GTX 460 superclocked

[Frizz]

I am planning to build a system with :
1x GT 240
1x GTX 460

How do I have to modify my app_info.xml to run those two cards with x32f?

The GTX 460 can run more than one task in parallel, right? Was it 3? How do I have to configure this?

[Jason G]

I am planning to build a system with :
1x GT 240
1x GTX 460

How do I have to modify my app_info.xml to run those two cards with x32f?

The GTX 460 can run more than one task in parallel, right? Was it 3? How do I have to configure this?

OK, firstly, if you are going to mix those cards (Fermi & pre fermi) there is an issue with the 3.1 Cuda runtime (nVidia bug in the CUFFT library) ... The solution is to use the Cuda 3.0 build of x32f instead.

For installation, I would use the Lunatics Installer first, BUT not start Boinc yet ... Then follow the same procedure as I gave perryjay below, but modify the app info to suti the Cuda 3.0 executable names & DLLs.

There is a complication if you plan to use a 64 bit OS.  You'll need a different, more complicated aistub.  I'll provide that if the situation arises.

Also bear in mind that no programming information has been released on the GTX 460 yet, either publicly or to registered developers under NDA, so there *may be* significant differences that will require some re-engineering of key code kernels to see what they are really capable of.

Cheers, Jason

[Frizz]

Thanks Jason,

For installation, I would use the Lunatics Installer first, BUT not start Boinc yet ... Then follow the same procedure as I gave perryjay below, but modify the app info to suti the Cuda 3.0 executable names & DLLs.

(1) As far as I understood the GF10x GPUs can run more than one WU at a time, no? How do I have to configure this in app_info.xml?

(2) I was thinking about getting the 768MB version (slightly less power consumption) - but if they run more than one WU at a time I might run out of memory when I crunch 3 at a time. No?

[Ghost0210]

Hi Frizz23,

Stop Boinc. Then in your app_info.xml there will be a segment like this

<app_version>
        .....
            <type>CUDA</type>
            <count>1</count>
        </coproc>
        .....
</app_info>

Change the <count>1</count> to either <count>0.5</count> or <count>0.33</count> if you either want 1,2 or three tasks to run on the fermi card
Restart Boinc and you should have multiple tasks running on the same card 

As far as I know the Fermi build takes about 300MB per task - so you should be able to run 2 tasks. Sure Jason will correct me if I'm wrong on this  On my 465 I get the best through put with two tasks running at a time (994MB)

Bit later: I've attached my app_info for reference (cuda 3.1 build make sure to change if you use the 3.0 build also remembered the plan class is different(fermi31) so this will need changing as well)

[Frizz]

Thanks Ghost,

As far as I know the Fermi build takes about 300MB per task - so you should be able to run 2 tasks. Sure Jason will correct me if I'm wrong on this  On my 465 I get the best through put with two tasks running at a time (994MB)

How come 2 x 300MB = 994MB? ;-)

Hmmm ... that would mean the 765MB version would be OK for me (they consume about 10 watt less and are approx. 40 EUR cheaper).

How long does an average WU take to crunch on your GTX 465?

[Ghost0210]

Could be wrong on the memory usage 
But thats what I've found on this card 1 shorty ~3min 20, 2 @ a time ~5min 50, 3 at a time just over 7 minutes 30, this is about a 20% increase over the stock fermi app

[Jason G]

I find 2 at a time to be best on my 480, stock is better with 3 at a time because it uses less cores.  Probably as things progress I'll get a single task to use the whole cards, so back to 1-2 tasks will become optimal. 

The 460 is a bit unknown at this point, because there is a big question mark on how to use 48 cores per multiprocessor effectively, given that it is 3 half warps.  We probably won't know much more about how to handle that the best until further documentation becomes available, but to my mind may require some significant adaptation of kernel geometries to get the most from.  All indications are that these are going to be extremely popular cards, so it will be done eventually, whatever changes (if any) are needed.

Jason

[Frizz]

But thats what I've found on this card 1 shorty ~3min 20, 2 @ a time ~5min 50, 3 at a time just over 7 minutes 30, ...

and:

On my 465 I get the best through put with two tasks running at a time

OK ... when I do the math 3 WUs at a time is the optimum, no?

Completion time for 3 WUs:
1 at a time: 3 x 3:20 -> 10:00
2 at a time: 1.5 x 5:50 -> 8:45
3 at a time: 1 x 7:30 -> 7:30

[Jason G]

OK ... when I do the math 3 WUs at a time is the optimum, no?

Completion time for 3 WUs:
1 at a time: 3 x 3:20 -> 10:00
2 at a time: 1.5 x 5:50 -> 8:45
3 at a time: 1 x 7:30 -> 7:30

If the tasks used for measurement were all the same angle range yes, that's fine.  Quite possibly the best 'for now', because of those big question marks over using the 460 architecture effectively.

Don't forget about cpu overhead & Bus contention though ... so always best to measure.  Things don't 'always' scale nicely when you try cram in more work than the hardware can handle.

[Frizz]

Would you say its OK if I go for the 768MB version?

[Ghost0210]

All these task times were done through Boinc so it could have been a mix of any VHAR AR ranges
I also don't think I explained the math that well either the times I gave were for each task to complete

[Jason G]

Would you say its OK if I go for the 768MB version?

Either would be 'OK'

The 1 Gig  version is 'better' though.  Enough so in the memory bandwidth to justify the price difference.  Also If you have the budget, the top eVGA model comes with a higher factory overclock than other brands, along with rear exhaust, so is also worth a price premium if you can stretch that far.

Jason

[Frizz]

I also don't think I explained the math that well either the times I gave were for each task to complete

This is what I thought. And this is why I came to this conclusion:

Completion time for 3 WUs:
1 at a time: 3 x 3:20 -> 10:00
2 at a time: 1.5 x 5:50 -> 8:45
3 at a time: 1 x 7:30 -> 7:30

[Jason G]

Average each config over 6 similar(but preferably slightly different to represent reality) tasks & it should be clearest.

That's the difference between a measured bench & a synthetic one 

[Ghost0210]

Also have to compare the time it takes to complete one task, when running one at a time, to the time it takes to complete each task when running mulitple tasks to get an accurate picture of the exact throughput.

@Jason, if you want we can move this over to the "GTX460 Superclocked" thread , to keep this one clear for your testing? [Mod: Done]


[Edit]