Milkyway@Home - GPU & CPU performance stats wanted, any capable h/w, old or new!

Page 4 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

deerslayer

Lifer
Jan 15, 2001
10,153
0
76
Lol, did you find out if your GPU was nr max load?

I'm not sure if I did it or not. I have the flu right now. My wife locked me in quarantine (the basement) all weekend. I'm back to work now, but I still feel like I got hit by a truck.
 

GLeeM

Elite Member
Apr 2, 2004
7,199
128
106
I started up again on the OCed 7950 - 753 tasks so far and no mod fit type?

I'll keep watching :)
 

biodoc

Diamond Member
Dec 29, 2005
6,256
2,238
136
I'm not sure if I did it or not. I have the flu right now. My wife locked me in quarantine (the basement) all weekend. I'm back to work now, but I still feel like I got hit by a truck.

Hope you are feeling better soon deerslayer. :thumbsup:

Quarantined with my computers sounds like a good retirement plan for me! :biggrin:
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
I started up again on the OCed 7950 - 753 tasks so far and no mod fit type?

I'll keep watching :)
if you're truly having trouble getting MW@H Separation Modified Fit tasks and, you're certain that the MW@H Separation Modified Fit application box is checked (via your Milkyway@Home web preferences), the next thing to do would be to make sure the other GPU applications (MW@H and MW@H Separation) are unchecked...that way MW@H Separation Modified Fit tasks are the only kind of tasks you get.
 

deerslayer

Lifer
Jan 15, 2001
10,153
0
76
Hope you are feeling better soon deerslayer. :thumbsup:

Quarantined with my computers sounds like a good retirement plan for me! :biggrin:

I'll feel better when spring gets here. I'm tired of looking at my bike sitting in the corner of the garage looking all sad and lonely.
 

GLeeM

Elite Member
Apr 2, 2004
7,199
128
106
the next thing to do would be to make sure the other GPU applications (MW@H and MW@H Separation) are unchecked

I selected only mod fit and if no WUs for selected send others and got this message in log:
2/12/2014 7:15:31 PM | Milkyway@Home | Message from server: Catalyst driver version is not OK for CAL application with this GPU

CAL: ATI GPU 1: AMD Radeon HD 7870/7950/7970 series (Tahiti) (CAL version 1.4.1741

The above line I copied from the beginning of BOINC event log. Is my driver version too new or too old?

I have Catalyst Version 13.4
 

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
I use 13.1 atm, I hear 13.9 works, which I'll be trying soon.

New time for my HD 5850, GPU now at 800 MHz, 223s for 213.76s :), 3.22% clock speed increase cut WU times by 3.46%.
I won't add that time as I'm still tweaking & I don't want to spam the table with my 5850 enteries ;).

HD 5850, GPU 825 MHz, 213.76 WUs in 216s.

Wonder if I can get it to beat a 5870? ;)

Gleem
Any luck with the 7950?
 
Last edited:

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
Right, I've just tried 13.4, 13.9 & 14.1b v1.6 & they all give errors on MW v1.02 WUs on my 5850 & from what I gather on the relevant MW forum thread they will do on 6900s too.

It's not so clear with the 7900s though, their seemed to be mixed results, IIWU I would try 13.1.
Which ever ones you do try make sure you un-install all grx related s/w using the cat manager>reboot to safe mode> use Display Driver Uninstaller (or equiv., free d/l) to 'deep' clean it>reboot> install new driver> reboot once more! Without doing this I could not get 13.1 working.
 

GLeeM

Elite Member
Apr 2, 2004
7,199
128
106
Is my driver version too new or too old?
It's not so clear with the 7900s though, their seemed to be mixed results, IIWU I would try 13.1.
I guess I should have added that I probably would not go back to older version unless it would give MUCH better ppd (it takes 46 seconds for 159.86 points running one WU)

I run other GPU projects on the computer. Right now the 6870 is running Collatz (trying to move from 94 to 100M), a few days ago both GPUs were running Moo! Wrapper which I have to go back to for three days to move to 25M (I'm at 22M). After moving MW from 48 to 50M, there is Seti 2.1 to 2.5M, then PG 35 to 50M, POEM if it starts again 21 to 25M, and then everything to the next Milestone ad infinitum :)
 

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
HD 5850, GPU 850 MHz, 210s for 213.76 WUs.

Gleem
I didn't notice nor am I aware of any performance difference between the different drivers, but then I wasn't particularly looking for it either. I did crunch some WUs on the newer drivers but I don't think I could identify them from the others.
However 13.1 certainly does allow crunching of all WU types.

........I've only just realised your having problems with WUs that aren't normally a problem! lol. With post 13.1 drivers its usually the MW 1.02 WUs that error out not MW Sep. mod fit, don't know what's going on with your machine then :confused: , I'm gonna check out the MW forums.........
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
Gigabyte 7970, core @ 1000MHz (factory OCed), VRAM @ 850MHz (underclocked), BOINC v7.2.28, Catalyst 13.9, Win7 x64: ~69s

it is interesting that i'm getting the same average run time while running 1 tasks at a time as i did while running 2 tasks at a time. naturally, my current sample population of run times have a much lower variance and standard deviation than the previous sample population that i collected while running 2 tasks at a time on each GPU.

also Mark, i forgot to missed your question about my underclocking the VRAM earlier - VRAM clock has virtually no effect on MW@H task run times, so i underclock them to reduce power consumption (and thus temps). back when i first acquired some 69xx series GPUs, i was able to underclock the memory all the way down to 150MHz using early drivers. a mid-6xxx series driver update ended the days of such low underclocking, and from then on i was only able to underclock the VRAM as low as 700-800MHz...and with the 7970's i can underclock as low as 800-850MHz depending on the brand (and thus BIOS) of the card.


PS - these cards are air cooled, but if it gets cool enough, i'll leave the windows open tonight and bump the core clock up to 1100MHz...
 

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
Re VRAM, ah cool, worth knowing :), I think I'll produce a few benchmarks to see how it affects (or not) my 5850 (fancy posting a time for VRAM @default?). Btw, surely downclocking it to 150 MHz must of hurt WU times??

Re your single WU times vs 2xWUs, that's just bizarre! How is the even possible? :confused: lol. What's the GPU load at when crunching 1 WU?
I'll post the time anyway, thx :thumbsup:

Apparently cat 14.1 b significantly speeds up 7950s for MW, see updated table.
 
Last edited:

Fardringle

Diamond Member
Oct 23, 2000
9,183
751
126
In relation to the comments from a few people that said they never get the 213 point work units, I have had my HD 5670 set to run ONLY the Modified Fit task for over a week and in that time it hasn't received a single work unit from the project. (I don't want it to crunch full time right now so only set it to run the benchmark to get numbers for this thread.)

So maybe there is a conflict somewhere or an incompatible setting or something that makes it so a few of us just can't get those 'benchmark' work units?
 

GLeeM

Elite Member
Apr 2, 2004
7,199
128
106
So maybe there is a conflict somewhere or an incompatible setting or something that makes it so a few of us just can't get those 'benchmark' work units?
If in MW@H Preferences you say yes to "If no work for selected applications is available, accept work from other applications?" then you will probably get a message in the BOINC Event Log like I did: "Message from server: Catalyst driver version is not OK for CAL application with this GPU"
My driver it too old, I guess. Yours probably is also.
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
Re VRAM, ah cool, worth knowing :), I think I'll produce a few benchmarks to see how it affects (or not) my 5850 (fancy posting a time for VRAM @default?). Btw, surely downclocking it to 150 MHz must of hurt WU times??
actually no, even at a VRAM clock of 150MHz my MW@H task run times did not suffer...while a MW@H GPU task may require ~300MB VRAM, each task actually requires very little VRAM bandwidth. yeah i'll put the VRAM back to defualt and take an average here shortly...

Re your single WU times vs 2xWUs, that's just bizarre! How is the even possible? :confused: lol. What's the GPU load at when crunching 1 WU?
I'll post the time anyway, thx :thumbsup:
GPU load is currently at 99%, just as it was when i was running 2 tasks at a time. as far as the run time average being the same (at least to a decimal place or two), you have to remember that the advantage of running two MW@H GPU tasks at a time comes from the overlap of the two tasks during the few seconds that one of them is finishing its calculations on the CPU only, thus allowing the 2nd task to begin before the first task is truly 100% complete, and thus not allowing the GPU to sit idle during those few seconds (as would happen when running only one task at a time). as i stated in previous post, there is no increase in IPC efficiency to be had by running multiple MW@H GPU tasks at once, unlike many other BOINC projects. that is to say, MW@H GPU tasks take exactly twice as long to finish while running 2 at a time than they do running one at a time, if you ignore the few seconds overlap. so f you think about it, it makes sense that the average run time for tasks run 1 at a time vs 2 at a time are the same.

Apparently cat 14.1 b significantly speeds up 7950s for MW, see updated table.
i don't know that i'll be messing w/ beta drivers anytime soon, but we'll see...
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
ok, i feel like something about the work units has changed within the last several hours - i put the VRAM clock back to the reference value of 1375MHz, and my average run time increased to ~85s:

Gigabyte 7970, core @ 1000MHz (factory OCed), VRAM @ 1375MHz (factory/reference), BOINC v7.2.28, Catalyst 13.9, Win7 x64: ~85s

i don't know if the average run time jumped up to 85s just as soon as i increased the VRAM clock to 1375MHz, or if it happened some time after that...so i underclocked the VRAM down to 850MHz again, but my run times stayed right around the 85s mark (and did not go back down to ~69s). thus i suspect something about the work units has changed within the last several hours server side...hopefully i'm wrong for the benchmark's sake, but it wouldn't make sense if my 7970s were taking 85s to crunch at a core clock of 1000MHz. can anyone else confirm that their average run time for the 213.76-point tasks has gone up slightly within the last few hours?


TIA,
Eric
 
Last edited:

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
Amazing that MW needs so little bandwidth!

My 5850 is back down to 775 MHz, it benchmarked at 231s originally, currently 213.76s are taking ~230s, so pretty much the same. So no change in the WUs.

I don't know what's going on with your times, lol, very odd, are you sure your read the right WU times?
And yea I agree that increasing VRAM speed won't slow down WUs!
Maybe it's power throttling?? What is your power tune limit set to?

Oh btw I was getting mixed up about 1 WU vs 2 WU times, I thought you meant the total time for 2 WUs :$ lol.

Fardringle
What drivers are you using with the 5670?
My son's 4830 is on Cat 12.6 & crunches all WUs just fine.

Just a sec, does the 5670 have Double Precision?.............. Nope! just remembered that & confirmed on HD 5000 series wiki
Your grx card doesn't support DP so can't run MW.

Gleem
IIRC you using Cat 13.4? I think if anything, the driver is too new for MW!
 
Last edited:

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
I don't know what's going on with your times, lol, very odd, are you sure your read the right WU times?
And yea I agree that increasing VRAM speed won't slow down WUs!
Maybe it's power throttling?? What is your power tune limit set to?
yeah, i'm confident that i was looking at the 213.76-point WU's...i also considered the possibility that the GPUs might be throttling, but i'm not so sure that's the problem. as i said before, GPU load is always at 99% unless i'm at the end of a task and all the calculations are being done on the CPU - only then do i see GPU usage dip below 99%. powertune was set to 0% b/c i've never had issues before...so i see no reason why throttling problems would show up now. just to be safe though, i've since increased powertune to +20%, but still my average run time is hovering right around 85s. i'll have to look more into it this evening...
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
um yeah...i kinda screwed up there...

something funky must have happened while piecing my two machines back together, b/c what i thought was Cat 13.9 is actually Cat 12.1...i didn't realize it right away, and had to rely on GPU-Z to divulge the driver info (for some reason when i use the Intel IGP to run the display, AMD driver info doesn't show up entirely, or at all, in the CCC)...so that may explain my slower ~85s run times all of the sudden. anyways, i think my other mobo/SSD/HDD has the right BIOS settings and Cat 13.9 drivers on them...looks like i have another several hours of taking apart and putting back together my machines tonight :rolleyes:...
 

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
Don't you just need to install the 13.9 drivers?

Although I'm still confused, wasn't the latter 69s time done after the rebuild & thus also on the 12.1 drivers??
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
yes, technitcally all i need to do is install the 13.9 drivers...but i reorganized my components anyways b/c i wanted them exactly as they were before i took them apart. at any rate, the ~69s average run time was in fact calculated on the machine that now has the 12.1 drivers on it. i really don't know what's going on here, but i'll get to the bottom of it eventually. in fact i'd like to test 13.9 vs 12.1 just to see if that makes the difference between an average run time of 69s and 85s...unfortunately i went to fire up my other computer last night, and a funky burning smell came from the PSU and the system failed to power up...so it looks like i may have to RMA my SeaSonic X750...
 

Sunny129

Diamond Member
Nov 14, 2000
4,823
6
81
ugh...i'm an idiot. it turns out my current driver is 13.4, not 12.1...apparently i can't read LOL. now i'm as confused as ever. i looked in my driver folder and can't even find 13.9, so i must have never had 13.9 installed on this machine before...which means i've probably been running 13.4 for quite some time now. at any rate, i'm still trying to unravel the mystery...


*EDIT* - question...is everyone else getting fairly consistent run times for the 213.76 point tasks (i.e. a variance of no more than a few seconds give or take)? i ask b/c i'm consistently seeing run times of either ~60s or ~90s, but i have no idea what the ratio of ~60s to ~90s tasks i have, so i'm reluctant to calculate another average just yet. by the way, i see no evidence of GPU throttling in MSI Afterburner.

ok, due to the ~30s variance in my run times, and the fact that no run times actually fell anywhere between ~60s and ~90s (all were either ~60s or ~90s), i chose to use all completed and validated tasks whose time reported fell between 3:00:00 UTC and 4:00:00 UTC for my average...it ended up being a sample size of 30 tasks, instead of just 5 tasks. also of note is that in the process of reorganizing my hardware, i'm now crunching on a Sapphire 7970...the results are as follows:

Sapphire 7970, core @ 950MHz (factory OCed), VRAM @ 1375MHz (factory/reference), BOINC v7.2.28, Catalyst 13.4, Win7 x64: ~73s

something i found interesting is that this average run time (despite the oddly-behaved variance in my run times) is right in line with the average run time i initially got with the Gigabyte 7970 (~69s at a core clock of 1000MHz, 50MHz faster than the Sapphire GPU) before i started getting the ~85s run times that didn't make sense for a 7970.
 
Last edited:

Assimilator1

Elite Member
Nov 4, 1999
24,118
507
126
Uh-oh, I think the benchmark WU now has a problem! Seems you were right afterall Sunny! Don't know why you got hit so early though :confused:

This has only occurred today, but I now see that my 213.76 WUs with the GPU at 825 MHz completes in either 216s (as before) or 144s!
I was wondering if the overclock was suddenly unstable, however that's looking very unlikely as all the 'quick' 213.76's are being done in 144s +/- 0.1s. Just incase I'm testing now at 800 MHz.

Strangely most of the quick validated 213.76's I've crunched have apparently only been crunched by me! Bar 2 of them, the 1 user I checked (who also has a 5800 card) also has a similar variance, long ones taking ~248s & short ones ~168s (see here, TRAILBLAZER). By those times I reckon he's got a 5850 too.

So it now seems their can be 2 times for the 213.76 WUs!
Well seeing as their are only 2 different speed WUs it is still useable as a benchmark, so far! But if they add more different speed 213.76s WUs then it will scupper the benchmark.

For now, I'll say pick the 'long' 213.76 WU times & average 5 of them.

[update] Looking at another cruncher whose validated 1 of 'my' WUs further confirms the variance, DokV, he's got a 7900 series based card & is either getting ~50 or 77s on 213.76 WUs.

Sunny
I'll take the longer of your times ;).

[update2] Looking at 4-5 users the difference between the short & long WUs is ~50% +/- 5%, so the long ones take approx. 50% longer than the short 213.76's.

[update3] 5850 @800 MHz, 213.76 WUs ~223s & ~149s. So definitely not an overclocking issue.
 
Last edited: