4-Way GPU computing


log in

Advanced search

Questions and Answers : GPU applications : 4-Way GPU computing

Author Message
danieloconnor
Volunteer tester
Send message
Joined: 16 Nov 05
Posts: 17
Credit: 1,871,252
RAC: 4,198
Canada
Message 1491728 - Posted: 20 Mar 2014, 7:35:14 UTC

Hello,

I am considering purchasing a motherboard that supports 4-way SLI; the Asus Z87-WS. At the moment, I do not own any 4-way SLI graphics cards and would like to know if I could use 4 non-SLI video cards on the board to processes work-units via CUDA. Or, will I need to use SLI to crunch for SETI with this many graphics cards?
____________

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2879
Credit: 6,455,861
RAC: 2,769
Bulgaria
Message 1491797 - Posted: 20 Mar 2014, 11:16:06 UTC - in response to Message 1491728.

SLI doesn't matter for BOINC / SETI@home

In fact in the past (a few years ago, old NVIDIA drivers) it was necessary to disable SLI for BOINC to detect properly the number of GPUs
With current NVIDIA drivers BOINC works with SLI enabled/disabled

(I'm not sure if the speed of CUDA is lower with SLI enabled, I think the video RAM usage may be higher with SLI)
____________



- ALF - "Find out what you don't do well ..... then don't do it!" :)

danieloconnor
Volunteer tester
Send message
Joined: 16 Nov 05
Posts: 17
Credit: 1,871,252
RAC: 4,198
Canada
Message 1491891 - Posted: 20 Mar 2014, 16:29:02 UTC - in response to Message 1491797.

Great, thank you!

So basically I can plug in as many GPUs and my system will hold and Boinc will use them? Or, is there a limit to how many Boinc will utilize?
____________

Profile ZalsterProject donor
Volunteer tester
Avatar
Send message
Joined: 27 May 99
Posts: 949
Credit: 41,718,371
RAC: 69,657
United States
Message 1492103 - Posted: 20 Mar 2014, 19:12:19 UTC - in response to Message 1491891.
Last modified: 20 Mar 2014, 20:09:07 UTC

Someone did the math, it was like 42 GPUS or something like that (something to do with how bonic's code is written)...anyway, it's what ever your system is able to hold. Depending on number of Cores on your Chip and power requirements of your MB and what your PSU can give. Most I've seen is 8 GPUs on a system with 12 cores. At some point there is diminishing returns.

danieloconnor
Volunteer tester
Send message
Joined: 16 Nov 05
Posts: 17
Credit: 1,871,252
RAC: 4,198
Canada
Message 1492596 - Posted: 21 Mar 2014, 8:41:12 UTC - in response to Message 1492103.
Last modified: 21 Mar 2014, 8:42:32 UTC

Wow, 42 is a lot of GPUs. I only have room for 4 and I'm sure that will do a lot more crunching than my laptop with 2 CPU cores. Do you know much about bandwidth needed for GPU functionality? I know the CPU communicates to the GPUs but I'm not sure how often and how much communication actually takes place. I imagine a more powerful CPU wouldn't have issues with older video cards but how would I know if an i7-4770 has enough bandwidth or how much bandwidth it has to not get bogged down by 4 GPUs?
____________

Profile ZalsterProject donor
Volunteer tester
Avatar
Send message
Joined: 27 May 99
Posts: 949
Credit: 41,718,371
RAC: 69,657
United States
Message 1492701 - Posted: 21 Mar 2014, 14:10:26 UTC - in response to Message 1492596.

Yeah, that was just an academic discussion on how many GPUs. No one was really going to build one that big. Much easier to build several smaller rigs. I don't know the answers to your questions. What I can tell you is that the amount of data being conducted through the PCI slot isn't so great but it is very frequent. Seti is an intense read/write program. In most discussions, Gamers will talk about the PCI slots with PCIx16, x8, x4..etc dealing with bandwidth. That doesn't really apply with seti since it is small amount of data each time. If you don't have a high end MB with top of the line PCIx16 slot, you can still run seti. I run several GPUs thru PCIx1 risers off the MB. The main thing is having enough cores on your chip to feed the GPUs. There are discussions on how many core per GPU. Usually it as follows. 1 core per each GPU. Are you planning on only using the GPUs for Seti or are you also planning on using the Cores to crunch as well? I'll give you an example. Some people prefer to only use the GPUs in their computers, so they will have 3 GPUs on a quad core and restrict Bonic to only use the GPUs and not get any work units for the CPU.(this is done in the preference on your main account page) Others may have a 12 core (6 physical and 6 virtual) and run 8 GPUs with 8 cores attached leaving 4 free. Those 4 free cores are then crunching CPU oriented work units. Confused? Hopefully not. There is talk about leaving 1 core free for the systems usage. That is up to you. If all you are doing is crunching, then it really shouldn't matter since there isn't a lot of other things going on at the same time. However, if you are planning on using this computer for other things, then you will need to restrict Bonic not to use all of the cores for it's own purposes. Otherwise it will become bogged down and slow or even stall while doing other things. You can do this in the pull down section of the Bonic Manager. Under tools go to computing perferences, click on Processor usage. Second to last line says "On multiprocessor systems, use at most XX % of the processors" Here you need to do some math. If you are using a 4 core then use at most 75% . If you are using 8 care then use at most 87.5%. That will leave 1 core free for OS system. If you don't care since it's only going to crunch then leave it at 100%. This section has nothing to do with mating 1 core to 1 GPU.. when people say save 1 core for 1 GPU and tell you to go there...It does not work that way...I've tried when I first started and realized it wasn't true. Example. If you changes the values and tell it use only 50% because you want to save 2 cores for 2 GPUS and want to run the other 2 cores for crunching on a 4 core, both the CPU and GPU will only use the same 2 of the 4 cores and the remaining 2 core won't be touched at all....You would be better off just using 100% of all cores. Only when you use an optimized programs can you adjust how many cores are being utilized. If you plan on using an optimized program like Lunatics, then it will install it's own preferences into your data folder. There it tends to run 0.4 CPU cores per 1 GPU i believe. If you plan on modifying the code then you will have to play with it to find the right combination.( and by play with it, I mean rewriting code) Ok, sorry for being long winded. Hope this helps.

danieloconnor
Volunteer tester
Send message
Joined: 16 Nov 05
Posts: 17
Credit: 1,871,252
RAC: 4,198
Canada
Message 1493212 - Posted: 22 Mar 2014, 7:56:26 UTC - in response to Message 1492701.

That wasn't confusing at all. The motherboard I will be using supports 4 GPUs at 8x each, 32gb of RAM, and any Intel LGA 1150 processor (with graphics). To start, I plan on running the motherboard and CPU alone; until my budget allows for some random video cards with CUDA cores.

I have another question. You mentioned 1 CPU-core per GPU is good to keep the work flowing between both, but do hyper-threaded cores count as cores? An example is a 4-core processor with hyper-threading has 8 cores. I suppose it wouldn't really matter if the CPU only has 4 cores and the motherboard only holds 4 GPUs. As I said, my main goal at the moment is to choose the correct processor and power supply for the motherboard I am using; to provide adequate bandwidth and proper amount of wattage for the system.

There has been no desire to overclock the CPU, then again, that may be something to take into consideration. There is an Intel i7 4770K that also fits within my budget (it's the overclocking version of the CPU I am considering). But again, not much thought has been given to overclocking the CPU. Do you have any experience with a multi-GPU system with an overclocked CPU?
____________

Profile ZalsterProject donor
Volunteer tester
Avatar
Send message
Joined: 27 May 99
Posts: 949
Credit: 41,718,371
RAC: 69,657
United States
Message 1493224 - Posted: 22 Mar 2014, 8:40:46 UTC - in response to Message 1493212.

Hyperthreading cores are considered cores. My example of the 12 cores was based on a 6 physical core and 6 virtual cores that feed 8 GPUs and left 4 core free for crunching. I have never OC'ed my cores. I know I have the potential to do so but I've never done it. OCing the GPUs is talked about alot, but you have to be careful when doing that, as at some point it starts to thrash the results and you end up with all invalids results. As far as PSUs. If you are really concern you can do a google search for eXtreme Power Supply Calculator lite. Should show you a page called extreme outervision. One of the best calculators it shows how how much power you need. It goes into Motherboard types, chips, ram, fans, HDs, USBs, GPU cards by make and model and number said cards. That will give you an idea on Power supplies.

danieloconnor
Volunteer tester
Send message
Joined: 16 Nov 05
Posts: 17
Credit: 1,871,252
RAC: 4,198
Canada
Message 1493523 - Posted: 22 Mar 2014, 19:18:29 UTC - in response to Message 1493224.

I forgot to ask in my previous post, how does 16x, 8x, and 4x work on PCIE slots? I remember in the days a 4x CD-ROM drive was much slower than an 8x and 16x. Is this similar with PCIE slots? And, how does one go about making sure a GPU supports the read/write times if read/write times is what this measures?
____________

Profile ZalsterProject donor
Volunteer tester
Avatar
Send message
Joined: 27 May 99
Posts: 949
Credit: 41,718,371
RAC: 69,657
United States
Message 1493784 - Posted: 23 Mar 2014, 0:52:54 UTC - in response to Message 1493523.

Here is a link on how PCI slots work. They are important in Gaming, not so much with Seti. Yes larger number means more bandwidth of information, good for gaming, not so important for seti. Don't worry about Read/write times, we aren't dealing with that when we talk about GPU and PCI slots. That is taken care of by Bonic and Seti's program. Main thing, Make sure use similar GPUs, have enough power for all of them and enough fans to cool you unit. If your GPU will fix in a slot and there is room, then you can use them. Once you starting getting into modifing your computer, we can talk about risers and such but that is for a different time.


http://computer.howstuffworks.com/pci-express.htm

Profile ZalsterProject donor
Volunteer tester
Avatar
Send message
Joined: 27 May 99
Posts: 949
Credit: 41,718,371
RAC: 69,657
United States
Message 1494035 - Posted: 23 Mar 2014, 15:42:53 UTC - in response to Message 1493784.

When I spoke of read/write, I was referring to Harddrive or SSD. Sorry I confused you.

danieloconnor
Volunteer tester
Send message
Joined: 16 Nov 05
Posts: 17
Credit: 1,871,252
RAC: 4,198
Canada
Message 1494066 - Posted: 23 Mar 2014, 18:02:02 UTC - in response to Message 1494035.

No worries. Thank you for the tutorial on PCI/E. I'm now searching the web for the best price for the Asus Z87-WS. I looked at the online documentation for this motherboard and found that Asus provides their User Manuals online. I didn't think of this earlier and found that the manuals answers a lot of questions I was having.

Pretty exciting stuff.
____________

Questions and Answers : GPU applications : 4-Way GPU computing

Copyright © 2014 University of California