Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Generalized Fermat Prime Search :
List of Compatible Nvidia GPUs
Author |
Message |
Michael Goetz Volunteer moderator Project administrator
 Send message
Joined: 21 Jan 10 Posts: 13804 ID: 53948 Credit: 345,369,032 RAC: 4,797
                              
|
The GeneferCUDA software only runs on an Nvidia GPU that has double precision floating point hardware. If you wish to do GFN work units on your CPU, you can crunch the "Short tasks" on BOINC or smaller values of N on the PSA Generalized Fermat Number Prime Search. At this time, there's no software for ATI GPUs, and none is currently planned as far as I know.
The following GPUs are compatible with GeneferCUDA:
- GTX Titan
- Geforce GTX 600 and GT 600M series GPUs
- Geforce GTX 500 and 500M series GPUs (*)
- Geforce GTX 400 and 400M series GPUs
- Geforce GTX 260 or above only (260, 275, 280, 285, 295)
- Tesla C2000 series GPUs
- Tesla C1000 series GPUs
- Quadro: See Nvidia's list or the PrimeGrid Wiki list below.
Of particular note is that all GeForce 300 series GPUs, all mobile GeForce 200 series GPUs, and the GeForce GTS 250 and below GPUs do not have double precision hardware and therefore can not be used.
This list will change over time. An up to date list can be found on Nvidia's List of CUDA capable GPUs. You can also view the PrimeGrid's Wiki GPU compatibility list. Any GPU listed as having compute capability 1.3, 2.0, or 2.1 may be used. A GPU may not be used if it has CC 1.0, 1.1, or 1.2, or if it is not on the list.
(*) Special note about the GTX 550 TI: You may need to underclock the memory clock or use exceptional cooling methods to get this card to operate reliably.
____________
My lucky number is 75898524288+1
| |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
(*) Special note about the GTX 560 TI: For reasons that are not well understood, it seems that the GTX 560 TI encounters "maxErr exceeded" errors at lower levels than do other video cards. At this time, there's no solution nor workaround for this problem.
The above should refer to the GTX 550 Ti. The GTX 560 Ti versions do not appear to have this problem. The problem may also exist on certain GTX 460 versions (those with 1024mb of video RAM with the 192-bit bus).
____________
141941*2^4299438-1 is prime!
| |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
Quadro cards that are compatible (i.e., with Double Precision) include:
Quadro 6000
Quadro 5000
Quadro 4000
Quadro 2000
Quadro 600
Quadro FX 5800
Quadro FX 4800
Quadro FX 3800
Quadro CX
Quadro 1000M
Quadro 2000M
Quadro 3000M
Quadro 4000M
Quadro 5000M
Quadro 5010M
____________
141941*2^4299438-1 is prime!
| |
|
Michael Goetz Volunteer moderator Project administrator
 Send message
Joined: 21 Jan 10 Posts: 13804 ID: 53948 Credit: 345,369,032 RAC: 4,797
                              
|
(*) Special note about the GTX 560 TI: For reasons that are not well understood, it seems that the GTX 560 TI encounters "maxErr exceeded" errors at lower levels than do other video cards. At this time, there's no solution nor workaround for this problem.
The above should refer to the GTX 550 Ti. The GTX 560 Ti versions do not appear to have this problem. The problem may also exist on certain GTX 460 versions (those with 1024mb of video RAM with the 192-bit bus).
Thanks. This is what happens when I get on the boards 1st thing in the morning, specifically BEFORE I find the coffee.
As far as you know, has this problem actually been observed on the that particular 460 version? (That's the v2 version of the 460? Sometimes I can't keep up with Nvidia's marketing department.) The more I think about this problem, the less inclined I am to believe that the memory architecture could break the software. But it can affect the speed of the software, and that could cause the FFT library to construct its execution plan differently. That, in turn, could affect the floating point instructions that are used, and the number of floating point instructions that are used, which could cause the rounding errors to be greater.
Even if that's correct, it doesn't get us a lot closer to solving the problem.
____________
My lucky number is 75898524288+1 | |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
Thanks. This is what happens when I get on the boards 1st thing in the morning, specifically BEFORE I find the coffee.
As far as you know, has this problem actually been observed on the that particular 460 version? (That's the v2 version of the 460? Sometimes I can't keep up with Nvidia's marketing department.) The more I think about this problem, the less inclined I am to believe that the memory architecture could break the software. But it can affect the speed of the software, and that could cause the FFT library to construct its execution plan differently. That, in turn, could affect the floating point instructions that are used, and the number of floating point instructions that are used, which could cause the rounding errors to be greater.
Even if that's correct, it doesn't get us a lot closer to solving the problem.
Unfortunately (or maybe fortunately), I don't have one of those 460's to observe. I think that they are somewhat rare, at least compared to the 550's and original 460's. I mentioned it hoping that someone with that version could confirm/deny similar errors.
As for the memory architecture thing, I think you may be right in your thinking. I know that the fix for the 550 on other software (NVidia got some complaints when they first came out regarding some issues with some software) was a driver fix. Guess I need to thumb back through the driver release notes to see if I can find when (and maybe what was done) they addressed the issue.
____________
141941*2^4299438-1 is prime!
| |
|
|
This is a nice spreadsheet that I refer to quite a bit. It's a good one to bookmark.
The Model numbers in Bold are DP cards.
https://docs.google.com/spreadsheet/ccc?authkey=COmlndUK&key=0AqCMiOAyW0eBdHdWRDlkZ2xHeUxDQ0E1TVQtVkI3ckE&hl=en&authkey=COmlndUK#gid=0
____________
| |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
(*) Special note about the GTX 560 TI: For reasons that are not well understood, it seems that the GTX 560 TI encounters "maxErr exceeded" errors at lower levels than do other video cards. At this time, there's no solution nor workaround for this problem.
The above should refer to the GTX 550 Ti. The GTX 560 Ti versions do not appear to have this problem. The problem may also exist on certain GTX 460 versions (those with 1024mb of video RAM with the 192-bit bus).
For those interested in a fuller description of the odd memory structure on GTX 550 Ti cards, see here.
____________
141941*2^4299438-1 is prime!
| |
|
|
This is a nice spreadsheet that I refer to quite a bit. It's a good one to bookmark.
The Model numbers in Bold are DP cards.
https://docs.google.com/spreadsheet/ccc?authkey=COmlndUK&key=0AqCMiOAyW0eBdHdWRDlkZ2xHeUxDQ0E1TVQtVkI3ckE&hl=en&authkey=COmlndUK#gid=0
Am I just blind or did I miss seeing the 550 and 560 plain and ti versions?
____________
@AggieThePew
| |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
This is a nice spreadsheet that I refer to quite a bit. It's a good one to bookmark.
The Model numbers in Bold are DP cards.
https://docs.google.com/spreadsheet/ccc?authkey=COmlndUK&key=0AqCMiOAyW0eBdHdWRDlkZ2xHeUxDQ0E1TVQtVkI3ckE&hl=en&authkey=COmlndUK#gid=0
Am I just blind or did I miss seeing the 550 and 560 plain and ti versions?
520-560 models came out after the higher models were released. I would guess that the spreadsheet is dated before the former were released. The spreadsheet is also missing the Quadro CX (which is basically identical to the FX 4800).
____________
141941*2^4299438-1 is prime!
| |
|
|
This is a nice spreadsheet that I refer to quite a bit. It's a good one to bookmark.
The Model numbers in Bold are DP cards.
https://docs.google.com/spreadsheet/ccc?authkey=COmlndUK&key=0AqCMiOAyW0eBdHdWRDlkZ2xHeUxDQ0E1TVQtVkI3ckE&hl=en&authkey=COmlndUK#gid=0
Am I just blind or did I miss seeing the 550 and 560 plain and ti versions?
520-560 models came out after the higher models were released. I would guess that the spreadsheet is dated before the former were released. The spreadsheet is also missing the Quadro CX (which is basically identical to the FX 4800).
ahhh. thanks for the udpate | |
|
|
This is a nice spreadsheet that I refer to quite a bit. It's a good one to bookmark.
The Model numbers in Bold are DP cards.
https://docs.google.com/spreadsheet/ccc?authkey=COmlndUK&key=0AqCMiOAyW0eBdHdWRDlkZ2xHeUxDQ0E1TVQtVkI3ckE&hl=en&authkey=COmlndUK#gid=0
Am I just blind or did I miss seeing the 550 and 560 plain and ti versions?
520-560 models came out after the higher models were released. I would guess that the spreadsheet is dated before the former were released. The spreadsheet is also missing the Quadro CX (which is basically identical to the FX 4800).
ahhh. thanks for the udpate
I guess I need to find out who owns it and have them update it.
____________
| |
|
Michael Goetz Volunteer moderator Project administrator
 Send message
Joined: 21 Jan 10 Posts: 13804 ID: 53948 Credit: 345,369,032 RAC: 4,797
                              
|
I guess I need to find out who owns it and have them update it.
The bigger problem with that spreadsheet is that it just lists "BOINC compatible" cards, which is only useful if all you care about is that the GPU will run on at least one project, but you don't care about *which* project. While it lists single or double precision (albeit not quite correctly), it doesn't list the specific qualities of the cards. Certain projects could require some CC level other than 1.0 or 1.3, which is all the spreadsheet tells you. (I vaguely remember that AP26 required CC 1.1, but I could be wrong.) I could easily imagine apps being written for 2.0 that need Fermi to run. It's trivial to tell which GeForce cards are Fermi, but can you tell from that list which Quadro cards have a Fermi architecture?
If you care about the card running on a particular project, then that spreadsheet doesn't have the necessary information.
The spreadsheet also isn't just missing newer data, it has incorrect data on it. It lists all the 400 series mobile GeForce GPUs as single precision when they're actually double precision.
Handy? Yes. Accurate? Not so much. I'll stick with Nvidia's and AMD's own lists rather than rely on an unknown third party.
____________
My lucky number is 75898524288+1 | |
|
|
Handy? Yes. Accurate? Not so much. I'll stick with Nvidia's and AMD's own lists rather than rely on an unknown third party.
I value your input Michael. Thanks!
____________
| |
|
Tyler Project administrator Volunteer tester Send message
Joined: 4 Dec 12 Posts: 1077 ID: 183129 Credit: 1,365,637,185 RAC: 1
                        
|
Is it worth it to use my Nvidia GT 520 1024mb for primegrid sieving? Or is it not worth the decreased performance and display lag..
____________
275*2^3585539+1 is prime!!! (1079358 digits)
Proud member of Aggie the Pew
| |
|
Michael Goetz Volunteer moderator Project administrator
 Send message
Joined: 21 Jan 10 Posts: 13804 ID: 53948 Credit: 345,369,032 RAC: 4,797
                              
|
Is it worth it to use my Nvidia GT 520 1024mb for primegrid sieving? Or is it not worth the decreased performance and display lag..
Did you mean sieving or running Genefer? I ask because you said 'sieving' but you're asking in a thread relating to compatibility with the GeneferCUDA program.
If you meant Genefer, then you can run Genefer's benchmarks on both the GT 520 and your CPU and then make a decision. The 520 is pretty slow, and especially if you used app_info to use the AVX version of Genefer on your CPU you might find that the CPU can run 4 GFN tasks faster than the GPU can run 4 GFN tasks.
As for sieving, I'll leave that question to others more familiar with the PPS sieve.
____________
My lucky number is 75898524288+1 | |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
Is it worth it to use my Nvidia GT 520 1024mb for primegrid sieving? Or is it not worth the decreased performance and display lag..
As for sieving, I'll leave that question to others more familiar with the PPS sieve.
I think your 520 is a 48 shader card. That would be not too bad on PPS sieve. I'd estimate it at about 12,500 to 15,000 seconds per task depending on your clocks. That should be a bit more than 15,000 to a bit less than 25,000 credits per day if running 24/7.
| |
|
Tyler Project administrator Volunteer tester Send message
Joined: 4 Dec 12 Posts: 1077 ID: 183129 Credit: 1,365,637,185 RAC: 1
                        
|
Is it worth it to use my Nvidia GT 520 1024mb for primegrid sieving? Or is it not worth the decreased performance and display lag..
As for sieving, I'll leave that question to others more familiar with the PPS sieve.
I think your 520 is a 48 shader card. That would be not too bad on PPS sieve. I'd estimate it at about 12,500 to 15,000 seconds per task depending on your clocks. That should be a bit more than 15,000 to a bit less than 25,000 credits per day if running 24/7.
Ok cool. If would get me THAT much credit, I'll run it. I thought that it'd get me around 1 or 2 thousand credit per day.
____________
275*2^3585539+1 is prime!!! (1079358 digits)
Proud member of Aggie the Pew
| |
|
Tyler Project administrator Volunteer tester Send message
Joined: 4 Dec 12 Posts: 1077 ID: 183129 Credit: 1,365,637,185 RAC: 1
                        
|
Is it worth it to use my Nvidia GT 520 1024mb for primegrid sieving? Or is it not worth the decreased performance and display lag..
As for sieving, I'll leave that question to others more familiar with the PPS sieve.
I think your 520 is a 48 shader card. That would be not too bad on PPS sieve. I'd estimate it at about 12,500 to 15,000 seconds per task depending on your clocks. That should be a bit more than 15,000 to a bit less than 25,000 credits per day if running 24/7.
PPS Sieve is taking around 10500 secs each. I calculated it. at 1% In to the project it was 1:45 in. So I did 1.75*100 to get how many minutes then did that number*60 to get the seconds. Did I do that right?
____________
275*2^3585539+1 is prime!!! (1079358 digits)
Proud member of Aggie the Pew
| |
|
Scott Brown Volunteer moderator Project administrator Volunteer tester Project scientist
 Send message
Joined: 17 Oct 05 Posts: 2329 ID: 1178 Credit: 15,582,675,723 RAC: 15,114,994
                                           
|
Seems correct, but on the PPS sieve I'd wait to about 5% or 10% to get a more accurate estimate. Also, note that any significant GPU tasks being done for regular computer use (e.g., gaming, video editing, etc.) can and will slow down the overall time to complete a unit.
| |
|
Tyler Project administrator Volunteer tester Send message
Joined: 4 Dec 12 Posts: 1077 ID: 183129 Credit: 1,365,637,185 RAC: 1
                        
|
Seems correct, but on the PPS sieve I'd wait to about 5% or 10% to get a more accurate estimate. Also, note that any significant GPU tasks being done for regular computer use (e.g., gaming, video editing, etc.) can and will slow down the overall time to complete a unit.
I noticed my SGS LLR Tasks are taking a few minutes longer now.
EDIT: First PPS (Sieve) took 2 hours and 50 minutes. While running 4 Sophie German tasks.
____________
275*2^3585539+1 is prime!!! (1079358 digits)
Proud member of Aggie the Pew
| |
|
Message boards :
Generalized Fermat Prime Search :
List of Compatible Nvidia GPUs |