Milkyway@Home - GPU performance statistics

[Sunny129]

If you want to lower your temps in M@H, downlock your ram. My 4870 ram is running at 175mhz in stead of the default 900 with no effect on completion times. However, this doesn't work on Collatz projects, because those ones need a lot of memory bandwidth unlike M@H.

thanks for the tip...i'm gonna give that a try. my 5870 fluctuates between the mid 40's and upper 50's over the course of the day (though it probably averages ~55°C), but the stock blower has to run at 60% in order to keep it that cool under a 100% MW@H load. i don't know what that is in terms of RPM b/c my monitoring software only gives me a percentage readout, but i will say that at the currrent speed, it sounds like a small vacuum or a dust buster - in other words, LOUD. its noticeably quieter at 50%, but then the temps creep up into the upper 60's...and with the blower at only ~21% (auto setting) it is virtually silent, but temps will then reach the upper 70's to low 80's under full load.

*EDIT* - i seem to be having trouble trying to downclock the GPU memory. i'm using MSI Afterburner, and every time i lower the memory clock and hit the apply button, the memory clock jumps back up to the default 1200MHz. i did some research and found that its actually somewhat of a common problem with quite a few possible causes. i do believe i set the program options properly, and i made the appropriate changes to the MSIAfterburner.cfg file. i don't have a non-reference GPU w/ non-reference VRM's, so there's no reason i shouldn't be able to both overclock/underclock and overvolt/undervolt. my guess is that there are communication issues between the particular version of MSI i'm using and the particular version of Catalyst GPU drivers i'm using. i guess i'll have to experiment w/ different versions of each and see if that fixes my inability to overclock and overvolt with MSI Afterburner...

 

[Bradtech519]

thanks for the tip...i'm gonna give that a try. my 5870 fluctuates between the mid 40's and upper 50's over the course of the day (though it probably averages ~55°C), but the stock blower has to run at 60% in order to keep it that cool under a 100% MW@H load. i don't know what that is in terms of RPM b/c my monitoring software only gives me a percentage readout, but i will say that at the currrent speed, it sounds like a small vacuum or a dust buster - in other words, LOUD. its noticeably quieter at 50%, but then the temps creep up into the upper 60's...and with the blower at only ~21% (auto setting) it is virtually silent, but temps will then reach the upper 70's to low 80's under full load.

*EDIT* - i seem to be having trouble trying to downclock the GPU memory. i'm using MSI Afterburner, and every time i lower the memory clock and hit the apply button, the memory clock jumps back up to the default 1200MHz. i did some research and found that its actually somewhat of a common problem with quite a few possible causes. i do believe i set the program options properly, and i made the appropriate changes to the MSIAfterburner.cfg file. i don't have a non-reference GPU w/ non-reference VRM's, so there's no reason i shouldn't be able to both overclock/underclock and overvolt/undervolt. my guess is that there are communication issues between the particular version of MSI i'm using and the particular version of Catalyst GPU drivers i'm using. i guess i'll have to experiment w/ different versions of each and see if that fixes my inability to overclock and overvolt with MSI Afterburner...

My 4890 was really hot.. I ordered a used 5850 with a Zalman on it.. Went from 90-95c temps on my 4890 with stock to 50-60c load gpu temps in my case. When I'm crunching my cpu gets hotter than my GPU.. CPU is still stock and GPU is zalman. I think the 40-50 is worth the investment.

 

[Sunny129]

thanks for the thread bump...i almost forgot to post up on my progress.

so last night i uninstalled the Catalyst 11.1 drivers i've been using for some time and installed the 11.3 drivers that were just released by AMD a few days ago. first and foremost, i no longer have problems getting the Catalyst Control Center to work correctly like i did with all the previous driver versions i've tried. secondly, i no longer have problems overclocking/underclocking the GPU core and GPU memory w/ MSI Afterburner. however, this particular advent was contingent upon the Catalyst Control Center working correctly. in other words, i have to open the CCC and enable "ATI Overdirve" (the OCing features in the CCC). note that i am not limited to directly manipulating the GPU core and memory clocks via ATI Overdrive in the CCC - i am also able to do this via MSI Afterburner so long as ATI Overdirve is enabled in the CCC. if i disable ATI Overdrive, then i can no longer manipulate the GPU core and memory clocks via MSI Afterburner. strange, i know...but while MSI Afterburner doesn't work properly independently, it does work properly when ATI Overdrive is enabled, and that's good enough for me. otherwise i'll have to spend hours playing with different versions of Catalyst drivers (and possibly different versions of MSI Afterburner) to find a combination of the two that will allow MSI Afterburner to work independently.

my second experiment involved downclocking the GPU memory since i now have the ability to manipulate it in either ATI Overdrive or MSI Afterburner. while i know its common knowledge on the MW@H message boards, i don't know if its been mentioned here in the ATDC forum - MW@H tasks require very little system memory (and consequently, very little GPU memory) when transferring to and from the GPU at initialization/completion. as a result, one can downclock his/her GPU memory frequency significantly without decreasing the crunching efficiency of his/her GPU or increasing the run times of his/her MW@H tasks. the GPU's memory bandwidth alone is more than adequate to make up for any decrease in GPU memory clock frequency when crunching MW@H tasks. the major benefit of downclocking your GPU's memory is that it'll drop your GPU temps a few °C. that may not sound like alot, but it can seem like a world of difference when you factor in fan speed and noise level. initially, when i was unable to manipulate my GPU memory clock and had to remain at the default 1200MHz, i would have to run my GPU blower at 60% just to keep temps at 55°C or below. with the GPU memory now downlocked to 600MHz, the highest temp i see is 51°C (with the blower still at 60%). when i drop the blower speed to 40%, the highest temp i see is 57°C, which is perfectly acceptable. so its nice to downclock your GPU memory and be able to either 1) decrease your GPU temps, or 2) decrease your fan/blower speed and thus reduce noise.

note that this is only for MW@H and other GPU projects/apps whose tasks require very little memory at initialization and completion. Collatz Conjecture stands in sharp contrast to MW@H in that the data sets are huge in comparison, and can take up to ~300MB of system memory (and thus ~300MB of GPU memory) upon initialization and completion (as opposed to the only ~30MB of system and GPU memory taken up by a MW@H task). so the performance of projects like Collatz Conjecture will surely benefit from a higher memory clock and suffer from underclocked memory.

 

[Bradtech519]

Took my Diamond 5850 with zalman 1:21 to complete last WU from Milkyway. De_separation_13s

 

[Sunny129]

Took my Diamond 5850 with zalman 1:21 to complete last WU from Milkyway. De_separation_13s

how much credit did you earn for that particular WU? the reason i ask is b/c 1) as of a few days ago, MW@H started circulating new tasks that only net ~160 credits (as opposed to the ~213-credit tasks whose run times we've been documenting in the list on on page 1), and 2) your run time of 81s is on par with the ~213-credit WU's run times typical of a 5870 GPU or a 6950 GPU. perhaps the 81s WU you documented was one of the newer ~160-credit WU's?...b/c i can't imagine that a 5850 is finishing ~213-credit WU's in as little as 81s...i could be wrong though - could you please verify?

TIA,
Eric


EDIT - by the way, my 5870 is currently taking approx. 76s to finish the ~160-credit WU's

 

[Assimilator1]

Interesting about GPU RAM clock speeds & MW@H, however I use my card for gaming too so I don't want to cripple it, or have to keep changing speeds. Plus my old 4830 doesn't get that hot anyway .

Took my Diamond 5850 with zalman 1:21 to complete last WU from Milkyway. De_separation_13s

Yea as said, run time without granted credit numbers is meaningless

 

[tparin]

I'm clocking 3:19 on milkway & about 7:07 on Seti on a EVGA 570OC paired with Intel E6750 OC 'ed 3.2 ish running Abit IP 35 pro FSB 1600

 

[Sunny129]

I'm clocking 3:19 on milkway & about 7:07 on Seti on a EVGA 570OC paired with Intel E6750 OC 'ed 3.2 ish running Abit IP 35 pro FSB 1600

welcome to the forums . i was actually hoping a few more nVidia GPU crunchers would pop in and share their MW@H times. i think its common knowledge that ATI GPUs have the edge when it comes to MW@H, but i was never clear about the magnitude of the difference. likewise, i know nVidia GPUs rule the F@H realm, but again, i have no idea how much more efficient nVidia GPUs actually are.

what kind of SETI WU did you fininsh in 7:07? i take it this was a Multibeam task? that's about twice as fast as my 5870 . looks like nVidia has a leg up in S@H. although i can't complain that at ~15 min. per MB task, my GPU still finishes them 4-5 times as fast as a CPU core does. and i love that AP tasks seem to run approx. 18-20 times as fast on the GPU as they do on a CPU core.

 

[zzuupp]

Welcome, tparin!

 

[RussianSensation]

Took my Diamond 5850 with zalman 1:21 to complete last WU from Milkyway. De_separation_13s

There is more than 1 type of De_Separation_13. I think the time you are reporting is for a smaller unit than ~ 213. Otherwise, this is too fast for a 5850 since it nearly matches an HD5870, unless your 5850 is overclocked?

I'm clocking 3:19 on milkway & about 7:07 on Seti on a EVGA 570OC

~ 213 work unit? What's the overclock on the 570 SC (797mhz)?

Welcome to the Forums tparin!

 

[Assimilator1]

I'm clocking 3:19 on milkway & about 7:07 on Seti on a EVGA 570OC paired with Intel E6750 OC 'ed 3.2 ish running Abit IP 35 pro FSB 1600

Hi, welcome to AT forums

Btw WU times without granted credit numbers are meaningless , different WUs take different times (SETI & MW). And yea catching MW granted credit stats are a PITA .

 

[RussianSensation]

And yea catching MW granted credit stats are a PITA .

You don't need to "eye catch" or "estimate" the times. They are recorded.

1. Log into your account at http://setiathome.berkeley.edu/.
2. On the bottom of your account page will see the projects you currently participate in, listed under "Projects in which you are participating"
3. Click Milkyway@home to bring you to your Username statistics screen for MilkyWay.
4. Find a section labelled Computers. Click "View".
5. Under "Computers Beloging to *Your Username" screen, click "Tasks".

There you will see your Run time (sec) and Credit granted (213.76). :thumbsup:

 

[Sunny129]

You don't need to "eye catch" or "estimate" the times. They are recorded.

1. Log into your account at http://setiathome.berkeley.edu/.
2. On the bottom of your account page will see the projects you currently participate in, listed under "Projects in which you are participating"
3. Click Milkyway@home to bring you to your Username statistics screen for MilkyWay.
4. Find a section labelled Computers. Click "View".
5. Under "Computers Beloging to *Your Username" screen, click "Tasks".

There you will see your Run time (sec) and Credit granted (213.76). :thumbsup:

he knows that. i think what he's saying is that the data driven pages of the MW@H server only show valid tasks for a minute or two before wiping them from the "valid tasks" list. in other words, from the instant that a "ready to report" MW@H task gets reported to the server and disappears from your BOINC clients task list, you only have a minute or two to go to the MW@H website and see the results there before it disappears there too.

 

[Assimilator1]

Yep, sunny you got it dead on .

Stupid MW website clears them far too quickly!

 

[RussianSensation]

he knows that. i think what he's saying is that the data driven pages of the MW@H server only show valid tasks for a minute or two before wiping them from the "valid tasks" list. in other words, from the instant that a "ready to report" MW@H task gets reported to the server and disappears from your BOINC clients task list, you only have a minute or two to go to the MW@H website and see the results there before it disappears there too.

Ya, good point.

Hey Sunny, I am going to follow your suggestion and try downclocking the GDDR5 on my 6950. You think going down to 625mhz is safe? (that's the lowest setting I see in MSI Afterburner)

 

[Sunny129]

Ya, good point.

Hey Sunny, I am going to follow your suggestion and try downclocking the GDDR5 on my 6950. You think going down to 600mhz is safe?

hmm...i don't know. nobody has ever asked me that. i will say that i've never seen an instance in which downclocking a CPU, system memory, a GPU core, or GPU memory was unsafe and caused any damage. i would think the worst that could happen is that the memory becomes a bottleneck to the rest of the GPU or perhaps the entire system...but i would imagine that would only happen while playing GPU memory-intensive games or using video rendering software. for MW@H, the idea is that a WU requires so little GPU memory that it can be downclocked on certain video cards to a certain extent, and the GPU's memory quantity and bandwidth is more than enough to make up for the downclock. in other words, MW@H tasks seem to be more memory bandwidth-intensive than they are memory clock-intensive. b/c your 6950 and my 5870 are both 2GB, 256-bit cards, we both have more than enough memory and memory bandwidth to make up for any reduction in memory clock frequency.

i don't exactly know where the line is drawn, or at what point folks will start to notice a significant decrease in MW@H GPU crunching efficiency due to downclocking the GPU memory. but i would imagine that the less memory your GPU has, and the smaller the bandwidth of the memory bus is, the less you'll be able to downclock the memory without any adverse effects on crunching efficiency. but if i can cut my GPU memory freq. in half without increasing the time it takes to complete a MW@H WU on my 5870, then i would think you should be able to do the same on your 6950 without increasing crunch times. if you attempt it, be sure to report back. i've love to know how much your GPU temps drop.

 

[tparin]

The OC is 800 1600 2000 running @ 57c .

 

[RussianSensation]

i've love to know how much your GPU temps drop.

I downclocked the memory to the lowest possible which is 750mhz on my card 6950 card (won't allow the slider to go below that value in MSI Afterburner). The temperature dropped 2-3*C at 99% load vs. running the memory at 1250 mhz). So looks like there is not a lot of change for me. I will do further testing. Alternatively, when I bump my GDDR5 from 1250 to 1375mhz, the GPU temperatures don't rise either.

The OC is 800 1600 2000 running @ 57c .

That's 3 min 19 sec for a 213 work unit? Confirm and I will put it up on the front page. Seems unusually fast for a Fermi architecture.

 

[Sunny129]

I downclocked the memory to the lowest possible which is 750mhz on my card 6950 card (won't allow the slider to go below that value in MSI Afterburner). The temperature dropped 2-3*C at 99% load vs. running the memory at 1250 mhz). So looks like there is not a lot of change for me. I will do further testing. Alternatively, when I bump my GDDR5 from 1250 to 1375mhz, the GPU temperatures don't rise either.

thanks for following up. looks like your GPU memory frequency experiment results are right in line with mine. that is, your GPU temps dropped 2-3°C as a result of decreasing your GPU memory freq. by 500mhz (from 1250mhz to 750mhz)...and my GPU temps dropped ~4°C by reducing the GPU memory clock by 600mhz (from 1200mhz to 600mhz). granted, i suppose its not much of a difference...and perhaps its not worth downclocking the memory if you also use the GPU for other activities that require a higher GPU memory clock and you don't want to switch back and forth regularly.

one thing i never tried was overclocking my GPU memory. of course i would only do it to see if GPU temps rise. we know that if there's no lost efficiency while crunching MW@H tasks when downclocking the GPU memory, we can safely say that overclocking the GPU memory isn't going to affect MW@H crunching efficiency either.

 

[salvorhardin]

Added a 5850 to my computer here are some of the times I am getting:

points          4830(700-920)     4830(575-900)     5850(725-1000)
159.86                                250s                  93s   
213.76              275s              334s                 121s        
267.2                                 415s                 151s

The 4830@700 was run by itself with a single screen. The 4830@575 and the 5850 were run together in the same system with dual screens. I also ran the 5850 by itself with single display and it was about 2% faster when compared with running dual screens and 2 gpus.

 

[RussianSensation]

Added a 5850 to my computer here are some of the times I am getting:

points          4830(700-920)     4830(575-900)     5850(725-1000)
159.86                                250s                  93s   
213.76              275s              334s                 121s        
267.2                                 415s                 151s

The 4830@700 was run by itself with a single screen. The 4830@575 and the 5850 were run together in the same system with dual screens. I also ran the 5850 by itself with single display and it was about 2% faster when compared with running dual screens and 2 gpus.

Great! Thanks a lot of these results. I updated the chart to reflect your times.

 

[Sunny129]

ok RussianSesnsation, i'm calling on your expertise again...

a while back (on page 3 of this thread) we went over the calculations for determining the double precision performance of some ATI GPUs. i thought i had a good understanding of it after you explained it to me in detail, but now i'm not so sure anymore.

i'll use the 5870 as an example - if you'll recall, you calculated the double precision performance of this GPU using the following equation:

2 * 1600 * 850 * 1/5 = 544,000 FLOPs (or 544 GFLOPs), where 2 is the # of 64-bit floating point MUL or ADD operations per clock (which you quoted from THIS AT article), 1600 is the # of SMs, 850 is the clock frequency, and 1/5 is the fraction of single precision performance at which the 5870 performs double precision floating point operations. my question is this - why is the double precision performance 2 * 1600 * 850 * 1/5 = 544 GFLOPs, and not 2 * 1600 * 850 = 2720 GFLOPs (or 2.72 TFLOPs)? i could understand if 2 * 1600 * 850 = 2.72 TFLOPs were a measure of the 5870's single precision performance - then it would make sense to multiply by 1/5 to get the double precision performance. but 2 * 1600 * 850 = 2.72 TFLOPs isn't a measure of single precision performance - it IS the measure of double precision performance. after all, the 2 in that equation is the # of 64-bit (double precision) floating point MUL or ADD operations per clock, and not 32-bit (single precision) floating point MUL or ADD operations per clock (assuming that bit of information in the article isn't a typo).

is there perhaps a typo in AT's article? should that not be listed as the FP32 (single precision) performance? if the 5870's FP64 performance is truly only 544 GFLOPs, then single precision performance must be 5 times greater, or 2.72 TFLOPs. and if single precision performance is 2.72 TFLOPs, and given that we know the clock freq. is 850mhz and the GPU has 1600 SMs, then the single precision performance of the 5870 must only be 2 32-bit (single precision) floating point MUL or ADD operations per clock, making the 64-bit (double precision) performance 1/5 of that value, or 0.4 64-bit (double precision) floating point MUL or ADD operations per clock.

what am i missing/overlooking here?

TIA,
Eric

 

[sangyup81]

Just bought a new built by ati 4770 from ebay for $56 including shipping! I have it overclocked to 785 MHz from the stock 750 MHz.

I'll provide a 213 credit time once I get one of those validated. For some reason, I only get 159.86 and 320.63 credit work units lately. In case anyone is wondering I got 3:14.66 for the 159.86 and 6:15.27 for the 320.63.

EDIT: I have a result! 213.76 credits 250.66 seconds or ~ 4 minutes 11 seconds. Overclocking helps a little I guess =)