Welcome to MilkyWay@home

HD5870


Advanced search

Message boards : Number crunching : HD5870
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 7 · Next

AuthorMessage
Hans-Ulrich Hugi

Send message
Joined: 14 Sep 09
Posts: 7
Credit: 163,221,288
RAC: 0
100 million credit badge12 year member badgeextraordinary contributions badge
Message 31959 - Posted: 4 Oct 2009, 16:59:50 UTC

Today i did some testing with the 5870 at different clock speeds (Milkyway and Collatz). All results are under Win7 Ultimate 64bit (Build 7600) with a Q9650 @ 3.75 GHz, Boinc Client 6.10.11 and the 0.20 Milkyway SW (no command line parameters changed). Under Collatz higher clocks greatly improve the performance (see my post there). Here are the results for Milkyway:

Clocks 850 / 1200 (@stock)
===========================

GPU load max. 99% (GPU-Z 0.3.5)

GPU core clock: 850 MHz, memory clock: 300 MHz (Wrong: 1200 MHz!)

predicted runtime per iteration is 62 ms(33.3333 ms are allowed), dividing each iteration in 2 parts
borders of the domains at 0 800 1600
Calculated about 8.22242e+012 floatingpoint ops on GPU, 1.23583e+008 on FPU. Approximate GPU time 20.7792 seconds

CPU time: 1.65361 seconds, GPU time: 20.7792 seconds, wall clock time: 21.793 seconds, CPU frequency: 3.7371 GHz
CPU time: 1.70041 seconds, GPU time: 20.7948 seconds, wall clock time: 21.833 seconds, CPU frequency: 3.7371 GHz
CPU time: 1.71601 seconds, GPU time: 20.7792 seconds, wall clock time: 21.844 seconds, CPU frequency: 3.7371 GHz

Clocks 850 / 1320 (GPU @stock / Memory + 10%)
=============================================

GPU load max. 99% (GPU-Z 0.3.5)

GPU core clock: 850 MHz, memory clock: 300 MHz (Wrong: 1320 MHz!)

predicted runtime per iteration is 62 ms(33.3333 ms are allowed), dividing each iteration in 2 parts
borders of the domains at 0 800 1600
Calculated about 8.22242e+012 floatingpoint ops on GPU, 1.23583e+008 on FPU. Approximate GPU time 20.7636 seconds

CPU time: 1.66921 seconds, GPU time: 20.7636 seconds, wall clock time: 21.831 seconds, CPU frequency: 3.7371 GHz
CPU time: 1.70041 seconds, GPU time: 20.8104 seconds, wall clock time: 21.907 seconds, CPU frequency: 3.7371 GHz
CPU time: 1.70041 seconds, GPU time: 20.8104 seconds, wall clock time: 21.847 seconds, CPU frequency: 3.7371 GHz

Clocks 935 / 1200 (GPU + 10% / Memory @stock)
==============================================

GPU core clock: 935 MHz, memory clock: 300 MHz (Wrong: 1200 MHz!)

predicted runtime per iteration is 56 ms (33.3333 ms are allowed), dividing each iteration in 2 parts
borders of the domains at 0 800 1600
Calculated about 8.22242e+012 floatingpoint ops on GPU, 1.23583e+008 on FPU. Approximate GPU time 20.1084 seconds

CPU time: 1.01401 seconds, GPU time: 20.1084 seconds, wall clock time: 21.15 seconds, CPU frequency: 3.7371 GHz
CPU time: 1.01401 seconds, GPU time: 20.1084 seconds, wall clock time: 21.15 seconds, CPU frequency: 3.7371 GHz
CPU time: 1.01401 seconds, GPU time: 20.0928 seconds, wall clock time: 21.125 seconds, CPU frequency: 3.7371 GHz

Clocks 935 / 1320 (both + 10%)
===============================

GPU load max. 99% (GPU-Z 0.3.5)

GPU core clock: 935 MHz, memory clock: 300 MHz (Wrong: 1320MHz!)

predicted runtime per iteration is 56 ms(33.3333 ms are allowed), dividing each iteration in 2 parts
borders of the domains at 0 800 1600
Calculated about 8.22242e+012 floatingpoint ops on GPU, 1.23583e+008 on FPU. Approximate GPU time 20.0772 seconds

CPU time: 1.04521 seconds, GPU time: 20.0772 seconds, wall clock time: 21.144 seconds, CPU frequency: 3.7371 GHz
CPU time: 0.998406 seconds, GPU time: 20.0928 seconds, wall clock time: 21.122 seconds, CPU frequency: 3.7371 GHz
CPU time: 0.998406 seconds, GPU time: 20.0928 seconds, wall clock time: 21.122 seconds, CPU frequency: 3.7371 GHz

I did hope to see a similar improvement with Milkyway as i got it with Collatz. But the result is only a lower CPU time a minimal better GPU time.
ID: 31959 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge13 year member badgeextraordinary contributions badge
Message 31963 - Posted: 4 Oct 2009, 19:11:47 UTC - in response to Message 31959.  

I did hope to see a similar improvement with Milkyway as i got it with Collatz. But the result is only a lower CPU time a minimal better GPU time.

You are complaining about crunching a WU (53 credits) in just about 20 seconds? ;)

As I've said several times already, the memory performance is not that important for MW. I guess you see virtually the same times when you clock it at 800MHz or something about that.
ID: 31963 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ExtraTerrestrial Apes
Avatar

Send message
Joined: 1 Sep 08
Posts: 204
Credit: 219,354,537
RAC: 0
200 million credit badge13 year member badge
Message 31965 - Posted: 4 Oct 2009, 19:17:23 UTC - in response to Message 31959.  

Thanks for the effort!

What comes to mind here is that 1200 MHz is the real stock memory clock, whereas 300 MHz is the power saving clock. Could it be that due to some reason your memory was not clocking up properly? You core was certainly not in power saving mode, otherwise it would have been much slower.
If the memory ran at real 300 MHz you may be in the memory-bandwidth limited regime with your 1600 shaders and therefore the core overclock wouldn't help much. Would be quite a coincidence, but maybe possible. On my 4870 MW scales perfectly with core clock.

MrS
Scanning for our furry friends since Jan 2002
ID: 31965 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Hans-Ulrich Hugi

Send message
Joined: 14 Sep 09
Posts: 7
Credit: 163,221,288
RAC: 0
100 million credit badge12 year member badgeextraordinary contributions badge
Message 31971 - Posted: 4 Oct 2009, 23:05:39 UTC

@ExtraTerrestrial Apes
The memory clocks are misread. Probably you noticed that it always reads 300 MHz, if the card runs at 1200 or 1320 MHz. But GPU-Z shows the correct speed. And the 5870 correctly reduces the memory/gpu clocks after about 5 seconds on idle.
ID: 31971 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Donnie
Avatar

Send message
Joined: 19 Jul 08
Posts: 67
Credit: 272,086,462
RAC: 0
200 million credit badge13 year member badge
Message 31973 - Posted: 5 Oct 2009, 1:40:29 UTC - in response to Message 31971.  
Last modified: 5 Oct 2009, 1:58:38 UTC

I guess we can expect a 50% reduction in credit because these new cards are extremely fast! You heard the writting on the wall here first!! DA will never stand for this type of technological advance without penalizing everyone else!!

[edit] even if the server could handle it?[/edit]

ID: 31973 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profilekashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
100 million credit badge13 year member badge
Message 31975 - Posted: 5 Oct 2009, 2:26:42 UTC

At default memory speed you had about 3.27% faster runtime with a 10% increase in clockspeed. I have noticed that although Milkyway performance scales reasonably well between different cards according to the number of shaders multiplied by the clockspeed, it doesn't always scale exactly with clockspeed on the same card. The change in runtime due to change in clockspeed can sometimes be partly in steps according to how the task is subdivided and this can also be affected by how the kernel frequency f is set. These performance steps are sometimes at different speeds or f values for different length tasks when they are available and can also be different for different model cards. The load on the CPU can also affect the GPU processing performance.

Because the priority is not correctly set until Collatz v2.05 is released you have already seen a very exaggerated example of this step behaviour at Collatz. Above a certain overclocked speed the kernel stoppped being subdivided allowing the HD 5870 to reach full load which unlocked its true performance.

Your HD 5870 @ 935 MHz completes a MilkyWay task in almost exactly half the time of my HD 4890 @ 950 MHz, so HD 5870 MilkyWay performance is basically scaling according to the formula and is excellent. My card can complete a task a few seconds faster if I add some parameters or leave an extra CPU core free, but that makes it run too hot when processing 2 tasks concurrently so I just use the default settings now.

It is possible that the new fast memory link retraining feature is the cause of the memory always being reported as 300 MHz.

When applications are new or I have new hardware I like to try and extract the maximum efficiency, but after I have played with the settings for a while I end up relaxing things a bit and do not concern myself with extracting the last few seconds of performance.

Thank you for sharing your MilkyWay results with your new HD 5870.
ID: 31975 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profileverstapp
Avatar

Send message
Joined: 26 Jan 09
Posts: 589
Credit: 497,834,261
RAC: 0
300 million credit badge12 year member badge
Message 31989 - Posted: 5 Oct 2009, 20:19:20 UTC

Possibly of interest to the more asymmetric among you.
http://www.theinquirer.net/inquirer/news/1557395/nvidia-bastille-stormed-hackers
Cheers,

PeterV

.
ID: 31989 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ExtraTerrestrial Apes
Avatar

Send message
Joined: 1 Sep 08
Posts: 204
Credit: 219,354,537
RAC: 0
200 million credit badge13 year member badge
Message 31990 - Posted: 5 Oct 2009, 20:42:14 UTC

If the memory speed is correct - what else causes this bad scaling? 21.8s @ 850 MHz should result in 18.9s @ 935 MHz. Are we seeing some CPU limitation here? Did you leave at least one physical core free while doing these tests?

MrS
Scanning for our furry friends since Jan 2002
ID: 31990 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileThe Gas Giant
Avatar

Send message
Joined: 24 Dec 07
Posts: 1947
Credit: 240,884,648
RAC: 0
200 million credit badge13 year member badge
Message 31999 - Posted: 6 Oct 2009, 0:43:59 UTC

Holy cow, that 5870 is quick! Almost makes me want to go out and purchase one......almost. Be nice to see the difference between the 5850 and the 5870.
ID: 31999 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profileverstapp
Avatar

Send message
Joined: 26 Jan 09
Posts: 589
Credit: 497,834,261
RAC: 0
300 million credit badge12 year member badge
Message 32000 - Posted: 6 Oct 2009, 1:32:02 UTC

5870 on order, in conjunction with a 6600->i7 upgrade.
My builder says there is a lack of stock...
Cheers,

PeterV

.
ID: 32000 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Divide Overflow
Avatar

Send message
Joined: 16 Feb 09
Posts: 109
Credit: 11,089,510
RAC: 0
10 million credit badge12 year member badge
Message 32004 - Posted: 6 Oct 2009, 2:38:59 UTC

If you're going to crunch 24x7, I'd suggest waiting 'till some custom cooling solutions are on the market. The improved cooling efficiency can give you longer life span and much lower fan noise at load.
ID: 32004 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ExtraTerrestrial Apes
Avatar

Send message
Joined: 1 Sep 08
Posts: 204
Credit: 219,354,537
RAC: 0
200 million credit badge13 year member badge
Message 32040 - Posted: 6 Oct 2009, 19:50:17 UTC
Last modified: 6 Oct 2009, 19:51:08 UTC

Regarding cooling: I'm sure I couldn't use one with stock cooling.. but all my current GPUs have Accelero S1s with 2 x 120 mm fans each. I read in forums that the mounting holes are compatible to the old ones, so one would *just* need to take special care of the power circuitry. For example use the back plate of the stock cooler to cover them - if this is possible, as with the 4870.

Regarding 5850 speed: 5870@850 MHz (stock) should do about 228.2k credits/day, whereas a 5850@725 MHz should do 175.2k credits/day (link).

MrS
Scanning for our furry friends since Jan 2002
ID: 32040 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Liranan

Send message
Joined: 20 Aug 09
Posts: 8
Credit: 41,276,077
RAC: 0
30 million credit badge12 year member badge
Message 32077 - Posted: 7 Oct 2009, 13:36:36 UTC

Right now I have two 4870's running Collatz and I'm getting about 70,000 points a day. I need a new video card for the third PC and I'm waiting for the 5870's to drop in price and I'll get one for myself.

I'm still waiting for someone at MW to create a programme that takes the changes that AMD made to the drivers eight months ago into account. Until then I will run Collatz on my video cards and MW on my 955BE @ 3.7GHz and C2D E8400.
ID: 32077 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge13 year member badgeextraordinary contributions badge
Message 32082 - Posted: 7 Oct 2009, 14:54:25 UTC - in response to Message 32077.  

I'm still waiting for someone at MW to create a programme that takes the changes that AMD made to the drivers eight months ago into account. Until then I will run Collatz on my video cards and MW on my 955BE @ 3.7GHz and C2D E8400.

Noone of the MW staff will do it. After all they have not that much done on the ATI app so far ;).

But to say the truth, I already built versions working with newer drivers without the need to copy the dll files. The result is that I have now every app twice, one version for old drivers and one for Catalyst 9.2 and up. As the BOINC clients 6.10.8 and up are able to submit all necessary information, the best thing would be if the MW project server would be updated to distribute the right version automatically to the clients as Collatz is doing. Installing an application with an app_info.xml isn't less work than to copy those dlls. It doesn't make much sense otherwise.

I've made them available to the project some days ago, so maybe we are just a server software update away from automatic downloads of the ATI app to capable hosts.
ID: 32082 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfilePaul D. Buck

Send message
Joined: 12 Apr 08
Posts: 621
Credit: 161,934,067
RAC: 0
100 million credit badge13 year member badge
Message 32087 - Posted: 7 Oct 2009, 17:05:38 UTC - in response to Message 32082.  

I've made them available to the project some days ago, so maybe we are just a server software update away from automatic downloads of the ATI app to capable hosts.

Way cool!

Some times when my health is bad I cannot monitor my machines and make sure that they are running the correct versions and so automatic updates is way cooler.

You do good work ... :)

Can we vote on which project we want you to fix next? :)
ID: 32087 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ExtraTerrestrial Apes
Avatar

Send message
Joined: 1 Sep 08
Posts: 204
Credit: 219,354,537
RAC: 0
200 million credit badge13 year member badge
Message 32095 - Posted: 7 Oct 2009, 20:45:33 UTC

Nabu, if it's not clear by now: Cluster Physik is the guy who wrote and optimized this neat ATI app and he's not part of the MW team ;)

MrS
Scanning for our furry friends since Jan 2002
ID: 32095 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileZanth
Avatar

Send message
Joined: 18 Feb 09
Posts: 158
Credit: 89,775,020
RAC: 8,109
50 million credit badge12 year member badge
Message 32113 - Posted: 8 Oct 2009, 5:01:26 UTC - in response to Message 32082.  


But to say the truth, I already built versions working with newer drivers without the need to copy the dll files. The result is that I have now every app twice, one version for old drivers and one for Catalyst 9.2 and up. As the BOINC clients 6.10.8 and up are able to submit all necessary information, the best thing would be if the MW project server would be updated to distribute the right version automatically to the clients as Collatz is doing. Installing an application with an app_info.xml isn't less work than to copy those dlls. It doesn't make much sense otherwise.

I've made them available to the project some days ago, so maybe we are just a server software update away from automatic downloads of the ATI app to capable hosts.



That's great news! Looking very much forward to it! :)
ID: 32113 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ExtraTerrestrial Apes
Avatar

Send message
Joined: 1 Sep 08
Posts: 204
Credit: 219,354,537
RAC: 0
200 million credit badge13 year member badge
Message 32125 - Posted: 8 Oct 2009, 7:50:10 UTC

Here it is. Don't forget to take a look at the readme ;)

MrS
Scanning for our furry friends since Jan 2002
ID: 32125 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileBeyond
Avatar

Send message
Joined: 15 Jul 08
Posts: 383
Credit: 503,272,427
RAC: 28,163
500 million credit badge13 year member badge
Message 32136 - Posted: 8 Oct 2009, 16:50:15 UTC - in response to Message 32125.  

Here it is. Don't forget to take a look at the readme ;)

MrS

Thanks ET! v0.20b dropped 3 seconds\WU on both my machines.
ID: 32136 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ExtraTerrestrial Apes
Avatar

Send message
Joined: 1 Sep 08
Posts: 204
Credit: 219,354,537
RAC: 0
200 million credit badge13 year member badge
Message 32146 - Posted: 8 Oct 2009, 19:34:27 UTC

I'm seeing similar crunching times. Could be that GPU utilization was improved in your case, whereas mine already was at 99%.

MrS
Scanning for our furry friends since Jan 2002
ID: 32146 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 7 · Next

Message boards : Number crunching : HD5870

©2021 Astroinformatics Group