Milkyway@Home - GPU performance statistics

[RussianSensation]

***I am creating a performance table for Milkway@Home for others to see***
***Feel free to drop in and share your results and I will update the chart as necessary (i.e., GTX 260 216, 275, 280, 285, GTX470/480/580 and HD4850, 4870, 4890, 5830, 5850, 5870, 6950, 6970, 6990, etc.)

Current GPU statistics ~ Time to Complete 1 WU ~ 213 credits in size: (Updated March 30, 2011):

1. 6950 2GB (1536 SPs @ 880mhz GPU) = 1m:11s (current leader)
2. 5870 1GB (stock) = 1m:26s - 1m:30s
3. 6950 2GB (stock) = 1m:27s
4. 5850 1GB (stock) = 1m:49s - 2m:01s
5. 4890 1GB (960mhz GPU) = 2m:26s
6. 4890 1GB = 2m:39s
7. 4870 512mb = 3m:11s
8. 4850 (700mhz GPU) = 3m:24s
9. 4850 = 3m:44 - 3m:46s
10. 4770 = 4m:16s
11. 4830 (700mhz GPU) = 4m:35s
12. 4830 (600mhz GPU) = 5m:11s
13. 4830 stock = 5m:34s
14: GTX480 = 5m:45s - 5m:59s
15. GTX295 = 6.5-7 min* extrapolated for a single unit (i.e., 13-14 minutes to complete 2 WUs)

Note: For the times to be included in the chart, please specify that your time is for the workunit which corresponds to ~213 credits (i.e., 213.75, 213.76, etc.).



------------------

It would be interesting to compare ATI vs. NV in this project since from what I've read NV stomps the wazzoo out of ATI in Folding @ Home.

Title changed as per request.
AnandTech Moderator Fullmetal Chocobo

 

[filibusterman]

I am currently getting ~108k a day on my 5850

 

[filibusterman]

I am getting about 2.00min WU completion with my XFX 5850 XXX FWIT.

 

[RussianSensation]

Hey,

Thanks for the reply. I would have imagined that a 5850 would be closer to 1 min 30 - 40 seconds. You are going to pass me anyday now

 

[biodoc]

My 4870's finish a WU in about 3 min 40 sec.
My GTX 295 finish WUs in about 13-14 min. (2 cores).

The ATI cards seem to have a significant advantage on MilkyWay.

 

[RussianSensation]

My GTX 295 finish WUs in about 13-14 min. (2 cores).

The ATI cards seem to have a significant advantage on MilkyWay.

13-14 min: (1) Is it running 1 workunit on 2 cores simultaneously? (2) or is each GPU running its own separate WU? In other words are you getting 2 WUs completed in 13-14 minutes (~213x2 = 426 credits)?

Ya, it seems that he algorithm / programming for Milkyway@Home is a completely different one from Folding@Home where ATI falls flat on its face.

This just goes to show then that the program itself is far more important towards GPGPU performance than the architecture itself! Shocking isn't it :hmm:

Thanks for the info on your part. I will update the original Post.

 

[ElFenix]

Ya, it seems that he algorithm / programming for Milkyway@Home is a completely different one from Folding@Home where ATI falls flat on its face.

yup

You may remember that NV included 30 double units in GT200 along with the 240 single precision shaders. Well, ATIs RV770 has 160 5-way VLIW units and all of them can run 1 or 2 doubles each clock. That's such a massive advantage, it just plain wouldn't make sense to use NV cards here.

http://www.gpugrid.net/forum_thread.php?id=705

so

This just goes to show then that the program itself is far more important towards GPGPU performance than the architecture itself! Shocking isn't it :hmm:

not really. i'd say they're equally important. if what you need to do requires double precision, then the ati architecture is better than the gt200 series. i think the new nvidia hardware is all double precision.

 

[mk]

HD 4770 is not on your list but it's finishing the current batch of WU's in about 4min 16s.

 

[dajeepster]

my 5870 does a wu in @ 1m 30s whereas my 5850 does a wu in @ 2m on the same system

I'm running catalyst 10.3 and win7 64bit on an i7 940 @4Ghz

 

[biodoc]

13-14 min: (1) Is it running 1 workunit on 2 cores simultaneously? (2) or is each GPU running its own separate WU? In other words are you getting 2 WUs completed in 13-14 minutes (~213x2 = 426 credits)?

(2) is correct for my GTX295.

 

[RussianSensation]

HD 4770 is not on your list but it's finishing the current batch of WU's in about 4min 16s.

Thanks for the time ! Happy crunchin'

 

[zzuupp]

one more for you:
4850 approximately 3m 45s

 

[Assimilator1]

Maybe I should look at running this on my HD 4830 (not on your list btw).

How much CPU power does it take to run the GPU client? (as in % of 1 core or % of a quad).

 

[mk]

How much CPU power does it take to run the GPU client? (as in % of 1 core or % of a quad).

Very little. BOINC Manager reports the GPU task status as "Running (0.05 CPUs + 1.00 ATI GPUs)" whereas according to Windows Task Manager a 17-19min Collatz WU needs a couple of seconds of CPU time.

There's practically no reduction to your CPU crunching power.

 

[Assimilator1]

Cool , what's a Collatz WU though? Is that still MW@H?

 

[mk]

Cool , what's a Collatz WU though? Is that still MW@H?

Collatz Conjecture is a different project. Milkyway is down atm so I couldn't check the CPU usage.

Milkyway seems to have more frequent server issues and due to the short queue of short WUs when the servers are down you run out of work very quickly. Collatz WU queue lasts longer and rare server problems are usually fixed in a few hours. Collatz also runs a little cooler although the credits are lower as well (on a 4770 it's something like 70k vs. 55k a day).

 

[Assimilator1]

Ok thx for the info , but maths DC projects aren't my thing.

 

[waffleironhead]

What are the requirements for the GPU for milkyway@home? Could I try on my onboard 780g?

(I would have to remove my x1950pro from the mobo, but would be willing to try just for fun)

 

[RussianSensation]

What are the requirements for the GPU for milkyway@home?

The GPU must be able to perform double-precision Floating Point operations.

For example, 5450, 5550, 5570, 5670, 5750, 5770 are not double-precision. GPUs such as 4850, 4870, 4890, 5850, 5870, GTX260, GTX275, etc. are.

I doubt your 780G is DP capable

 

[Assimilator1]

In that case my 4830 can too

So how much CPU power does it take to run M@H on a GPU?

 

[mk]

In that case my 4830 can too

So how much CPU power does it take to run M@H on a GPU?

Practically nothing. According to Task Manager ~1s CPU time/WU.

 

[Assimilator1]

Thanks mk
I've only just noticed that you'd already answered me :$ lol, I thought that answer had been for collatz only.

Btw on another subject, I saw you stopped running DIMES some months back (last year?). Plan to run it again? (just me & Wiz left from TA now, lol).

 

[mk]

Btw on another subject, I saw you stopped running DIMES some months back (last year?). Plan to run it again? (just me & Wiz left from TA now, lol).

Ah yes, DIMES. Sorry about that guys. I decided to drop it last year when doing fresh OS installs. All my ASNodes and ASEdges were found well before 1M measurement mark then nothing in the next 10+M. Seemed like it hadn't been doing anything useful for quite a while so it felt like a good time to call it quits.

 

[Assimilator1]

OT/ I've found the same for ASs, not found anything in ages. I posted at the DIMES forum asking if their was any point me carrying on, they said yes because the smaller/finer measurements still help to build up the picture. So I'm still running it .

 

[Mr. Pedantic]

My 4870 512MB just finished its first task in 3:13. The second task took 3:12. The third onwards took 3:11.