Welcome to MilkyWay@home

CUDA for Milkyway@Home

Message boards : Application Code Discussion : CUDA for Milkyway@Home
Message board moderation

To post messages, you must log in.

1 · 2 · 3 · 4 . . . 6 · Next

AuthorMessage
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25013 - Posted: 11 Jun 2009, 18:17:36 UTC
Last modified: 11 Jun 2009, 18:18:29 UTC

Because it looks like the server can handle a bit more work, I've uploaded the CUDA code for plain milkyway@home, and it's in the code release directory. Hopefully we'll be able to just use CUDA code for regular milkyway@home and then we wouldn't even need to start up milkywayGPU@home.

I've also added two binaries, one single and one double precision for intel-based macs running OS X. These are also in the code release directory. If anyone wants to try and make some linux and windows binaries for people to use that would be appreciated.
ID: 25013 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Emanuel

Send message
Joined: 18 Nov 07
Posts: 280
Credit: 2,442,757
RAC: 0
Message 25026 - Posted: 11 Jun 2009, 18:43:04 UTC - in response to Message 25013.  

How will that work with the single precision cards? Won't the searches eventually reach a point where they won't be accurate enough anymore?
ID: 25026 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ztmike

Send message
Joined: 4 Jun 09
Posts: 45
Credit: 447,355
RAC: 0
Message 25029 - Posted: 11 Jun 2009, 18:47:55 UTC

I noticed a problem in the binaries and code I put on the code download page. I removed them to do some fixes, but they should be available tonight or at the latest tomorrow morning.


^^ From the front page..

Does this mean when you put it back up, that we Cuda users can download the app and try crunching on that? The download link you posted http://milkyway.cs.rpi.edu/milkyway/download/code_release/

Is a little confusing ..to say the least.
ID: 25029 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Brian Silvers

Send message
Joined: 21 Aug 08
Posts: 625
Credit: 558,425
RAC: 0
Message 25030 - Posted: 11 Jun 2009, 18:48:16 UTC - in response to Message 25013.  

Hopefully we'll be able to just use CUDA code for regular milkyway@home and then we wouldn't even need to start up milkywayGPU@home.


I'm very confused by this...

I thought you were wanting to issue much larger tasks to GPUs so that you could do things that CPUs wouldn't be able to handle quickly enough? What happened with that?

Also, unless you're seeing a relatively large stockpile of tasks waiting to be sent, and I don't think < 1000 qualifies as "large", then I think that there is a high probability that people will figure out that things are working better here now, bring more processing power over, and then we'll be back in the same situation...

Make sure that you're planning for a longer time horizon than just a few weeks....
ID: 25030 · 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
Message 25033 - Posted: 11 Jun 2009, 18:49:42 UTC - in response to Message 25013.  
Last modified: 11 Jun 2009, 18:52:37 UTC

Because it looks like the server can handle a bit more work, I've uploaded the CUDA code for plain milkyway@home, and it's in the code release directory. Hopefully we'll be able to just use CUDA code for regular milkyway@home and then we wouldn't even need to start up milkywayGPU@home.

I've also added two binaries, one single and one double precision for intel-based macs running OS X. These are also in the code release directory. If anyone wants to try and make some linux and windows binaries for people to use that would be appreciated.

I would be ready. I can release a single precision ATI version today if you want. But be warned, SP is about a factor 4 or more faster than DP per WU! The server should be prepared for that combined rush (a lot of CUDA GPUs + lower end ATI cards adding to the higher end ATIs already in use).

Have you thought about the credit issue? Just giving the single precision clients half that of the DP ones isn't going to cut it. The single precision variant actually does quite a bit less operations than the double precision one because all used functions have hardware implementations. For instance one doesn't have to do the quite expensive exp() in software (like in DP or also on CPUs [done in the math library]). All in all the number of really executed operations for the SP version (have to thoroughly count it, will tell you the result) is almost exactly halved.

I think it would be best to use exactly the same credit multiplier as SETI which also GPUGrid converges to (2.72 credits per SP TeraFlop, here at MW the current value for DP is 7.5, one may think about lowering it to 5.44, i.e. twice the SP value). Together with the lower count of the operations for SP it will most likely result in about the same credit/time rate as when using DP on the same GPU (at least on ATI, for nvidias GTX2xx cards SP would yield most likely a higher credit rate than DP). But how do you differentiate between the SP and DP versions? By using the applications version string? Or should one introduce an additional SP marker to the out file?
ID: 25033 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25034 - Posted: 11 Jun 2009, 18:50:07 UTC - in response to Message 25029.  

I noticed a problem in the binaries and code I put on the code download page. I removed them to do some fixes, but they should be available tonight or at the latest tomorrow morning.


^^ From the front page..

Does this mean when you put it back up, that we Cuda users can download the app and try crunching on that? The download link you posted http://milkyway.cs.rpi.edu/milkyway/download/code_release/

Is a little confusing ..to say the least.


Yeah, I'll have a binary for OS X that you can try and crunch with.
ID: 25034 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25035 - Posted: 11 Jun 2009, 18:51:41 UTC - in response to Message 25033.  
Last modified: 11 Jun 2009, 18:55:34 UTC


I would be ready. I can release a single precision ATI version today if you want. But be warned, SP is about a factor 4 or more faster than DP per WU! The server should be prepared for that combined rush (a lot of CUDA GPUs + lower end ATI cards adding to the higher end ATIs already in use).


Talking to my advisors it seems they don't want us to release a single precision application for GPUs for milkyway@home just yet. I think that will have to wait until we get milkyway GPU up and running. The single precision binaries I'm going to release are just for testing really (i'd like to see how many machines my code works on).


Have you thought about the credit issue? Just giving the single precision clients half that of the DP ones isn't going to cut it. The single precision variant actually does quite a bit less operations than the double precision one because all used functions have hardware implementations. For instance one doesn't have to do the quite expensive exp() in software (like in DP or also on CPUs [done in the math library]). That almost exactly halves the amount of really executed operations for the SP version (have to thoroughly count it, will tell you the result).


Yeah we're going to have to come up with a credit scheme for milkywaygpu that takes into account floating point and double precision apps (double precision getting more credit because their results are more accurate and useful.


I think it would be best to use exactly the same credit multiplier as SETI which also GPUGrid converges to (2.72 credits per SP TeraFlop, here at MW the current value for DP is 7.5, one may think about lowering it to 5.44, i.e. twice the SP value). Together with the lower count of the operations for SP it will most likely result in about the same credit/time rate as when using DP on the same GPU (at least on ATI, for nvidias GTX2xx cards SP would yield most likely a higher credit rate than DP). But how do you differentiate between the SP and DP versions? By using the applications version string? Or should one introduce an additional SP marker to the out file?

ID: 25035 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ztmike

Send message
Joined: 4 Jun 09
Posts: 45
Credit: 447,355
RAC: 0
Message 25036 - Posted: 11 Jun 2009, 18:55:25 UTC - in response to Message 25034.  

I noticed a problem in the binaries and code I put on the code download page. I removed them to do some fixes, but they should be available tonight or at the latest tomorrow morning.


^^ From the front page..

Does this mean when you put it back up, that we Cuda users can download the app and try crunching on that? The download link you posted http://milkyway.cs.rpi.edu/milkyway/download/code_release/

Is a little confusing ..to say the least.


Yeah, I'll have a binary for OS X that you can try and crunch with.


So no support for Windows yet? All my machines run Windows.. :(
ID: 25036 · 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
Message 25037 - Posted: 11 Jun 2009, 18:55:26 UTC - in response to Message 25035.  

Talking to my advisors it seems they don't want us to release a single precision application for GPUs just yet.

I'm probably not going to release one for CUDA either (i've just been using single precision because thats what my GPU supports). We've ordered some double precision nvidia GPUs to play around with so once we get those it'll probably be double precision from here on out.

So no chance for any GPU older than the GTX2xx line of nvidia or the mainstream series of ATI (besides the HD4700 series which supports DP)?
ID: 25037 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25038 - Posted: 11 Jun 2009, 18:56:19 UTC - in response to Message 25037.  

Talking to my advisors it seems they don't want us to release a single precision application for GPUs just yet.

I'm probably not going to release one for CUDA either (i've just been using single precision because thats what my GPU supports). We've ordered some double precision nvidia GPUs to play around with so once we get those it'll probably be double precision from here on out.

So no chance for any GPU older than the GTX2xx line of nvidia or the mainstream series of ATI (besides the HD4700 series which supports DP)?



I edited my post :) We're going to have floating point on milkywayGPU whenever I get the backend up and running, but not here for the time being.
ID: 25038 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25039 - Posted: 11 Jun 2009, 18:56:20 UTC - in response to Message 25037.  

Talking to my advisors it seems they don't want us to release a single precision application for GPUs just yet.

I'm probably not going to release one for CUDA either (i've just been using single precision because thats what my GPU supports). We've ordered some double precision nvidia GPUs to play around with so once we get those it'll probably be double precision from here on out.

So no chance for any GPU older than the GTX2xx line of nvidia or the mainstream series of ATI (besides the HD4700 series which supports DP)?



I edited my post :) We're going to have floating point on milkywayGPU whenever I get the backend up and running, but not here for the time being.
ID: 25039 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25042 - Posted: 11 Jun 2009, 18:57:36 UTC - in response to Message 25036.  

I noticed a problem in the binaries and code I put on the code download page. I removed them to do some fixes, but they should be available tonight or at the latest tomorrow morning.


^^ From the front page..

Does this mean when you put it back up, that we Cuda users can download the app and try crunching on that? The download link you posted http://milkyway.cs.rpi.edu/milkyway/download/code_release/

Is a little confusing ..to say the least.


Yeah, I'll have a binary for OS X that you can try and crunch with.


So no support for Windows yet? All my machines run Windows.. :(


I don't have a windows box to compile and test it on, so no. Another student who should be taking over most of the GPU work, Anthony, does I think. So we should have those when he gets up to speed (or when one of our users releases one).
ID: 25042 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile borandi
Avatar

Send message
Joined: 21 Feb 09
Posts: 180
Credit: 27,806,824
RAC: 0
Message 25045 - Posted: 11 Jun 2009, 19:05:33 UTC - in response to Message 25034.  

I noticed a problem in the binaries and code I put on the code download page. I removed them to do some fixes, but they should be available tonight or at the latest tomorrow morning.


^^ From the front page..

Does this mean when you put it back up, that we Cuda users can download the app and try crunching on that? The download link you posted http://milkyway.cs.rpi.edu/milkyway/download/code_release/

Is a little confusing ..to say the least.


Yeah, I'll have a binary for OS X that you can try and crunch with.


Travis I think you misunderstood ztmike here.

The CUDA code release can only be tested on the test WUs ztmike, not for the main project itself (yet).

Also Travis, if you are going to deploy the CUDA DP at normal MW, will this not require a BOINC upgrade server side? This would be so people can select 'Use CUDA app' in their preferences, akin to SETI, rather just randomly issue CUDA Wus to CUDA machines, like in Aqua@Home.

It would be interesting if MW_GPU changed its name to MW_SP :) Such that the older cards (and some of the new cards) could go over there.
ID: 25045 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Phil Rosa-Leeke

Send message
Joined: 30 Mar 09
Posts: 1
Credit: 385,419
RAC: 0
Message 25160 - Posted: 12 Jun 2009, 15:59:03 UTC - in response to Message 25045.  
Last modified: 12 Jun 2009, 16:01:51 UTC

Hi there,

Right, first things first, I'm no computer buff so all this tech talk to me is gibberish. All I want to know is if the up-coming CUDA thingy will work on my computer. What I have is the following:

Hardware Overview:

Model Name: iMac
Model Identifier: iMac7,1
Processor Name: Intel Core 2 Duo
Processor Speed: 2.4 GHz
Number Of Processors: 1
Total Number Of Cores: 2
L2 Cache: 4 MB
Memory: 1 GB
Bus Speed: 800 MHz

ATI Radeon HD 2600 Pro:

Chipset Model: ATI,RadeonHD2600
Type: Display
Bus: PCIe
PCIe Lane Width: x16
VRAM (Total): 256 MB
Vendor: ATI (0x1002)
Device ID: 0x9583
Revision ID: 0x0000
ROM Revision: 113-B2250F-212
EFI Driver Version: 01.00.212
Displays:
iMac:
Resolution: 1680 x 1050
Depth: 32-Bit Color
Core Image: Hardware Accelerated
Main Display: Yes
Mirror: Off
Online: Yes
Quartz Extreme: Supported
Built-In: Yes
Display Connector:
Status: No Display Connected

Just a simple Yes or No will do thanks. I'm running the 6.2.18 version of BOINC and it's pushing out 1280+ Avg. Work Units at the mo. If the new fandangled CUDA will work on my system, will it be up to doing more or will I be better off leaving the BOINC running and forgetting the CUDA?

All the best and keep up the good work,

Phil
PORTUGAL
ID: 25160 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 25162 - Posted: 12 Jun 2009, 16:05:47 UTC - in response to Message 25160.  
Last modified: 12 Jun 2009, 16:06:41 UTC

CUDA is for NVidia graphics cards. You have an ATI graphics card. There might already be a GPU application for you (unless it doesn't support double precision).

When OpenCL comes out, your card should be supported by that. But I don't think OpenCL comes out until Apple releases Snow Leopard.
ID: 25162 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Glenn Rogers
Avatar

Send message
Joined: 4 Jul 08
Posts: 165
Credit: 364,966
RAC: 0
Message 25163 - Posted: 12 Jun 2009, 16:11:33 UTC - in response to Message 25033.  

Yes Please Cluster, can you make it if possible so it works on X series ATI cards? It would be greatly appreciated...

Cheers
Glenn
ID: 25163 · 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
Message 25168 - Posted: 12 Jun 2009, 16:57:21 UTC - in response to Message 25163.  

Yes Please Cluster, can you make it if possible so it works on X series ATI cards? It would be greatly appreciated...

All ATI HD series will work with the MW GPU project (most likely even the integrated chipsets like 780G), where single precision will be enough. In the moment it is simply not possible, because they don't support the required precision.

But for the X series there may be no simple solution as they are not supported by the Stream SDK. One would have to write some DirectX shaders (or use the old BrookGPU) to get it to work and I don't know if they support the required features.
ID: 25168 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile arkayn
Avatar

Send message
Joined: 14 Feb 09
Posts: 999
Credit: 74,932,619
RAC: 0
Message 25190 - Posted: 12 Jun 2009, 23:26:50 UTC - in response to Message 25162.  

Apple is only going to support OpenCL on a very limited amount of GPU's.

Graphics and OpenCL

One of the big features in Snow Leopard is the ability to tap into many more computing resources at once, and part of this is the inclusion of OpenCL, which allows the system to make use of the graphics processor for everyday tasks when it's not in use by other programs. According to the Leopard technical specifications page, this ability will require a computer with at least the following graphics processors:

NVIDIA Geforce 8600M GT, GeForce 8800 GT, GeForce 8800 GTS, Geforce 9400M, GeForce 9600M GT, GeForce GT 120, GeForce GT 130.
ATI Radeon 4850, Radeon 4870

Most of these processors were released in the last year, which means that while Snow Leopard will run on any Intel system, its full speed and computing potential will only happen on the following models released so far:
MacBook Pro: Mid-2007 and later
MacBooks: Late 2008 models and later
iMacs: Only the 2009 models and later
Mac Mini: Only the 2009 models and later
Mac Pro: All models with upgrades, though only the 2009 models will support it out of the box

The following models will run the OS, but will not benefit from the enhanced power available from the GPU:
MacBook Pro: Prior to 2007
MacBook: Prior to 2008
iMacs: Prior to 2009
Mac Mini: All except the 2009 models

ID: 25190 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ztmike

Send message
Joined: 4 Jun 09
Posts: 45
Credit: 447,355
RAC: 0
Message 25213 - Posted: 13 Jun 2009, 0:39:01 UTC

Is anybody even compiling/testing M@H Cuda app for Windows? Not sure why Windows was not supported first, as most people have that OS..
ID: 25213 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
SATAN
Avatar

Send message
Joined: 27 Feb 09
Posts: 45
Credit: 305,963
RAC: 0
Message 25243 - Posted: 13 Jun 2009, 11:01:39 UTC

Well i've finally managed to get CUDA working properly on the Mac Pro. Not bad considering it's a slow old 8800GT.

So looking forward to the next code release, to see if I can finally get Milkyway GPU working on this damn thing.
Mars rules this confectionery war!
ID: 25243 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
1 · 2 · 3 · 4 . . . 6 · Next

Message boards : Application Code Discussion : CUDA for Milkyway@Home

©2024 Astroinformatics Group