Are multiple gpu's necessary for CUDA...

WraithX · 2012-03-20, 03:19

Hello everyone,

I've tried looking through this (sub-)forum for the answer, but can't seem to find mention of it (I'm probably just bad at skimming). I've seen it mentioned in a readme file in the CUDA toolkit, but I want to make sure that this is what everyone here is doing.

Is it really necessary to have multiple (nVidia) video cards to run CUDA programs? The docs mention having one for display only, and the other can run CUDA programs. Is this how everyone here is set up? I'm about to get the parts to build a new system, but need to find out if I should get another nVidia video card or not. Does it matter how high, or low, end the display video card is? I'll be using Windows 7 x64 as my OS. Any tips/tricks/words of advice would be greatly appreciated.

flashjh · 2012-03-20, 03:23

Quote:

Originally Posted by WraithX

Hello everyone,

I've tried looking through this (sub-)forum for the answer, but can't seem to find mention of it (I'm probably just bad at skimming). I've seen it mentioned in a readme file in the CUDA toolkit, but I want to make sure that this is what everyone here is doing.

Is it really necessary to have multiple (nVidia) video cards to run CUDA programs? The docs mention having one for display only, and the other can run CUDA programs. Is this how everyone here is set up? I'm about to get the parts to build a new system, but need to find out if I should get another nVidia video card or not. Does it matter how high, or low, end the display video card is? I'll be using Windows 7 x64 as my OS. Any tips/tricks/words of advice would be greatly appreciated.

Nope, just one will work fine. If you push your video card really hard it will slow down your display. Gaming won't work very well (or at all) with CUDA running.

kladner · 2012-03-20, 03:26

Quote:

Originally Posted by WraithX

Hello everyone,

I've tried looking through this (sub-)forum for the answer, but can't seem to find mention of it (I'm probably just bad at skimming). I've seen it mentioned in a readme file in the CUDA toolkit, but I want to make sure that this is what everyone here is doing.

Is it really necessary to have multiple (nVidia) video cards to run CUDA programs? The docs mention having one for display only, and the other can run CUDA programs. Is this how everyone here is set up? I'm about to get the parts to build a new system, but need to find out if I should get another nVidia video card or not. Does it matter how high, or low, end the display video card is? I'll be using Windows 7 x64 as my OS. Any tips/tricks/words of advice would be greatly appreciated.

I am running my display and two instances of mfaktc on a single 460 GTX. I have tried using the onboard AMD GPU for the display. But this removes the possibility of Photoshop using the GPU for acceleration. I have also used that single GPU to run mfaktc as well as CUDALucas. Your Mileage May Vary, but I get along with using the same GPU for multiple purposes.

EDIT: I don't do any high-demand gaming these days.

Dubslow · 2012-03-20, 04:29

Depending on how hard your CUDA prog saturates your GPU, you should be able to get a display and run the prog just fine in tandem. For instance, mfaktc typically saturates my 460 to 85-90%, depending, and this as well as low priority for mfaktc means I can actually run TF2 at 1920x1080, though mfaktc throughput drops to 60%. If I were to run two instances of mfaktc and push my GPU to 100%, then I'm sure my display would get very laggy. You could probably even hit 95-98% usage and still have a working display for everyday tasks (not games).

flashjh · 2012-03-20, 04:33

Quote:

Originally Posted by Dubslow

Depending on how hard your CUDA prog saturates your GPU, you should be able to get a display and run the prog just fine in tandem. For instance, mfaktc typically saturates my 460 to 85-90%, depending, and this as well as low priority for mfaktc means I can actually run TF2 at 1920x1080, though mfaktc throughput drops to 60%. If I were to run two instances of mfaktc and push my GPU to 100%, then I'm sure my display would get very laggy. You could probably even hit 95-98% usage and still have a working display for everyday tasks (not games).

Yes, I run CUDALucas on a 580 which drives 99%. I can use the display for anything day-to-day, just no games.

Can you tell us the video card you're planning to use?

LaurV · 2012-03-20, 06:23

Quote:

Originally Posted by Dubslow

I can actually run TF2 at 1920x1080...

??

Trial Factoring... 2 ??

Dubslow · 2012-03-20, 07:53

Quote:

Originally Posted by LaurV

??

Trial Factoring... 2 ??

Team Fortress 2, a well known/popular game.

https://www.google.com/?q=TF2

WraithX · 2012-03-20, 11:52

Quote:

Originally Posted by flashjh

Yes, I run CUDALucas on a 580 which drives 99%. I can use the display for anything day-to-day, just no games.

Can you tell us the video card you're planning to use?

I'll be getting a 3GB GTX 580, but I won't be playing any games on it.

Do you think that it works with one video card because of how CL,mfaktc,etc actually use the card? Is the maximum run time mentioned below related to how long a kernel runs? Maybe CL,mfaktc,etc all launch a large number of short lived kernels, thus avoiding the windows "watchdog" timeout? Actually, I plan to use this for gpu_ecm if I can get that working in Win7 x64.

Is there a driver version or CUDA toolkit version that is recommended? Should I get the latest drivers or one of the older ones? Should I stick with CUDA toolkit 4.1.28, or get one of the older ones? Is it easy to switch between different versions? I guess for the toolkit, I can just put them in different directories and change my path variable. But I don't think it'll be that easy with the drivers.

nVidia CUDA FAQ from:
http://developer.nvidia.com/cuda-faq

Quote:

Q: What is the maximum kernel execution time?
On Windows, individual GPU program launches have a maximum run time of around 5 seconds. Exceeding this time limit usually will cause a launch failure reported through the CUDA driver or the CUDA runtime, but in some cases can hang the entire machine, requiring a hard reset.

This is caused by the Windows "watchdog" timer that causes programs using the primary graphics adapter to time out if they run longer than the maximum allowed time.

For this reason it is recommended that CUDA is run on a GPU that is NOT attached to a display and does not have the Windows desktop extended onto it. In this case, the system must contain at least one NVIDIA GPU that serves as the primary graphics adapter.

LaurV · 2012-03-20, 14:21

That is a good choice and it worth the money, except when you buy a Mars or some other "limited edition" thingies where you pay more for the brand then for the goods. If you use it as primary display, when is "maxed" (like CudaLucas with -aggressive and no -t switch, GPU load 99%) then your screen will move "in steps" and there is no way to watch any video. Even editing become a bit difficult, you type 2-3-5 keys and nothing appears on screen, then after the 3rd, 5th key, everything appears at once. If you use CudaLucas in polite mode (GPU load about 75-78% with -t, and about 83% without -t, if you don't know, -t is checking the errors at every iteration, this is "pausing" the big FFT calculus, giving some "breaks" to GPU, which can do other things in this time), then you can play videos, do whatever you like, except playing "fold-it". I don't know about graphic-intensive games, I better like playing gomoku :D, but AutoCad, Protel, SolidWorks, etc, run smooth and fast, like no CudaLucas (almost). In between as performance, when you use CudaLucas with -aggressive AND -t, the GPU load is between 89-92%, you still can not play "foldi-it", and some "heavy-model" Autocad will move in steps, but for daily office work there is no slowdown visible. Mfaktc does not count too much, even with 3 or 4 copies in the same time, you don't feel it too much, as mfaktc is usually bottleneck-ed by the CPU, not by GPU. You can not maximize gtx580 with mfaktc, unless you use 3 or 4 CPU cores for it, then of course your "fold-it" (or related) won't move smooth.

bcp19 · 2012-03-20, 21:20

Quote:

Originally Posted by WraithX

Hello everyone,

I've tried looking through this (sub-)forum for the answer, but can't seem to find mention of it (I'm probably just bad at skimming). I've seen it mentioned in a readme file in the CUDA toolkit, but I want to make sure that this is what everyone here is doing.

Is it really necessary to have multiple (nVidia) video cards to run CUDA programs? The docs mention having one for display only, and the other can run CUDA programs. Is this how everyone here is set up? I'm about to get the parts to build a new system, but need to find out if I should get another nVidia video card or not. Does it matter how high, or low, end the display video card is? I'll be using Windows 7 x64 as my OS. Any tips/tricks/words of advice would be greatly appreciated.

I have 1 system that has 2 GPUs in it and 5 with 1. Being a fairly avid gamer, I tried using the 2 GPU system to play Star Wars:TOR, which looked awesome on the highest setting, but the 2nd GPU would lose around 1/2 of it's computing capability when I left it running mfaktc. Being retired and spending more hours than I probably should playing, I ended up switching my gaming to one of the single GPU systems, since it was less of a hit on output. So my recommendation would be to use a single video card unless you plan to use both for computing.

nucleon · 2012-03-21, 11:10

I play League of Legends on my PC with 4x instances of mfaktc running. 60fps during the game - no trouble. 2x GTX560Ti. Desktop extended to both cards.

Obviously mfaktc slows down a bit. But no troubles.

-- Craig

2012-03-20, 03:19	#1
WraithX Mar 2006 2³×59 Posts	Are multiple gpu's necessary for CUDA... Hello everyone, I've tried looking through this (sub-)forum for the answer, but can't seem to find mention of it (I'm probably just bad at skimming). I've seen it mentioned in a readme file in the CUDA toolkit, but I want to make sure that this is what everyone here is doing. Is it really necessary to have multiple (nVidia) video cards to run CUDA programs? The docs mention having one for display only, and the other can run CUDA programs. Is this how everyone here is set up? I'm about to get the parts to build a new system, but need to find out if I should get another nVidia video card or not. Does it matter how high, or low, end the display video card is? I'll be using Windows 7 x64 as my OS. Any tips/tricks/words of advice would be greatly appreciated.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Using multiple PCs	numbercruncher	Information & Answers	18	2014-04-17 00:17
Multiple quotes	mfgoode	Forum Feedback	11	2005-10-16 00:30
Multiple Installations.	A_Pryme	Software	35	2005-02-16 05:51
Multiple systems/multiple CPUs. Best configuration?	BillW	Software	1	2003-01-21 20:11
Multiple OS'es on one box w/ Prime95	penguin22	Software	4	2002-10-05 16:10

2012-03-20, 04:29	#4
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Depending on how hard your CUDA prog saturates your GPU, you should be able to get a display and run the prog just fine in tandem. For instance, mfaktc typically saturates my 460 to 85-90%, depending, and this as well as low priority for mfaktc means I can actually run TF2 at 1920x1080, though mfaktc throughput drops to 60%. If I were to run two instances of mfaktc and push my GPU to 100%, then I'm sure my display would get very laggy. You could probably even hit 95-98% usage and still have a working display for everyday tasks (not games).

2012-03-20, 14:21	#9
LaurV Romulan Interpreter Jun 2011 Thailand 2·19·241 Posts	That is a good choice and it worth the money, except when you buy a Mars or some other "limited edition" thingies where you pay more for the brand then for the goods. If you use it as primary display, when is "maxed" (like CudaLucas with -aggressive and no -t switch, GPU load 99%) then your screen will move "in steps" and there is no way to watch any video. Even editing become a bit difficult, you type 2-3-5 keys and nothing appears on screen, then after the 3rd, 5th key, everything appears at once. If you use CudaLucas in polite mode (GPU load about 75-78% with -t, and about 83% without -t, if you don't know, -t is checking the errors at every iteration, this is "pausing" the big FFT calculus, giving some "breaks" to GPU, which can do other things in this time), then you can play videos, do whatever you like, except playing "fold-it". I don't know about graphic-intensive games, I better like playing gomoku :D, but AutoCad, Protel, SolidWorks, etc, run smooth and fast, like no CudaLucas (almost). In between as performance, when you use CudaLucas with -aggressive AND -t, the GPU load is between 89-92%, you still can not play "foldi-it", and some "heavy-model" Autocad will move in steps, but for daily office work there is no slowdown visible. Mfaktc does not count too much, even with 3 or 4 copies in the same time, you don't feel it too much, as mfaktc is usually bottleneck-ed by the CPU, not by GPU. You can not maximize gtx580 with mfaktc, unless you use 3 or 4 CPU cores for it, then of course your "fold-it" (or related) won't move smooth.

2012-03-21, 11:10	#11
nucleon Mar 2003 Melbourne 5×103 Posts	I play League of Legends on my PC with 4x instances of mfaktc running. 60fps during the game - no trouble. 2x GTX560Ti. Desktop extended to both cards. Obviously mfaktc slows down a bit. But no troubles. -- Craig