Welcome to MilkyWay@home

GPU Requirements [OLD]

Message boards : Number crunching : GPU Requirements [OLD]
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 . . . 18 · Next

AuthorMessage
Bofferbrauer

Send message
Joined: 19 Jul 10
Posts: 3
Credit: 1,018,032
RAC: 0
Message 41035 - Posted: 22 Jul 2010, 20:32:00 UTC - in response to Message 40295.  
Last modified: 22 Jul 2010, 20:34:08 UTC

I've just acquired an s/h pc with a PCI 16 express slot, but due to it being a small form factor case, standard length/height cards wont fit. Are there any double precision half height cards out there that could be used to crunch MW? I can get an HD4550 but I think its only single precision.

Thanks


I think there are some HD 4770/4730 in small form factor, but it's possible that they where just made for the OEM market

THE NEW HD 5670 WILL ALSO PERFORM DP MATH.


That's odd, considering the fact that the 57xx have NO DP support
ID: 41035 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Mario

Send message
Joined: 29 Jul 10
Posts: 1
Credit: 248,293
RAC: 0
Message 41156 - Posted: 30 Jul 2010, 13:41:53 UTC

What about my ATI Radeon HD 4600? It's not DP? Does that mean MilkyWay@home can't use it at all?
ID: 41156 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile The Gas Giant
Avatar

Send message
Joined: 24 Dec 07
Posts: 1947
Credit: 240,884,648
RAC: 0
Message 41161 - Posted: 30 Jul 2010, 15:49:54 UTC - in response to Message 41156.  

What about my ATI Radeon HD 4600? It's not DP? Does that mean MilkyWay@home can't use it at all?

Correct.
ID: 41161 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Mutiny32*

Send message
Joined: 13 Aug 10
Posts: 15
Credit: 122,278
RAC: 0
Message 41590 - Posted: 19 Aug 2010, 18:33:29 UTC

Just curious, what is the difficulty in recompiling the software to support newer GPUs like the GF104 or other cards that have double-precision floating point operation support? Why the big delay? I'd consider Fermi support to be important considering the massive performance increase it brings. That is the whole point of projects like this; to crunch a vast amount of information as quickly and efficiently as possible, is it not?

Even if on the back-end you guys are refactoring your programs to use OpenCL for better cross-hardware and cross-platform support, wouldn't it be worth it to update the old code to accept newer cards as a stop-gap solution? My GTX 460 can literally chew through anything I can currently throw at it without breaking a sweat. I'd love to see how it handles MW when it has WUs that don't just get discarded because the binary doesn't think it has a Compute Capability 1.3 or newer card (the GF104 actually has Compute Capability of 2.1).
ID: 41590 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile Werkstatt

Send message
Joined: 19 Feb 08
Posts: 350
Credit: 141,284,369
RAC: 0
Message 41593 - Posted: 19 Aug 2010, 18:58:01 UTC - in response to Message 41590.  

It looks like they are currently working on the nbody-simulation, this requires their full attention. I started with them 7 hours ago and have ~40 wu's ended with errors.
As Travis posted a couple of days ago, they are working on an gpu-app for the simulation as well http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=1875&nowrap=true#41548.

Alexander
ID: 41593 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Matt Arsenault
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 8 May 10
Posts: 576
Credit: 15,979,383
RAC: 0
Message 41596 - Posted: 19 Aug 2010, 19:56:34 UTC - in response to Message 41590.  

Even if on the back-end you guys are refactoring your programs to use OpenCL for better cross-hardware and cross-platform support, wouldn't it be worth it to update the old code to accept newer cards as a stop-gap solution?).


It would actually be less work to just finish the OpenCL. I haven't been able to get the existing CUDA to build since I started working on the project. I actually expect to have the OpenCL actually working later today, and ready to send out within a week or 2.
ID: 41596 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile Crunch3r
Volunteer developer
Avatar

Send message
Joined: 17 Feb 08
Posts: 363
Credit: 258,227,990
RAC: 0
Message 41597 - Posted: 19 Aug 2010, 20:02:56 UTC - in response to Message 41596.  

Even if on the back-end you guys are refactoring your programs to use OpenCL for better cross-hardware and cross-platform support, wouldn't it be worth it to update the old code to accept newer cards as a stop-gap solution?).


It would actually be less work to just finish the OpenCL. I haven't been able to get the existing CUDA to build since I started working on the project. I actually expect to have the OpenCL actually working later today, and ready to send out within a week or 2.


I'm guessing that for "better cross-hardware and cross-platform support" you're willing to sacrifice an awfull lot of performance on all those ATI cards crunching here.

Any bets on what that factor will be on 48x0 cards ?

Join Support science! Joinc Team BOINC United now!
ID: 41597 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile Werkstatt

Send message
Joined: 19 Feb 08
Posts: 350
Credit: 141,284,369
RAC: 0
Message 41598 - Posted: 19 Aug 2010, 20:14:59 UTC - in response to Message 41597.  

As I've learned from gpugrid, openCL works fine on nVidia, but is not yet optimized for the ati-cards. And it runs only on 58xx cards.
Are these infos true for the mw-apps as well?

Regards,
Alexander
ID: 41598 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Matt Arsenault
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 8 May 10
Posts: 576
Credit: 15,979,383
RAC: 0
Message 41599 - Posted: 19 Aug 2010, 20:21:24 UTC - in response to Message 41597.  


I'm guessing that for "better cross-hardware and cross-platform support" you're willing to sacrifice an awfull lot of performance on all those ATI cards crunching here.


No. With OpenCL, the main piece of the application is compiled and optimized at runtime for the local hardware. For more complicated GPGPU things, you have to make special efforts for speed on different hardware in the application. However, this is a pretty simple application and is mostly just a giant sum. There's not really anything special to do for speed on different GPUs.
ID: 41599 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Matt Arsenault
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 8 May 10
Posts: 576
Credit: 15,979,383
RAC: 0
Message 41600 - Posted: 19 Aug 2010, 20:31:50 UTC - in response to Message 41598.  

As I've learned from gpugrid, openCL works fine on nVidia, but is not yet optimized for the ati-cards. And it runs only on 58xx cards.
Are these infos true for the mw-apps as well?


This doesn't really make sense. OpenCL is a specification for an API. It doesn't actually mean anything to say if it's been optimized or not for Nvidia or ATI. It depends on the OpenCL implementation coupled with the hardware. I'm not sure about ATI's CL implementation, but Apple's, Nvidia's, and the in progress one for Gallium3d are all based on a clang front end. ATI probably does too, so to some degree everyone gets a similar subset of optimizations to start with. The actual speed depends on the implementation for particular hardware.
ID: 41600 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Mutiny32*

Send message
Joined: 13 Aug 10
Posts: 15
Credit: 122,278
RAC: 0
Message 41602 - Posted: 19 Aug 2010, 20:54:41 UTC - in response to Message 41596.  

Even if on the back-end you guys are refactoring your programs to use OpenCL for better cross-hardware and cross-platform support, wouldn't it be worth it to update the old code to accept newer cards as a stop-gap solution?).


It would actually be less work to just finish the OpenCL. I haven't been able to get the existing CUDA to build since I started working on the project. I actually expect to have the OpenCL actually working later today, and ready to send out within a week or 2.

The existing CUDA won't build? You mean with the 3.1.1 toolkit? Couldn't you just use the old toolkit and code to make an exception for the newer cards? From what I understand, Fermi will pretty much run Tesla code out of the box unless the original coding wasn't done properly. Nvidia even has compatibility documentation for Fermi at http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/NVIDIA_FermiTuningGuide.pdf and even a tuning guide at http://developer.nvidia.com.

Like someone else said, you're going to have a LOT of failed WU's since you're doing the nbody-simulation now and the program downloads them before checking the GPU and subsequently having a ton WU's crash and burn on Fermi machines. Yeah, I know you can turn off the use GPU, but what if someone has a bunch of GPUs that CAN do the work as well as some Fermi cards? That wouldn't be ideal for them.
ID: 41602 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile Crunch3r
Volunteer developer
Avatar

Send message
Joined: 17 Feb 08
Posts: 363
Credit: 258,227,990
RAC: 0
Message 41604 - Posted: 19 Aug 2010, 21:26:43 UTC - in response to Message 41600.  

As I've learned from gpugrid, openCL works fine on nVidia, but is not yet optimized for the ati-cards. And it runs only on 58xx cards.
Are these infos true for the mw-apps as well?


This doesn't really make sense. OpenCL is a specification for an API. It doesn't actually mean anything to say if it's been optimized or not for Nvidia or ATI. It depends on the OpenCL implementation coupled with the hardware. I'm not sure about ATI's CL implementation, but Apple's, Nvidia's, and the in progress one for Gallium3d are all based on a clang front end. ATI probably does too, so to some degree everyone gets a similar subset of optimizations to start with. The actual speed depends on the implementation for particular hardware.


Here's a PDF describung what he's talking about.. (GPUGrid experiments with OpenCL... epic fail)
Experiences porting from CUDA to OpenCL -> http://www.cse.scitech.ac.uk/disco/mew20/presentations/GPU_MattHarvey.pdf

Join Support science! Joinc Team BOINC United now!
ID: 41604 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile Werkstatt

Send message
Joined: 19 Feb 08
Posts: 350
Credit: 141,284,369
RAC: 0
Message 41605 - Posted: 19 Aug 2010, 22:39:40 UTC - in response to Message 41604.  


Here's a PDF describung what he's talking about.. (GPUGrid experiments with OpenCL... epic fail)
Experiences porting from CUDA to OpenCL -> http://www.cse.scitech.ac.uk/disco/mew20/presentations/GPU_MattHarvey.pdf


This is from GPUGRID-forum:

ATI 2.2 fixes most serious bugs and we will probably put out a test application for ATI. The performance is still poor for several reasons. ATI on one side and us on the other, will work to make it faster. Most likely, SDK 2.3 could be a better release performance wise, now that bugs are under controls.

gdf


They also talk about 48xx cards and OpenCL --> simply NO.

So if someone plans to upgrade to a newer GPU, watch the development of OpenCL. OpenCL can help us continue to crunch if a server fails and we need crunchies for our gpu's. But if nVidia is better siuted for that, a Fermi-card could be eventually a better choice.
I really look forward to the mw-app; it will give a first impression of what's going on.

Alexander
ID: 41605 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Anton Novomlinski

Send message
Joined: 21 Aug 10
Posts: 4
Credit: 3,189
RAC: 0
Message 41719 - Posted: 24 Aug 2010, 13:00:21 UTC

Hello!
I have ATI Radeon HD5570 & it's a great card!
It's a waste that boinc doesnt use it to calculate...!!!
Please advice how can I make Boinc use it as well.....
ID: 41719 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile Crunch3r
Volunteer developer
Avatar

Send message
Joined: 17 Feb 08
Posts: 363
Credit: 258,227,990
RAC: 0
Message 41720 - Posted: 24 Aug 2010, 13:22:34 UTC - in response to Message 41719.  

Hello!
I have ATI Radeon HD5570 & it's a great card!
It's a waste that boinc doesnt use it to calculate...!!!
Please advice how can I make Boinc use it as well.....


You can't use that card here on MW. However you can use it here -> http://boinc.thesonntags.com/collatz/index.php
or there -> http://dnetc.net


Join Support science! Joinc Team BOINC United now!
ID: 41720 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile uBronan
Avatar

Send message
Joined: 9 Feb 09
Posts: 166
Credit: 27,520,813
RAC: 0
Message 41730 - Posted: 24 Aug 2010, 22:40:37 UTC
Last modified: 24 Aug 2010, 22:46:15 UTC

Only all 58XX and up cards have double precision.
To begin with the 5830 but i think not many will buy it since for a bit more you can buy a 5850 which performs much better, and has more streaming cores.

5830 - 1120
5850 - 1440
5870 - 1600
5970 - 3200

Its sad that Ati decided not to implement double precision on the 5700 serie cards but i guess it has todo with the cost reduction for manufacture.

My good old 4770 still does well at processing its most of the time faster then my old 4830
Its new, its relative fast... my new bicycle
ID: 41730 · Rating: 0 · rate: Rate + / Rate - Report as offensive
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
Message 41733 - Posted: 24 Aug 2010, 23:41:16 UTC - in response to Message 41730.  
Last modified: 24 Aug 2010, 23:42:20 UTC

Only all 58XX and up cards have double precision.



I find this an unfortunately incomplete statement.

All ATI HD48xx and HD38xx series cards have double precision implemented and can successfully crunch Milkyway work quite readily.

When the server runs constantly, I have had an HD3850 crunching Milkyway work for over a year.
Go away, I was asleep


ID: 41733 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Anton Novomlinski

Send message
Joined: 21 Aug 10
Posts: 4
Credit: 3,189
RAC: 0
Message 41737 - Posted: 25 Aug 2010, 12:08:58 UTC

I still dont get why this project doesnt support Single Precision arithmetic as well..
ID: 41737 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Anton Novomlinski

Send message
Joined: 21 Aug 10
Posts: 4
Credit: 3,189
RAC: 0
Message 41739 - Posted: 25 Aug 2010, 16:20:58 UTC

Maybe there are other pointless
ID: 41739 · Rating: 0 · rate: Rate + / Rate - Report as offensive
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
Message 41740 - Posted: 25 Aug 2010, 16:55:54 UTC

Mapping THE Milkyway Galaxy requires a significant level of precision, to, i think, more than 16 places of decimals.

The way computer single precision is implemented will only give around 11 or 12 decimal places. this means the science cannot be carried out, unless the hardware is capable of double precision.

So, the GPU implementation of this project is only for ATI and nVidia cards capable of double precision, and the ATI GPUs have had this facility for several years. Even the Fermi nVidia GPU implementation has the double precision capability limited (crippled).

I am sure I will be heavily corrected by more knowledgeable posters, but I think this is the gist of the case?
Go away, I was asleep


ID: 41740 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 . . . 18 · Next

Message boards : Number crunching : GPU Requirements [OLD]

©2024 Astroinformatics Group