Advanced search

Message boards : Number crunching : Influence of the memory frequency on performance

Author Message
JM
Send message
Joined: 18 Mar 09
Posts: 5
Credit: 624,501,954
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwat
Message 28397 - Posted: 3 Feb 2013 | 20:47:12 UTC

Hi,
I have a GTX 980m and I was wondering if the memory frequency was affecting the crunching performance. I have noticed that a downclock from 1800MHz to 1300MHz is able to lower the GPU temperature by more than 5 degres. But do you think that it changes anything on the overall performance keeping the GPU core clock to the same frequency? I don't see any real difference from the last few results sended but maybe you know more about it technicaly speaking.
Thanks for your answers.

Profile MJH
Project administrator
Project developer
Project scientist
Send message
Joined: 12 Nov 07
Posts: 696
Credit: 27,266,655
RAC: 0
Level
Val
Scientific publications
watwat
Message 28400 - Posted: 3 Feb 2013 | 22:08:15 UTC - in response to Message 28397.

Derating the memory clock is fine. Done within reason, it should affect performance significantly.

MJH

JM
Send message
Joined: 18 Mar 09
Posts: 5
Credit: 624,501,954
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwat
Message 28403 - Posted: 3 Feb 2013 | 23:03:49 UTC - in response to Message 28400.
Last modified: 3 Feb 2013 | 23:04:15 UTC

Ok thanks.
And my mistake it is not a GTX 980m but a GTX 680m.
ok so could I monitor a possible loss of performance with the GPU usage?
By derating the memory, the GPU core usage (in %) actualy gain 1-2 points compared with the normal frequency. Any explanation?

werdwerdus
Send message
Joined: 15 Apr 10
Posts: 123
Credit: 1,004,473,861
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28406 - Posted: 4 Feb 2013 | 6:26:57 UTC - in response to Message 28403.

Ok thanks.
And my mistake it is not a GTX 980m but a GTX 680m.
ok so could I monitor a possible loss of performance with the GPU usage?
By derating the memory, the GPU core usage (in %) actualy gain 1-2 points compared with the normal frequency. Any explanation?

could be normal variation between 2 different work units
____________
XtremeSystems.org - #1 Team in GPUGrid

JM
Send message
Joined: 18 Mar 09
Posts: 5
Credit: 624,501,954
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwat
Message 28407 - Posted: 4 Feb 2013 | 9:04:59 UTC - in response to Message 28406.

Yes but in this case same working unit, only the graphic memory clock change..

Dylan
Send message
Joined: 16 Jul 12
Posts: 98
Credit: 386,043,752
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwat
Message 28418 - Posted: 5 Feb 2013 | 23:25:03 UTC

Done within reason, it should affect performance significantly.




I am a bit confused by this, is this statement saying memory clock adjustments won't affect performance, or that they will? In other words, was the word "should" here supposed to be "shouldn't"?

Profile MJH
Project administrator
Project developer
Project scientist
Send message
Joined: 12 Nov 07
Posts: 696
Credit: 27,266,655
RAC: 0
Level
Val
Scientific publications
watwat
Message 28420 - Posted: 5 Feb 2013 | 23:50:35 UTC - in response to Message 28418.

Correct: "shouldn't"

MJH

Dylan
Send message
Joined: 16 Jul 12
Posts: 98
Credit: 386,043,752
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwat
Message 28427 - Posted: 6 Feb 2013 | 20:41:04 UTC

Thanks. I'll try adjusting my memory clock on my 670 once the beta tasks are over and I'll post the results I get.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1112
Credit: 6,162,416,256
RAC: 0
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28525 - Posted: 15 Feb 2013 | 3:12:30 UTC

Tried an experiment with my 650 TI, underclocked the memory 300MHz and then increased a little at a time until it was OCed by 300MHz. With every memory clock increase there was a corresponding decrease in completion time. All WUs were the Noelia long type. Total time difference was over 2 hours from lowest to highest memory speed.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28528 - Posted: 15 Feb 2013 | 14:07:55 UTC - in response to Message 28525.

I agree with Beyond's observations.

A slight adjustment won't make much difference of course. If GPU-related task crashes are an issue, reducing the GDDR5 clocks a bit is the easiest method of improving stability. Reducing the clocks from 3000MHz by 100MHz won't make a big impact on runtime; it's only 3% slower, so if it stabilizes the GPU it's well worth it.
It's probably the case that if you can find stability at a reduced Voltage, even at a slightly reduced memory clock that you will be more efficient in terms of the GPU's performance per Watt. However if you include the rest of the systems overheads you're really shooting yourself in the foot. You would get more value from undervolting a CPU.

It's also going to be the case that some cards will yield more from a GDDR OC due to the different bus widths, and thus GPU memory bandwidth. There is going to be some core/shader count/rates vs bandwidth factor, but with varying task requirements it's probably better to be cautious at this time, especially if you are running the test app.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1112
Credit: 6,162,416,256
RAC: 0
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28537 - Posted: 15 Feb 2013 | 20:46:26 UTC

The particular 650 TI card I bought (MSI Power Edition) has huge memory headroom, the chips are rated much higher than they are clocked by default. SO I feel pretty confident in the memory OC on this particular card. So far no mishaps. As always YMMV.

Post to thread

Message boards : Number crunching : Influence of the memory frequency on performance

//