Advanced search

Message boards : Graphics cards (GPUs) : AMD support on gpugrid, 2014 request

Author Message
EmSti [BlackOps]
Send message
Joined: 21 Apr 12
Posts: 4
Credit: 37,802,925
RAC: 0
Level
Val
Scientific publications
watwatwatwatwat
Message 34743 - Posted: 20 Jan 2014 | 22:16:47 UTC

Last year there was a 2013 request, so why not for 2014. Last year's answer was basically nothing was planned at the time because of some difficulty with the amd drivers and we could support Donate gpugrid's sister project.

Now that donate seems to be dead is there any plans to revisit AMD gpus on gpugrid?

A guy can hope.

John C MacAlister
Send message
Joined: 17 Feb 13
Posts: 181
Credit: 144,871,276
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 34746 - Posted: 21 Jan 2014 | 11:25:12 UTC - in response to Message 34743.

Last year there was a 2013 request, so why not for 2014. Last year's answer was basically nothing was planned at the time because of some difficulty with the amd drivers and we could support Donate gpugrid's sister project.

Now that donate seems to be dead is there any plans to revisit AMD gpus on gpugrid?

A guy can hope.


Hi, EmSti:

As I have three AMD GPUs on two different machines,I would also like to see this project using AMD GPUs, but I think the cost is a deterrent....Given the state of the Spanish economy at this time, I think nothing will happen.

John

Stefan
Project administrator
Project developer
Project tester
Project scientist
Send message
Joined: 5 Mar 13
Posts: 348
Credit: 0
RAC: 0
Level

Scientific publications
wat
Message 34747 - Posted: 21 Jan 2014 | 13:26:57 UTC
Last modified: 21 Jan 2014 | 13:28:10 UTC

We would probably need to hire a new guy to do just this and the relative cost/advantage will probably be too low at this point. I am sure there are people out there who would like to contribute their AMDs but it doesn't seem probable that it will be implemented in the near future.

But believe me that the moment it becomes an easy thing to implement we will do it. It's not like we purposely want to miss out on computational power :)

TJ
Send message
Joined: 26 Jun 09
Posts: 815
Credit: 1,470,385,294
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34748 - Posted: 21 Jan 2014 | 16:03:35 UTC

Yes AMD GPU's have power and are cheaper then nVidia's.
My old 5470 still outperforms a GTX690 on MilkyWay@home!
Would be very awesome if we could use them here. I think the programming is not that hard. You could contact Einstein@home. They are very skilled and very helpful.

____________
Greetings from TJ

_Ryle_
Send message
Joined: 7 Jun 09
Posts: 24
Credit: 1,138,093,416
RAC: 102,609
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34751 - Posted: 21 Jan 2014 | 17:03:56 UTC - in response to Message 34748.

TJ, I don't think you can compare it like that. Milkyway was always faster on AMD GPUs. It's to do with the applications, running OpenCL i think.

Anyways, it would be nice with some AMD cards here, but I can understand the challenges facing it.

RaymondFO*
Send message
Joined: 22 Nov 12
Posts: 72
Credit: 14,040,706,346
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34755 - Posted: 22 Jan 2014 | 2:28:27 UTC - in response to Message 34748.
Last modified: 22 Jan 2014 | 2:28:42 UTC

Yes AMD GPU's have power and are cheaper then nVidia's.
My old 5470 still outperforms a GTX690 on MilkyWay@home!
Would be very awesome if we could use them here. I think the programming is not that hard. You could contact Einstein@home. They are very skilled and very helpful.



TJ,

The reason why a 5470 can outperform a GTX690 at Milkyway@home ("MW") is because the MW tasks are double precision ("DP") intensive and the 600 and 700 NVIDIA series have had their DP capabilities crippled. I am being nice about using the word crippled. The 500 series was the last NVIDIA card series to have relatively decent (I did not say good) DP performance, and from what I have read this situation may not change going forward.

The TITAN cards on paper have very good DP performance, however when you engage the TITAN DP capabilities, the GPU clock speed will decrease to a point of not being all that impressive. Again, I am being nice about this topic. The one project I have found the TITAN DP capabilities to be worthwhile/superior is on PrimeGrid's Genefer "World Record" and "Short" tasks using their OpenGL tasks.

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 34758 - Posted: 22 Jan 2014 | 11:22:30 UTC - in response to Message 34751.

The applications used at GPUGrid are ACEMD versions, designed and maintained for CUDA research. For GPUGrid to use AMD GPU's (or Intel's) the research would need to be different and the staff would need to learn how to use different apps, written for OpenCL/GL. For existing PhD researchers this isn't feasible. It would first take the senior team to explore and develop research possibilities, seed the research with post-grad diplomas/MSc's and then begin a PhD project. However, even that might need to be supported with experienced OpenCL/GL researchers; Post Docs. It's a big undertaking to do it right. There is little point throwing out a buggy OpenCL/GL app that runs 1/5th as fast and the staff can't support. If existing researchers wanted to expand their experience/expertise and use other apps then using CPU apps might be feasible for some.
____________
FAQ's

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

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34783 - Posted: 23 Jan 2014 | 8:44:39 UTC

Well now that we know hoping for an OpenCL app here is pointless let's move on to the next phase which in 2013 was, IIRC, a pointless discussion about why CUDA is better than OpenCL and why AMD and Intel should just throw in the towel, bow before their NVIDIA masters and beg for license to use CUDA.

I found the answer to the third question in the FAQ at this site rather interesting. Renowned hi-perf parallel computing guru Wen-mei W. Hwu says in a nutshell that from a programmer's perspective OpenCL sucks and CUDA rules. So, AMD, on your knees, kiss the ring and repeat after me, "Oh great NVIDIA my liege, I your lowly and unworthy vassal do humbly beg for the privilege of using CUDA, a privilege for which I agree to pay 2,000 goats and 1,000 plump, plucked, poulets yearly."
____________
BOINC <<--- credit whores, pedants, alien hunters

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34809 - Posted: 24 Jan 2014 | 22:37:41 UTC - in response to Message 34748.

My old 5470 still outperforms a GTX690 on MilkyWay@home!

Nope, that card can't run MW at all since it doesn't have any DP capabilities. You probably meant HD5870?

@Raymond: well, all GCN chips below Tahiti also crunch DP only at 1/16 SP, which is also crippled. Nvidia is not alone with this.

@SK: I don't think it's this bad. Parallela, the firm behind ACEMD could port the app to OpenCL (again), if they can afford the development. Ideally this part of the project would involve professionals rather than students below PhD level ;)

MrS
____________
Scanning for our furry friends since Jan 2002

TJ
Send message
Joined: 26 Jun 09
Posts: 815
Credit: 1,470,385,294
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34819 - Posted: 25 Jan 2014 | 16:02:27 UTC - in response to Message 34809.

My old 5470 still outperforms a GTX690 on MilkyWay@home!

Nope, that card can't run MW at all since it doesn't have any DP capabilities. You probably meant HD5870?
MrS


Yep that's one I meant. You are right again :)
____________
Greetings from TJ

Stefan
Project administrator
Project developer
Project tester
Project scientist
Send message
Joined: 5 Mar 13
Posts: 348
Credit: 0
RAC: 0
Level

Scientific publications
wat
Message 34831 - Posted: 27 Jan 2014 | 9:51:53 UTC - in response to Message 34809.
Last modified: 27 Jan 2014 | 9:52:02 UTC

Parallela, the firm behind ACEMD


Acellera :)
Yes but for a company to do it there usually has to be some profit in it. From what I understand there is quite low interest in OpenCL in the industry.

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34832 - Posted: 27 Jan 2014 | 14:32:07 UTC - in response to Message 34831.

Low interest in OpenCL in the High Performance Parallel Computing industry tells me the industry sees less profit in OpenCL and more profit in CUDA. There must be a reason why they think that way. Maybe they're just "silly". Maybe they see CUDA as the least expensive and/or easiest path to a profitable product. Maybe NVIDIA pays kickbacks. When the more expensive GPUs use CUDA and the less expensive GPUs use OpenCL they must have a darn good reason for going with CUDA.

____________
BOINC <<--- credit whores, pedants, alien hunters

RaymondFO*
Send message
Joined: 22 Nov 12
Posts: 72
Credit: 14,040,706,346
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34839 - Posted: 28 Jan 2014 | 18:59:58 UTC - in response to Message 34809.
Last modified: 28 Jan 2014 | 19:08:20 UTC

[quote]

@Raymond: well, all GCN chips below Tahiti also crunch DP only at 1/16 SP, which is also crippled. Nvidia is not alone with this.

MrS



Do you know if the DP issue also applies to the R9 280X video cards as well? Thought the R9 280X was a rebranded 7970. Thank you for your input.

Ray

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34840 - Posted: 28 Jan 2014 | 22:43:53 UTC - in response to Message 34839.

[quote]

@Raymond: well, all GCN chips below Tahiti also crunch DP only at 1/16 SP, which is also crippled. Nvidia is not alone with this.

MrS



Do you know if the DP issue also applies to the R9 280X video cards as well?


I assume you're asking if R9 280X cards are also DP crippled. The short answer to that is "yes, but not by a factor of 16". How do I know or, more important, how can you know on your own?

Start with this comparison of AMD GPUs for a good summary based on official AMD data all in one convenient table. There is also a comparison of NVIDIA GPUs table.

The R9 280X (Tahiti) is in the section titled Volcanic Islands (Rx 200) Series since Tahiti is a volcanic island. Footnote 5 in that section says (bolding added by me)...

Double precision performance of Hawaii is 1/8 of single precision performance,[19] Tahiti is 1/4 of single precision performance, others 28 nm chip is 1/16 of single precision performance.


...which would lead one to believe R9 280 DP is crippled by a factor of 4 but that's not true. Let's take a look at the actual SP and DP processing power. For the R9 280 (Tahiti Pro) the table says the computational speeds in GFlops are:
Single precision Double precision 3046.4 761.6 3315.2 828.8


I assume the figures on the second line are speeds at maximum boost, "turbo mode" or whatever they call it. The main point is that for both lines SP / DP = 4 (single precision divided by double precision equals 4) which means on that model SP is 4X faster than DP which is consistent with footnote 5 however due to the fact DP requires 2X as many operations as SP, SP will always be 2X faster than DP even on non-crippled hardware therefore they've crippled DP speed by a factor of 2 on R9 280X.

Thought the R9 280X was a rebranded 7970.

Data in the table seems to indicate R9 280X is a rebranded 7970 but I could be wrong. 7970 and R9 280X are both code name Tahiti but I see there are Pro, XT and XT2 variants of Tahiti. All are 28nm tech so I would say they are all the same chip and the difference is just in the binning but there could be other stuff going on there that I am not aware of.

____________
BOINC <<--- credit whores, pedants, alien hunters

RaymondFO*
Send message
Joined: 22 Nov 12
Posts: 72
Credit: 14,040,706,346
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34845 - Posted: 30 Jan 2014 | 3:11:23 UTC - in response to Message 34840.

Dagorath,

Thank you for your input and assistance and I have bookmarked those two links for future reference. I do not believe you are wrong about the R9 280X being a rebranded 7970.

Ray

Profile Mumak
Avatar
Send message
Joined: 7 Dec 12
Posts: 92
Credit: 225,897,225
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 34846 - Posted: 30 Jan 2014 | 9:34:41 UTC

Yes, the R9 280(X) is just a rebranded HD 7950/7970 (Tahiti Pro/XT). They belong to the Southern Islands family.
The R9 290(X) (Hawaii) is from the Volcanic Islands series and unfortunately these have the DP performance crippled to 1/8. So even though the R9 290(X) has more shaders, in DP it will perform worser than the Tahiti.

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34850 - Posted: 31 Jan 2014 | 2:49:22 UTC - in response to Message 34846.

You're welcome. Ray. I like tables, lots of info all in one place.

Yes, the R9 280(X) is just a rebranded HD 7950/7970 (Tahiti Pro/XT). They belong to the Southern Islands family.
The R9 290(X) (Hawaii) is from the Volcanic Islands series and unfortunately these have the DP performance crippled to 1/8. So even though the R9 290(X) has more shaders, in DP it will perform worser than the Tahiti.


Paying more money but getting lower DP performance would be disappointing.

Now a little pedantic word mincing to test one's knowledge of how the crippling done by NVIDIA and AMD works. Mumak correctly states these have the DP performance crippled to 1/8. The word to is the proper choice and it would be incorrect to say these have the DP performance crippled by 1/8 since limitations of the hardware mean DP performance must always be 1/2 of SP. It would, however, be correct to say these have the DP performance crippled by 1/4. If one can understand that bit of pedantry then one understands a bit about quantifying the extent of DP crippling as well as a bit about word usage.

____________
BOINC <<--- credit whores, pedants, alien hunters

Profile Mumak
Avatar
Send message
Joined: 7 Dec 12
Posts: 92
Credit: 225,897,225
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 34852 - Posted: 31 Jan 2014 | 7:20:46 UTC - in response to Message 34850.

OK, to be more precise ;-)
Hawaii-XT/PRO have the DPFP_RATE (Double-precision floating point rate) field of CC_GC_SHADER_ARRAY_CONFIG/GC_USER_SHADER_ARRAY_CONFIG registers set to 2, which means 1/4 of original DP performance (=1/8 of SP), where the workstation series have no such limit and the setting is 0 (=full), which means 1/2 of SP :-DD

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34853 - Posted: 31 Jan 2014 | 9:04:04 UTC - in response to Message 34852.

You win! I can't out precise that one :-)

____________
BOINC <<--- credit whores, pedants, alien hunters

Post to thread

Message boards : Graphics cards (GPUs) : AMD support on gpugrid, 2014 request

//