Welcome to MilkyWay@home

GPU Requirements [OLD]


Advanced search

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

To post messages, you must log in.

Previous · 1 . . . 14 · 15 · 16 · 17 · 18 · Next

AuthorMessage
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
50 million credit badge13 year member badge
Message 49099 - Posted: 29 May 2011, 16:06:11 UTC - in response to Message 49098.  
Last modified: 29 May 2011, 16:07:49 UTC


Can anyone tell me whether an NV GTX460 is good for double precision. Most of my cards are ATI.


I use one GTX460, it works fine for MW, but needs > 12min / wu.



Such an NV card would give me flexibility on ATI biased GPU projects (DNETC, Milkyway, Collatz) but shine at PrimeGrid as well. I can them move them to new projects or combinations of projects (including SETI and Einstein) as I feel like.

I can then dedicate the quad CPUs to supporting the GPUs and CPU only projects like Malaria, FreeHAL and WUProp
Go away, I was asleep


ID: 49099 · Rating: 0 · rate: Rate + / Rate - Report as offensive
ProfileWerkstatt

Send message
Joined: 19 Feb 08
Posts: 350
Credit: 139,344,691
RAC: 93
100 million credit badge13 year member badge
Message 49105 - Posted: 29 May 2011, 19:26:19 UTC - in response to Message 49099.  


Such an NV card would give me flexibility on ATI biased GPU projects (DNETC, Milkyway, Collatz) but shine at PrimeGrid as well. I can them move them to new projects or combinations of projects (including SETI and Einstein) as I feel like.

I can then dedicate the quad CPUs to supporting the GPUs and CPU only projects like Malaria, FreeHAL and WUProp


The card crunches for me at:
Seti, Einstein, MW, Collatz and Gpugrid. And I hope, in the future dna also.
@ Seti and Einstein you should run at least two wu's at the same time to get an acceptable GPU-usage. Works perfect on 1024MB-Cards.
ID: 49105 · Rating: 0 · rate: Rate + / Rate - Report as offensive
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
50 million credit badge13 year member badge
Message 49106 - Posted: 29 May 2011, 20:47:00 UTC
Last modified: 29 May 2011, 20:47:45 UTC

I will try to remember to get the 1GB RAM version and see how it goes withh the right app_info file
Go away, I was asleep


ID: 49106 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Traills

Send message
Joined: 16 Aug 10
Posts: 6
Credit: 3,885,804
RAC: 0
3 million credit badge11 year member badge
Message 49107 - Posted: 29 May 2011, 22:04:32 UTC

I set up a new machine this week and started MW on it. The WUs that have run on the CPUs have been fine, but the one that has started on the GPU looks to be in trouble. Before I abort it, I want to ask whether there is something that can be done to save the other 11 cuda WUs that are queued.

Host machine is an i7 with 8 CPUs running Windows 7 SP1. Graphics card is an OEM nVidia GeForce GT 420 with driver 267.24, cuda driver is 3.2.1. The WU having problems is 29477371, and on my machine the MW app is 0.52 cuda_opencl.

Trouble symptom: The WU has just passed 3 hrs run time and shows 0% Progress in the BOINC Manager. (Is the progress meter supposed to update when a GPU job is running?) It and 3 other of the queued cuda WUs downloaded with an expected time of 2:00:49. On the other hand,someone else's machine running this WU with XP and an ATI card finished in 238 sec (11 CPU sec).

Thank you
ID: 49107 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Link
Avatar

Send message
Joined: 19 Jul 10
Posts: 357
Credit: 16,324,081
RAC: 0
10 million credit badge11 year member badge
Message 49108 - Posted: 29 May 2011, 22:22:08 UTC - in response to Message 49097.  

It would be nice to have these conveniently labelled(graphics card list) as to which ones are available with AGP or not. My roommate uses BOINC but has an older Motherboared with AGP and an AMD 3100+ in it and would be interested to know which were available for what style without hunting the internet. We could do that but it would be more convenient labelled for those in the future wondering the same thing.

He might want to seriously consider an inexpensive MB that supports PCIe instead of trying to cobble a long extinct AGP GPU into the present one. The AGP cards that are available are several generations old, very slow, relatively expensive and use way more power for their output compared to the current offerings.


Unfortunately there are no socket A boards with PCIe (was looking once for that myself), so he would also need new CPU and RAM. And since new motherboards generate the voltage for CPU from 12V and not 5V like socket A boards did, it's more likely that he would also need new PSU.
.
ID: 49108 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Sunny129
Avatar

Send message
Joined: 25 Jan 11
Posts: 271
Credit: 346,072,284
RAC: 0
300 million credit badge11 year member badgeextraordinary contributions badge
Message 49110 - Posted: 30 May 2011, 2:13:40 UTC - in response to Message 49107.  

Trouble symptom: The WU has just passed 3 hrs run time and shows 0% Progress in the BOINC Manager. (Is the progress meter supposed to update when a GPU job is running?) It and 3 other of the queued cuda WUs downloaded with an expected time of 2:00:49. On the other hand,someone else's machine running this WU with XP and an ATI card finished in 238 sec (11 CPU sec).

hmm...i wish i could do more to address your problem, but my knowledge with that kind of troubleshooting is limited. i'm sure some of the volunteer developers and testers will have some suggestions.

i will say this though - i wouldn't read into the fact that your wingman's ATI GPU will be able to crunch through this task much faster than your nVidia GPU...that's just the nature of the MW@H project - it just works better on ATI hardware architecture for the time being. while i'm not familiar with where the GeForce GT 420 slots into nVidia's GPU evolution, it does sound like a somewhat dated GPU in comparison to nVidia's most recent generations of crunchers, the GTX4xx and GTX5xx series...and even these nVidia GPUs don't hold a candle to ATI's HD69xx, HD58xx, HD48xx, HD47xx, and perhaps even some of the HD3xxx series in MW@H crunching performance. so i wouldn't expect an nVidia GPU that is several (or even just a few) generations old to be a 10th as fast as any current ATI GPU at MW@H.

to give you an example, over on the AnandTech distributed computing forum we TeAm AnandTech members have been compiling a list of the run times of various GPUs specifically for the de_separation tasks that earn ~213 credits. it turns out that while an ATI HD 5870 GPU (ATI's previous generation flagship GPU) can complete such a task in approx. 1.5 minutes, it takes the nVidia GTX 480 (nVidia's previous generation flagship GPU) approx. 6 minutes to complete the very same kind of task. now we unfortunately don't have any run times for your GPU in our database, nor do we have any for an ATI HD5830, the slowest of the HD 58xx series GPUs. but we do have the HD 5850 showing run times of approx. 2 minutes. its also unfortunate that the MW@H server doesn't glean more accurate information on a host's hardware, b/c all we can tell by looking at that information is that your wingman is using two HD 5800 1GB series GPUs - he could be running something as slow as a 5830, or he have a 5870 running 2 tasks simultaneously...its impossible to tell. the point though, is that if ATI's slowest previous generation single-GPU HD series video card is ~3 times as fast as nVidia's fastest previous generation single-GPU video card at MW@H, then its bound to take quite a bit longer than that on your GPU.

just how much slower, i don't know...but your GeForce GT 420 truly is supposed to be taking roughly 2 hours per task, then a MW@H task would take on the order of 60 times as long to complete as it would on say, ATI's HD 5850 GPU. hopefully someone can help you with your troubles soon so you can see for sure how long it should take your GPU to crunch a MW@H task.
ID: 49110 · Rating: 0 · rate: Rate + / Rate - Report as offensive
robertmiles

Send message
Joined: 30 Sep 09
Posts: 211
Credit: 34,251,373
RAC: 9,632
30 million credit badge12 year member badgeextraordinary contributions badge
Message 49111 - Posted: 30 May 2011, 2:26:03 UTC

I know a little about the GT 420. Probably actually more recent than the GTX 400 series, but still about the lowest-end member of the 400 generation. Expect it to be rather slow for crunching. Fermi-based, so it's able to do Milkyway@Home calculations.

http://www.geeks3d.com/20100903/nvidia-geforce-gt-420-specifications/
ID: 49111 · Rating: 0 · rate: Rate + / Rate - Report as offensive
robertmiles

Send message
Joined: 30 Sep 09
Posts: 211
Credit: 34,251,373
RAC: 9,632
30 million credit badge12 year member badgeextraordinary contributions badge
Message 49112 - Posted: 30 May 2011, 3:01:46 UTC
Last modified: 30 May 2011, 3:09:18 UTC

I just looked up the GT 420 specifications and the specifications of my GTS 450.

http://www.nvidia.com/object/product-geforce-gt-420-oem-us.html

http://www.nvidia.com/object/product-geforce-gts-450-us.html

The most important things to compare to estimate the relative speeds are the clock rates and the number of CUDA cores.

Roughly the same clock rates.

Only 48 CUDA cores for the GT 420 compared to 192 for the GTS 450.

Therefore, expect the GT 420 to take roughly 4 times as long per workunit as the 15 to 20 minutes it takes my GTS 450, if nothing else interferes.
ID: 49112 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Sunny129
Avatar

Send message
Joined: 25 Jan 11
Posts: 271
Credit: 346,072,284
RAC: 0
300 million credit badge11 year member badgeextraordinary contributions badge
Message 49113 - Posted: 30 May 2011, 3:43:42 UTC - in response to Message 49112.  

The most important things to compare to estimate the relative speeds are the clock rates and the number of CUDA cores.

excellent point. i almost forgot the basic rule of thumb - all else being equal, one can make rough comparisons based on (shader clock)*(shader count). and thanks for finding that info on the GT 420 Robert.

with that in mind, we can go back and consider the GT4 420 and the GTX 480 i mentioned above. both have a core clock of 700mhz, thus both have the same shader clock (twice that of the core clock on Fermi GPUs). however, the GTX 480 has 10 times as many cores/shaders (480 vs 48). thus, the GTX 480 will be ~10 times as fast as your GT 420, and an ATI HD 5850 GPU for instance will be ~30 times as fast as your GT 420. so if your wingman's 5800 series GPU happens to be a 5850 running 1 task at a time, and it takes him ~238s (~4 minutes) to complete a task, then a very general estimate of the amount of time it should take your GT 420 approx. to complete a task is ~120 minutes, or ~2 hours, just as your BOINC manager predicted.

but remember, this a very general and very crude estimate. there are other factors that come into play when comparing ATI run times to nVidia run times. first of all, MW@H requires a double precision floating point (FP64) capable GPU. ATI's HD 58xx series' FP64 performance is limited to 1/5 of its single precision performance. ATI's HD 69xx series' FP64 performance is limited to 1/4 of its single precision performance. nVidia's Fermi architecture both the GTX 4xx and GTX 5xx series, and probably some earlier ones as well) limits FP64 performance to 1/8 that of its single precision performance. also, something to consider is that ATI's Cypress and Cayman architectures (the HD 58xx and HD 69xx series) could perform 2 FP64 operations per shader clock. nVidia's Fermi CUDA architecture is different in that some of its shaders perform the same as ATI's shaders, and some perform better. that is to say, while some of a Fermi GPU's shaders can perform 2 FP64 operations per shader clock, other shaders on the GPU can perform up to 4 FP64 operations per shader clock. so you see its very difficult to get an apples vs apples comparison with an ATI GPU and an nVidia GPU.
ID: 49113 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Traills

Send message
Joined: 16 Aug 10
Posts: 6
Credit: 3,885,804
RAC: 0
3 million credit badge11 year member badge
Message 49114 - Posted: 30 May 2011, 5:28:04 UTC

Thanks, Sunny & Robert for your comments. I figured anything OEM was going to be relatively slow. Now though it is up to 10 hrs elapsed and still showing 0% Progress so I have to think something is amiss besides the GT 420 using an old GPU design. I found something in an earlier thread about some conditions resulting in the gpu clock rate slowing way down under Win7 when a WU is restarted after a power-saver interruption, and then not recovering (message pasted below). I am speculating that that may be what is going on here. Being new to Win7 I didn't expect it to shut down running processes just because the keyboard was inactive overnight, but of course it did, so the MW cuda WU was put to sleep not long after I started it last night. Today I discovered that problem and reset Win7 not to shut anything off period but if that is what happened to this WU then the "damage" has been done. As a test I have suspended the WU and started another to see if it behaves closer to expectations (~2 hrs). If it ends up with the same problem I will probably have to disable MW use of the GPU. Too bad, but I didn't spec the machine mainly for Boinc; the GPU would be nice if it will behave but if it doesn't I'll be content with all the CPUs.

(Message 48986)
Posted 4 days ago by Tex1954
--------------------------------------------------------------------------------
There is an ongoing problem with CUDA tasks and the Clock Rates being dropped in Nvidia cards I've written it up and the problem is with Vista and Win7 both.

What happens, is the clock rate gets dropped to conserve power/heat etc. and never returns to high speed again. This always happens with DUAL Nvidia cards installed and seems only magic prevents it from happening on it's own most of the time. Doesn't matter what power settings are set, performance mode seems to help, but not totally correct it. Snoozing or Suspending tasks is a 95% guarantee the clocks with drop and never regain full speed again.
ID: 49114 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Chris S
Avatar

Send message
Joined: 20 Sep 08
Posts: 1391
Credit: 203,539,424
RAC: 9,943
200 million credit badge13 year member badge
Message 49120 - Posted: 30 May 2011, 12:20:38 UTC

I will try to remember to get the 1GB RAM version and see how it goes withh the right app_info file


John, 1 GB cards run a lot faster, a lot cooler, and are more over-clockable than 512mb cards of the same family, on any Boinc project you care to name. I will never buy a 512mb card again.
Don't drink water, that's the stuff that rusts pipes
ID: 49120 · Rating: 0 · rate: Rate + / Rate - Report as offensive
ProfileBeyond
Avatar

Send message
Joined: 15 Jul 08
Posts: 383
Credit: 510,838,469
RAC: 248,168
500 million credit badge13 year member badge
Message 49126 - Posted: 30 May 2011, 20:30:51 UTC - in response to Message 49108.  

It would be nice to have these conveniently labelled(graphics card list) as to which ones are available with AGP or not. My roommate uses BOINC but has an older Motherboared with AGP and an AMD 3100+ in it and would be interested to know which were available for what style without hunting the internet. We could do that but it would be more convenient labelled for those in the future wondering the same thing.

He might want to seriously consider an inexpensive MB that supports PCIe instead of trying to cobble a long extinct AGP GPU into the present one. The AGP cards that are available are several generations old, very slow, relatively expensive and use way more power for their output compared to the current offerings.

Unfortunately there are no socket A boards with PCIe (was looking once for that myself), so he would also need new CPU and RAM. And since new motherboards generate the voltage for CPU from 12V and not 5V like socket A boards did, it's more likely that he would also need new PSU.

First of all the 3100+ is not socket a, it's socket 754. But even buying a Phenom 9750 with PCIe MB for $99, add 2GB ram for $35, and a used HD4770 for $50 is going to be way faster and probably use less power than trying to find an AGP HD3870 at a non-silly rice, which will still be comparatively extremely slow and completely obsolete. Might make sense if you already have it, but if you're going to spend $$, not...
ID: 49126 · Rating: 0 · rate: Rate + / Rate - Report as offensive
ProfileBeyond
Avatar

Send message
Joined: 15 Jul 08
Posts: 383
Credit: 510,838,469
RAC: 248,168
500 million credit badge13 year member badge
Message 49127 - Posted: 30 May 2011, 20:42:00 UTC - in response to Message 49120.  

I will try to remember to get the 1GB RAM version and see how it goes withh the right app_info file

John, 1 GB cards run a lot faster, a lot cooler, and are more over-clockable than 512mb cards of the same family, on any Boinc project you care to name. I will never buy a 512mb card again.

If you're referring to the GTX 460, the 768MB version is not much slower than the 1GB version at least at PG and arguably more efficient for the power used. The 1GB is somewhat faster at GPUGrid. The 768MB card uses LESS power than the 1GB. Just stay away from the GTX 460 SE, which is seriously crippled. For MW, ATI/AMD is a far more efficient choice. I have 4 GTX 460 cards and they run very well at PG, about the same as an HD 5770 at Collatz and not so well at all here compared to the competition.
ID: 49127 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Traills

Send message
Joined: 16 Aug 10
Posts: 6
Credit: 3,885,804
RAC: 0
3 million credit badge11 year member badge
Message 49130 - Posted: 31 May 2011, 1:27:41 UTC - in response to Message 49114.  

Post mortem on the NVidia GT 420 ... second WU behaved same as first, running "forever" with 0% progress ... I then found / downloaded / installed a more recent driver v 270.61 after which all the MW cuda WUs failed on computational error within seconds. Now guessing the same thing happened with the old driver but that driver did not send the expected signal and BAM! thought the WUs were still running. I have with regret disabled GPU usage for MW. Will stay active with CPU jobs.

Meanwhile, the GPU does run SETI cuda fermi WUs correctly so the card isn't entirely pathological...

ID: 49130 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Sunny129
Avatar

Send message
Joined: 25 Jan 11
Posts: 271
Credit: 346,072,284
RAC: 0
300 million credit badge11 year member badgeextraordinary contributions badge
Message 49131 - Posted: 31 May 2011, 2:33:18 UTC - in response to Message 49130.  
Last modified: 31 May 2011, 2:36:42 UTC

Meanwhile, the GPU does run SETI cuda fermi WUs correctly so the card isn't entirely pathological...

i'm glad you could find an alternative project to crunch so that your GT 420 didn't go to waste. on that note, its a known fact that nVidia's Fermi architecture is more efficient with S@H Multibeam tasks than ATI's current architectures. out of curiosity, what is an example Multibeam 6.10 task run time on your GT 420? and if possible, could you provide that AR (angle range) of the task?

thanks,
Eric
ID: 49131 · Rating: 0 · rate: Rate + / Rate - Report as offensive
ProfileWerkstatt

Send message
Joined: 19 Feb 08
Posts: 350
Credit: 139,344,691
RAC: 93
100 million credit badge13 year member badge
Message 49149 - Posted: 6 Jun 2011, 17:00:01 UTC - in response to Message 49127.  

I will try to remember to get the 1GB RAM version and see how it goes withh the right app_info file

John, 1 GB cards run a lot faster, a lot cooler, and are more over-clockable than 512mb cards of the same family, on any Boinc project you care to name. I will never buy a 512mb card again.

If you're referring to the GTX 460, the 768MB version is not much slower than the 1GB version at least at PG and arguably more efficient for the power used. The 1GB is somewhat faster at GPUGrid. The 768MB card uses LESS power than the 1GB. Just stay away from the GTX 460 SE, which is seriously crippled. For MW, ATI/AMD is a far more efficient choice. I have 4 GTX 460 cards and they run very well at PG, about the same as an HD 5770 at Collatz and not so well at all here compared to the competition.

The old version of Afterburner gave numbers about memory usage. I've seen that most project need 200 - 300 MB's of ram, so where is the advantage of using a card with more ram?
I use a GTX460 with 768MB, I can run 3 Seti or 2 Einstein WU's in parallel without problems, that increases GPU-usage above 90%. This works also for GPUGrid; 2 parallel run ~15% faster than 2 wu's one after another.
Is there a technical explanation for 'Being faster with more RAM' ?
ID: 49149 · Rating: 0 · rate: Rate + / Rate - Report as offensive
ProfileBeyond
Avatar

Send message
Joined: 15 Jul 08
Posts: 383
Credit: 510,838,469
RAC: 248,168
500 million credit badge13 year member badge
Message 49153 - Posted: 6 Jun 2011, 17:18:44 UTC - in response to Message 49149.  

If you're referring to the GTX 460, the 768MB version is not much slower than the 1GB version at least at PG and arguably more efficient for the power used. The 1GB is somewhat faster at GPUGrid. The 768MB card uses LESS power than the 1GB. Just stay away from the GTX 460 SE, which is seriously crippled. For MW, ATI/AMD is a far more efficient choice. I have 4 GTX 460 cards and they run very well at PG, about the same as an HD 5770 at Collatz and not so well at all here compared to the competition.

The old version of Afterburner gave numbers about memory usage. I've seen that most project need 200 - 300 MB's of ram, so where is the advantage of using a card with more ram? I use a GTX460 with 768MB, I can run 3 Seti or 2 Einstein WU's in parallel without problems, that increases GPU-usage above 90%. This works also for GPUGrid; 2 parallel run ~15% faster than 2 wu's one after another.

Is there a technical explanation for 'Being faster with more RAM' ?

I certainly didn't say that, but here's more information.The 1GB non-SE GTX 460 has a wider memory interface which can make a difference in a few projects, but probably not in most. For my money though the 768MB cards offer the most bang for the buck in most projects. The 1GB SE model is crippled with only 288 cores.
ID: 49153 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Sunny129
Avatar

Send message
Joined: 25 Jan 11
Posts: 271
Credit: 346,072,284
RAC: 0
300 million credit badge11 year member badgeextraordinary contributions badge
Message 49179 - Posted: 8 Jun 2011, 2:04:49 UTC - in response to Message 49149.  

I will try to remember to get the 1GB RAM version and see how it goes withh the right app_info file

John, 1 GB cards run a lot faster, a lot cooler, and are more over-clockable than 512mb cards of the same family, on any Boinc project you care to name. I will never buy a 512mb card again.

If you're referring to the GTX 460, the 768MB version is not much slower than the 1GB version at least at PG and arguably more efficient for the power used. The 1GB is somewhat faster at GPUGrid. The 768MB card uses LESS power than the 1GB. Just stay away from the GTX 460 SE, which is seriously crippled. For MW, ATI/AMD is a far more efficient choice. I have 4 GTX 460 cards and they run very well at PG, about the same as an HD 5770 at Collatz and not so well at all here compared to the competition.

The old version of Afterburner gave numbers about memory usage. I've seen that most project need 200 - 300 MB's of ram, so where is the advantage of using a card with more ram?
I use a GTX460 with 768MB, I can run 3 Seti or 2 Einstein WU's in parallel without problems, that increases GPU-usage above 90%. This works also for GPUGrid; 2 parallel run ~15% faster than 2 wu's one after another.
Is there a technical explanation for 'Being faster with more RAM' ?

i guess the technical explanation would go something like this:

let's assume for example that we have a GTX 460 768MB GPU and a GTX 460 1GB GPU. both GPUs have the same number of shaders, the same memory bus bandwidth, and the same core & memory clock frequencies - the only difference is the amount of memory. now suppose that GPU usage sits at 25,% 50%, 75%, and 100% respectively when running 1, 2, 3, and 4 tasks of a particular project simultaneously. now suppose that GPU memory consumption is 250MB per task. while both GTX 460's can theoretically run 4 simultaneous tasks before maxing out the core/shaders, only the 1GB GPU has the memory capacity to handle 4 simultaneous tasks in reality without being bottlenecked. having 768MB of GPU memory for a quantity of simultaneous tasks that require a total of 1GB of GPU memory is quite the bottleneck, and would cause the GTX 460 768MB GPU to be slower than the GTX 460 1GB GPU.

now, would the 768MB GPU be any slower than the 1GB card while running only two or even three simultaneous tasks of the same hypothetical project? no, b/c both cards have the core/shader capacity and the memory capacity to handle 2 or 3 simultaneous tasks. are all DC projects' tasks going to consume the same quantities of core/shader usage and memory? of course not. so it'll depend on a project's core/shader and memory consumption per task as to whether the GTX 460 1GB GPU will be faster than the GTX 460 768MB GPU.
ID: 49179 · Rating: 0 · rate: Rate + / Rate - Report as offensive
ProfileBeyond
Avatar

Send message
Joined: 15 Jul 08
Posts: 383
Credit: 510,838,469
RAC: 248,168
500 million credit badge13 year member badge
Message 49197 - Posted: 10 Jun 2011, 19:54:41 UTC - in response to Message 49179.  

If you're referring to the GTX 460, the 768MB version is not much slower than the 1GB version at least at PG and arguably more efficient for the power used. The 1GB is somewhat faster at GPUGrid. The 768MB card uses LESS power than the 1GB. Just stay away from the GTX 460 SE, which is seriously crippled. For MW, ATI/AMD is a far more efficient choice. I have 4 GTX 460 cards and they run very well at PG, about the same as an HD 5770 at Collatz and not so well at all here compared to the competition.

i guess the technical explanation would go something like this:

let's assume for example that we have a GTX 460 768MB GPU and a GTX 460 1GB GPU. both GPUs have the same number of shaders, the same memory bus bandwidth, and the same core & memory clock frequencies - the only difference is the amount of memory. now suppose that GPU usage sits at 25,% 50%, 75%, and 100% respectively when running 1, 2, 3, and 4 tasks of a particular project simultaneously. now suppose that GPU memory consumption is 250MB per task. while both GTX 460's can theoretically run 4 simultaneous tasks before maxing out the core/shaders, only the 1GB GPU has the memory capacity to handle 4 simultaneous tasks in reality without being bottlenecked. having 768MB of GPU memory for a quantity of simultaneous tasks that require a total of 1GB of GPU memory is quite the bottleneck, and would cause the GTX 460 768MB GPU to be slower than the GTX 460 1GB GPU.

now, would the 768MB GPU be any slower than the 1GB card while running only two or even three simultaneous tasks of the same hypothetical project? no, b/c both cards have the core/shader capacity and the memory capacity to handle 2 or 3 simultaneous tasks. are all DC projects' tasks going to consume the same quantities of core/shader usage and memory? of course not. so it'll depend on a project's core/shader and memory consumption per task as to whether the GTX 460 1GB GPU will be faster than the GTX 460 768MB GPU.

The 768MB model has a 192-bit memory interface and the 1GB model is 256-bit. That's why the difference in speed, but only in some applications.
ID: 49197 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Sunny129
Avatar

Send message
Joined: 25 Jan 11
Posts: 271
Credit: 346,072,284
RAC: 0
300 million credit badge11 year member badgeextraordinary contributions badge
Message 49198 - Posted: 10 Jun 2011, 22:22:59 UTC - in response to Message 49197.  
Last modified: 10 Jun 2011, 22:25:36 UTC

The 768MB model has a 192-bit memory interface and the 1GB model is 256-bit. That's why the difference in speed, but only in some applications.

i stand corrected on the memory bus rates, and i can see how that difference alone can make the 1GB card faster in some apps. but as i pointed out above, the quantity of memory can also come into play in certain instances/apps, and make the 1GB card the faster card...especially if we're comparing apples-to-apples and forcing both cards to run the same number of simultaneous tasks (i.e. force the 768MB card to deal with a number of simultaneous tasks whose memory requirements can potentially fall between 768MB and 1GB).
ID: 49198 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Previous · 1 . . . 14 · 15 · 16 · 17 · 18 · Next

Message boards : Number crunching : GPU Requirements [OLD]

©2022 Astroinformatics Group