Using the valid results i have (57 No) and using razmatraz equation i get the following

RAC := (5533.42/(56271.9+1710))*86400 = 8245.46

Results range from 0.16 to 185.72 credit with times varying from 26.7 to 1063 respectively - so quite a range again

on average the card does a wu every 17 to 20 minutes so say 72 per day which would equate to 7200 if the average RAC per unit is 100.

our proposed calc is giving an average of about 114.5

Power no idea
____________
Tim

Profile Tim Norton
Volunteer tester
Avatar
Send message
Joined: 2 Jun 99
Posts: 835
Credit: 33,540,164
RAC: 0
United Kingdom
Message 1026224 - Posted: 18 Aug 2010, 20:54:49 UTC - in response to Message 1026219.

third one

is the same as previous post except its running at 650/700/1300 - same os and driver but has two monitors attached as its my main machine at high res but still runs very similar times

i only recently upped the clock to 650 and times improved to compare very well with the other GTX even though it is doing screen refresh etc while working on other stuff - not games

oh reason i have turned the memory down a lot - think stock way 1000 or so was to save on power/heat and these hosts were doing collatz and milkyway most of the time and memory is not a deciding factor on the overall speed for them - depending on what others find i may turn the memory back up in the future

right thats me done as the only other gpu i have is ATI! :)
____________
Tim

Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026229 - Posted: 18 Aug 2010, 21:14:34 UTC - in response to Message 1026224.

Thanks Tim, I will have to go through my cards when I get home from work. Can't spend too much time on here.

Profile razamatraz
Send message
Joined: 23 Oct 07
Posts: 142
Credit: 27,807,059
RAC: 0
Canada
Message 1026281 - Posted: 19 Aug 2010, 3:33:39 UTC - in response to Message 1026229.

So I just did my 3 cards.

#1: 9800 GT 512 Meg 600/900/1500 (Stock), on 197.13 Driver Win XP x64, FERMI app
with 23 units total time 35476 seconds, 2369 credits = RAC ~5816
this is on my dual GPU machine hooked to a 24" 19 * 12 screen.

#2: GTX 460 1 Gig 715/900/1430, on 258.96 Driver Win XP x64, FERMI app
with 25 units checked, total time 20768.4 seconds 2807 credits = a measly 11782 RAC
This is on the same machine, also driving a 24" 19 *12 monitor...and yes the drivers were different....the joys of XP x64...I've since updated, lets see if that helps.

#3: 9800 GT 512 Meg 600/900/1500 (Stock), on 197.45 Driver Win XP 32 bit Lunatics app
with 25 units checked, total time 30499 seconds, 2288 credits = 6540 RAC
this is on an old Pentium D at 3.4 Ghz driving a single 19" 12 * 10 monitor


On the 460 there were a couple of outliers, one really bad one, that had CPU like speed (mislabeled VLAR maybe) take that one unit out and the 460 jumps to 13000, but for now I will leave it at 11782...Another test to follow after the shutdown.

1 · 2 · 3 · 4 . . . 5 · Next

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

Copyright © 2014 University of California