ABP2 CUDA applications

Ver Greeneyes
Ver Greeneyes
Joined: 26 Mar 09
Posts: 140
Credit: 9,562,235
RAC: 0

RE: I've seen, this is

Message 96354 in response to message 96352

Quote:
I've seen, this is already talked over, but no real steps to greatly enhance the GPU output, aren't planned(?).


They plan to do everything they can. But utilizing GPUs effectively isn't easy, so give them some time; things will only get better.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,219
Credit: 137,297,177
RAC: 41,301

One ought not forget that

One ought not forget that GPU's are really, really good for ....... wait for it ...... graphics!! :-) :-)

So the speed benefit elsewhere is only going to be recovered out of those other scenarios to the extent that the problem can be massively paralleled, pipelined etc .... including setting up memory and whatnot. If one thread does indeed need the result from another then it will block, no bones about it. Can one factorise the problem/code into independent non-blocking threads???

A top fuel dragster may be fast, but it steers like a cow. Not your best choice for a trip to the shops [ and note the earlier implication about integer speeds for a floating point problem ].

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

DanNeely
DanNeely
Joined: 4 Sep 05
Posts: 1,358
Credit: 2,341,140,289
RAC: 3,307,249

RE: One ought not forget

Message 96356 in response to message 96355

Quote:

One ought not forget that GPU's are really, really good for ....... wait for it ...... graphics!! :-) :-)

So the speed benefit elsewhere is only going to be recovered out of those other scenarios to the extent that the problem can be massively paralleled, pipelined etc .... including setting up memory and whatnot. If one thread does indeed need the result from another then it will block, no bones about it. Can one factorise the problem/code into independent non-blocking threads???

eventually yes. What we need are GPUs like Larabe's. Each SP is a p1 core, and while quite pokey compared to an i7 is at least theoretically capable of running a WU on it's own. Once that's the case and the GPU can run hundreds of cut down WU's at a time the logistics get somewhat simpler.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,219
Credit: 137,297,177
RAC: 41,301

RE: eventually yes. What

Message 96357 in response to message 96356

Quote:
eventually yes. What we need are GPUs like Larabe's. Each SP is a p1 core, and while quite pokey compared to an i7 is at least theoretically capable of running a WU on it's own. Once that's the case and the GPU can run hundreds of cut down WU's at a time the logistics get somewhat simpler.


This is reminiscent of purpose built digital orrerys ( for planetary orbital prediction ) which I first saw featured in Scientific American some years ago - the idea there was to gauge the divergence of predictions with level of precision ( ie. chaos ).

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

autarkis
autarkis
Joined: 2 Feb 10
Posts: 1
Credit: 19,358
RAC: 0

RE: A top fuel dragster

Quote:
A top fuel dragster may be fast, but it steers like a cow. Not your best choice for a trip to the shops

... so are you saying I shouldnt be taking my dragster to pick up groceries?!? ;)

Here is another hand in the air for an OpenCL app. I too much prefer to have crunch actual science vs. finding numbers (based on my poor understanding of math theory, finding more examples doesnt prove anything). My goal is not to get more credits, but to do more science; seems that a cross-GPU enabled app will get more science done.

My sincere thanks to all the developers out there helping us to maximize our layman contributions to science!

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,219
Credit: 137,297,177
RAC: 41,301

RE: ... so are you saying I

Message 96359 in response to message 96358

Quote:
... so are you saying I shouldnt be taking my dragster to pick up groceries?!? ;)


:-) I've dreamt for some time about using, say, an Abrams M1A2 for local routes. :-)

[ Especially shopping centre carparks. I'd be unlikely to be on the wrong end of road-rage, cutting in or inappropriate finger gestures. Never leave home without a HE round up the tube ]

Quote:
Here is another hand in the air for an OpenCL app .... My sincere thanks to all the developers out there helping us to maximize our layman contributions to science!


Hear, hear!

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

Elphidieus
Elphidieus
Joined: 20 Feb 05
Posts: 233
Credit: 15,891,302
RAC: 11,315

Just curious... Is it

Just curious...

Is it normal to have BOINC reporting under Messages tab every other minute that CUDA tasks are being restarted...?

EDIT: Never mind. I think I've found the culprit. I think BOINC's finicky about the DCF variation as CUDA tasks are downloaded.

Elphidieus
Elphidieus
Joined: 20 Feb 05
Posts: 233
Credit: 15,891,302
RAC: 11,315

I've got three of these long

I've got three of these long CUDA workunits that are having error messages written all over them:

161525558
161525522
161525518

Any ideas...?

EDIT: Previous short CUDA workunits are also having the same error messages, but they were valid.

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,041
Credit: 745,253,462
RAC: 1,244,928

Is the Spy-Hill workround

Is the Spy-Hill workround still needed?

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4,031
Credit: 218,159,243
RAC: 49,950

RE: Is the Spy-Hill

Message 96363 in response to message 96362

Quote:
Is the Spy-Hill workround still needed?


AFAIK No. The Client uses only a single Shared Memory segment to communicate with the Apps; the communication between App and Graphics should be implemented with memory mapped files.

BM

BM

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.