GPU Requirements
log in

Advanced search

Message boards : Number crunching : GPU Requirements

1 · 2 · 3 · 4 . . . 15 · Next
Author Message
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 49250 - Posted: 13 Jun 2011, 13:55:13 UTC
Last modified: 10 Feb 2012, 9:01:28 UTC

Milkyway@Home requires a GPU supporting Double Precision arithmetic.

NVIDIA:
- Requires Compute Capability 1.3 and Above.
- For the GeForce 2xx series, this is the GTX 260 and above.
- Any Fermi based (GeForce GTX 4xx or 5xx) should support doubles.
- Any older GPUs (such as a GeForce 8xxx or 9xxx) will not work

AMD/ATI:
- The oldest GPUs that work are the ATI HD Radeon 38x0 series.
- In general laptop AMD GPUs do NOT support doubles despite similar branding (e.g. A Mobility Radeon 5870 is not the same as a normal Radeon 5870). Only the Mobility Radeon 48xx are the only current ATI laptop GPUs that have doubles.
- For the Radeon 6000 series, only the 69xx have doubles.


Examples (these lists are not all of the GPUs which should work)
Nvidia:


    GeForce GTX 590
    GeForce GTX 580
    GeForce GTX 570
    GeForce GTX 560 Ti
    GeForce GTX 560
    GeForce GTX 550 Ti
    GeForce GT 545
    GeForce GTX 480
    GeForce GTX 470
    GeForce GTX 465
    GeForce GTX 460
    GeForce GTS 450
    GeForce GT 430
    Geforce GTX 295
    Geforce GTX 285
    Geforce GTX 280
    Geforce GTX 275 (credits to Bruce)
    Geforce GTX 260
    Tesla S1070
    Tesla C1060
    Tesla M2090
    Tesla M2070
    Tesla M2050
    Tesla S2050
    Quadro Plex 2200 D2
    Quadro FX 5800
    Quadro FX 4800
    Quadro 5000s, 5000s, 4000s
    (Based on GT200 GPU)



AMD/ATI:

    AMD Radeon 7970
    AMD Radeon 7950
    AMD Radeon 6990
    AMD Radeon 6970
    AMD Radeon 6950
    AMD Radeon 6930
    ATI HD Radeon 5970 (credits to kashi)
    ATI HD Radeon 5870
    ATI HD Radeon 5850
    ATI HD Radeon 5830
    ATI HD Radeon 4890
    ATI HD Radeon 4870
    ATI HD Radeon 4850
    ATI HD Radeon 4830
    ATI HD Radeon 4770
    ATI HD Radeon 4830
    ATI HD Radeon 38x0
    (credits to cenit for the AMD documentation describing the products above)
    ATI Firestream 9270
    ATI Firestream 9250
    ATI Firestream 9170
    (credits to Cluster Physik)



For AMD, you can try checking the chart at the end of this if you have something not listed: http://developer.amd.com/sdks/amdappsdk/assets/AMD_APP_SDK_Getting_Started_Guide_v2.4.pdf

Profile skgiven
Avatar
Send message
Joined: 22 Dec 07
Posts: 35
Credit: 18,298,040
RAC: 95

Message 49465 - Posted: 20 Jun 2011, 22:48:19 UTC
Last modified: 20 Jun 2011, 22:58:25 UTC

You might want to mention that most of the NVidia GeForce cards are poor for double precision tasks, especially the Fermi cards.

May I suggest you add the GeForce GTX 465, GTX 550 Ti, GTX 560, GTX 560 Ti, GTX 570 and GTX 590, and if you are going to include the GeForce GT 430, you might as well include the GT 545 versions. Not sure about the GT 530 though, is it worth including?

While you are at it, and due to their 1/2 FP32 performance you might want to add the the extremely expensive,
Tesla's:
Tesla C2050 (515 gigaflops of double precision performance) 3GB
Tesla C2070 (Workstation) same as the C2050 but 6GB
These are similar to the GTX470 cards; 448:56:40, but without as much disabled, with more GDDR5 and with ECC - which when on further reduces performance.

and the even more expensive,
DataCentre Tesla's:
Tesla M2090 (665 DP Gigaflops)
Tesla M2070, Tesla M2050 and S2050 (515 DP GFlops)

Oh, and the,
Quadro Desktop's:
Quadro 6000, 5000, 4000 should all work.

    Relative SP to DP performance's:
    Tesla series: DP performance is 1/2 GF100 SP performance,
    Quadro 4000-6000: DP performance is 1/2 SP
    GTX 5xx: DP is 1/8 SP
    Quadro 600-2000: DP is 1/12 SP


Then there are all those OEM and Notebook GPU's ;)

Johnny The Dreamer
Send message
Joined: 16 Apr 10
Posts: 5
Credit: 47,386,311
RAC: 0

Message 49475 - Posted: 21 Jun 2011, 18:14:53 UTC
Last modified: 21 Jun 2011, 18:23:02 UTC

I see that HD4730 it's not listed here, nor is listet anywhere as non DP capable.

So my question is: Does it support DP or not? Did anyone tested it?

And yes, I googled it few times and did not found anything.

Thank you.

Update: onik1318 said that it does support it. You should add it to the list!

VictordeHollander
Send message
Joined: 9 Nov 10
Posts: 15
Credit: 59,289,544
RAC: 0

Message 49518 - Posted: 23 Jun 2011, 22:41:15 UTC - in response to Message 49465.

You might want to mention that most of the NVidia GeForce cards are poor for double precision tasks, especially the Fermi cards.

Tesla's:
Tesla C2050 (515 gigaflops of double precision performance) 3GB


The Tesla C2050 costs something like EUR 2.000+ (not sure how much that in US Doller is) and in comparison to the latest ATI/AMD cards has relatively poor DP performance.

For example:
AMD 6950 costs something like EUR 200 and delivers 563 Gflops in DP!

source: http://www.legitreviews.com/article/1488/1/

No wonder that the top 20 computers all use AMD/ATI GPUs.

robertmiles
Send message
Joined: 30 Sep 09
Posts: 205
Credit: 5,078,643
RAC: 2,439

Message 49521 - Posted: 24 Jun 2011, 3:48:28 UTC

I'd rewrite that to say the top 20 companies that need heavy double precision use and don't depend on function libraries written elsewhere. Nvidia still has the best selection of function libraries written elsewhere, especially for CUDA.

VictordeHollander
Send message
Joined: 9 Nov 10
Posts: 15
Credit: 59,289,544
RAC: 0

Message 49522 - Posted: 24 Jun 2011, 5:21:33 UTC - in response to Message 49518.


No wonder that the top 20 computers all use AMD/ATI GPUs.

I was referring to the top 20 computers on Milkyway@home =).

FruehwF
Send message
Joined: 28 Feb 10
Posts: 120
Credit: 109,830,617
RAC: 0

Message 49527 - Posted: 24 Jun 2011, 10:22:39 UTC

You might want to mention that most of the NVidia GeForce cards are poor for double precision tasks, especially the Fermi cards.


I thing the word "especially" isn't really ok in this sentence.

The GeForce GTX Cards do have 1/8 of SP Performance as DP Performance.
The Fermi and the Quattro Cards have 1/2. Thats 4 time more.

The ATI (AMD) cards have 1/5 (1/4 HD69xx) of SP performance as DP performance.
no matter if HD cards or Firestream Cards.

BUT the SP Perf. of an HD 6970 is 2703 GFlops
the GTX 580 reaches 1581 GFLops.
So we have
ATI 676 GFlops maximal DP Performance versus
Nividia 198 GFlops DP Perf for the Mainstream GPU's
but
515 GFLops with the professional cards from Nividia (Quattro 5000 & 6000 & Tesla C2050)

Profile BladeD
Avatar
Send message
Joined: 2 Nov 10
Posts: 731
Credit: 131,536,342
RAC: 0

Message 49531 - Posted: 24 Jun 2011, 14:33:38 UTC - in response to Message 49527.

Yes, he should have just said....

You might want to mention that most of the NVidia cards are poor for double precision tasks.
____________

Profile skgiven
Avatar
Send message
Joined: 22 Dec 07
Posts: 35
Credit: 18,298,040
RAC: 95

Message 49543 - Posted: 24 Jun 2011, 19:42:39 UTC

"You might want to mention that most of the NVidia GeForce cards are poor for double precision tasks, especially the Fermi cards".

I dribbled on and lost the ball...

As suggested, Except the Fermi cards would have been closer to the mark, if you consider the GeForce cards not to be full fat Fermi's, and "most of the NVidia GeForce cards are poor for double precision tasks" might have sufficed - there is probably not many Tesla's here and those that forked out for one would probably have read the manual.

Anyway, Matt has added some compatible cards to the main list.

Cheers guys,

Michael Corderoy
Send message
Joined: 14 Apr 11
Posts: 3
Credit: 286,714
RAC: 0

Message 49561 - Posted: 25 Jun 2011, 5:51:13 UTC - in response to Message 49250.
Last modified: 25 Jun 2011, 6:00:52 UTC

i have been getting messages saying my GPU lacks the needed compute capability.
then the message says i need 1.3, required for double precision math.
what would you suggest i have installed to so i dont get this message, without costing too much. thanks
____________

Michael Corderoy
Send message
Joined: 14 Apr 11
Posts: 3
Credit: 286,714
RAC: 0

Message 49562 - Posted: 25 Jun 2011, 5:51:17 UTC - in response to Message 49250.

NVIDIA:
- Requires Compute Capability 1.3 and Above.
What is meant by 1.3 and above?
I have NVIDIA
geforce 220 installed.
thanks.
____________

Zydor
Avatar
Send message
Joined: 24 Feb 09
Posts: 620
Credit: 100,587,625
RAC: 0

Message 49566 - Posted: 25 Jun 2011, 9:31:59 UTC - in response to Message 49562.
Last modified: 25 Jun 2011, 9:37:41 UTC

Your card is Compute Capability 1.2. The messages you are getting are saying you must have Compute Capability 1.3 inclusive or above to run the application. That will - as you have already guessed - mean a new card in order to run that application.

Compute Capability
The term indicates at a technical level what the card is capable of doing, it applies to NVidia Cards only, AMD cards have different terminology. The details of that do not matter to us mere mortels, but it is important to wrap the head around what card is capable of what. There is a good Wikipedia Summary on it - "read over" technical details, and absorb the overview bits in particular the "Supported GPUs" table (scroll down the wikipedia article two pages and you'll see it.

http://en.wikipedia.org/wiki/CUDA#Version_features_and_specifications

Any card in that table with Compute Capability 1.2 inclusive or below will not work for the application your card is trying to run. In the table the cards of 1.3 and above will run it. As always the table is "dynamic" and a couple of cards may be missing from the list - but it will be fine for what you need.

Scroll further down and you see the technical differences - worth doing even if you are non-tech as it will give an overall feel for it, and show that its more than "tech speak", there are solid everyday differences.

http://developer.nvidia.com/cuda-gpus - is a straightfoward Card - Compute Capability list on the NVidia Site

New Card Selection
Thats a doozy :) It depends on a lot - what you are going to do with the card, what your overall priorities are(eg, BOINC, Gaming, other non-BOINC Tasks), finance and current overall PC hardware spec. There unfortunately is no short cut to a simple statement "go buy XXX card", at least not if you are to avoid wasting money.

Part of that discussion is also whether you buy NVidia, or AMD. In a crude sense, outside BOINC for most people, there is not really much difference to everyday useage as the use of various NVidia or AMD specific languages doesnt come into play (eg for gaming) - easy to create such a list just for the sake of it, but in the broad scheme of things forget it, its just a question of the card power/cost ratio. Inside BOINC its a whole new ball game, and there are Big differences AMD Versus NVidia Versus Project Application, could write a book on it - some do rofl :)

Suggested Way Ahead
Go to the Number Crunching Forum, start a new thread "Suggestions for a New Graphics Card". In first post list what you want to use the card for, both inside and outside BOINC, the spec of the PC you want to run it on, and the budget level you are prepared to pay for it. That will keep detailed discussion off this thread.

This will probably seem long winded, but with card selection long term, unless you focus on this and get suggestions and reasons, you *will* end up with the wrong card and waste money long term.

Regards
Zy

Profile skgiven
Avatar
Send message
Joined: 22 Dec 07
Posts: 35
Credit: 18,298,040
RAC: 95

Message 49573 - Posted: 25 Jun 2011, 12:35:51 UTC - in response to Message 49566.

Compute Capable 1.2 cards have no double precision capabilities. These stripped down 40nm cards use to be good entry level cards elsewhere, being efficient for single precision jobs, but lacking DP they were never any use here; hence they are not on Matt's list of compatible cards.

Profile skgiven
Avatar
Send message
Joined: 22 Dec 07
Posts: 35
Credit: 18,298,040
RAC: 95

Message 49577 - Posted: 25 Jun 2011, 18:42:12 UTC - in response to Message 49250.

... and the AMD/ATI HD cards that don't do DP (well enough):

6850
6790
6670
6570
6450

5770
5750
5670
5570
5470

4770
4670
4650
4550
4350

Profile Crunch3r
Volunteer developer
Avatar
Send message
Joined: 17 Feb 08
Posts: 363
Credit: 258,180,004
RAC: 1

Message 49578 - Posted: 25 Jun 2011, 18:48:43 UTC - in response to Message 49577.

... and the AMD/ATI HD cards that don't do DP (well enough):

6850
6790
6670
6570
6450

5770
5750
5670
5570
5470

4770
4670
4650
4550
4350


wrong ... the 4770 can do DP...

____________

Join Support science! Joinc Team BOINC United now!

Profile Beyond
Send message
Joined: 15 Jul 08
Posts: 383
Credit: 501,817,389
RAC: 0

Message 49592 - Posted: 26 Jun 2011, 14:22:43 UTC - in response to Message 49578.

... and the AMD/ATI HD cards that don't do DP (well enough)

wrong ... the 4770 can do DP...

And do it well in regard to credits/watt.

Profile skgiven
Avatar
Send message
Joined: 22 Dec 07
Posts: 35
Credit: 18,298,040
RAC: 95

Message 49603 - Posted: 26 Jun 2011, 19:08:42 UTC - in response to Message 49592.

Good you spotted that, though Matt had included the 4770 on his list of good GPUs.
I just listed the non DP cards from the AMD pdf link which suggested the 4770 could not do DP.

FruehwF
Send message
Joined: 28 Feb 10
Posts: 120
Credit: 109,830,617
RAC: 0

Message 49606 - Posted: 26 Jun 2011, 20:13:00 UTC - in response to Message 49603.

Good you spotted that, though Matt had included the 4770 on his list of good GPUs.
I just listed the non DP cards from the AMD pdf link which suggested the 4770 could not do DP.


In this case, the list is wrong. The HD4770 has a RV740 GPU this is an RV770 with 8 Shader-Clusters and in 40 nm Technologie (RV770 uses 55 nm).

So it does DP as a lot of particpations demonstrate every day ;-)

Profile Beyond
Send message
Joined: 15 Jul 08
Posts: 383
Credit: 501,817,389
RAC: 0

Message 49611 - Posted: 26 Jun 2011, 22:56:06 UTC - in response to Message 49606.

The HD4770 has a RV740 GPU this is an RV770 with 8 Shader-Clusters and in 40 nm Technologie (RV770 uses 55 nm).
So it does DP as a lot of particpations demonstrate every day ;-)

I've had up to 10 HD 4770 cards running MilkyWay at any given time. They seem to run credits/watt pretty much the same as other 40nm ATIs.

Michael Corderoy
Send message
Joined: 14 Apr 11
Posts: 3
Credit: 286,714
RAC: 0

Message 49624 - Posted: 27 Jun 2011, 11:30:41 UTC - in response to Message 49566.

Thanks heaps, i will look into it amd get a decent card that will do the job with room to spare.
much appriciated.

1 · 2 · 3 · 4 . . . 15 · Next
Post to thread

Message boards : Number crunching : GPU Requirements


Main page · Your account · Message boards


Copyright © 2016 AstroInformatics Group