2 , 3 or 4 way sli

Tom Plummer
Tom Plummer
Joined: 12 Mar 13
Posts: 13
Credit: 1260114
RAC: 0

or should one go for the max

or should one go for the max cell processor and not wory about the work units being memeory bound on the compute power used for the project.

I notice that a 650 ti gets more average credit with more cores..but my question is if I put 2 or 3 of them in one mother board will they all act independent of each other and then will they will be memory bound using the project.

meaning if I have 1 gpu with less cores and more memory bandwidth will it perform better than 2 gpu with more cores and not as much memory bandwidth..

thanks for the input.

Neil Newell
Neil Newell
Joined: 20 Nov 12
Posts: 176
Credit: 169699457
RAC: 0

RE: ...how does the number

Quote:

...how does the number of cores and the pci interface affect the work units.

meaning that if I have 2 cards in system do I have to worry about the amount of
data that one card can use vs the pci bus transfer speed 16 x vs 8 vs 4..

The BRP application is bandwidth intensive, so it's recommended to put the cards in x16 PCIe2.0 slots. The faster the card, the more performance will be lost because of bandwidth limits.

Quote:
I notice that a 650 ti gets more average credit with more cores..but my question is if I put 2 or 3 of them in one mother board will they all act independent of each other and then will they will be memory bound using the project.

I've found that the performance of a single GPU is little affected by the addition of a second GPU. Also it's generally recommended to reserve a CPU core just to feed the GPUs (e.g. only use 3 cores for CPU work on a quad core). From the point of the GPU's, it doesn't seem to matter too much what sort of CPU (how powerful) it is feeding them. Try reserving one CPU core as a starting point, and then monitor the performance and make small changes to see if they help or not.

To reserve CPU cores, this also goes in the cc_config.xml text file referred to above.

For example, here is a cc_config.xml file I used on a quad core system with 2 GPUs:

   
   
   
     3
     1
   

Also, once you get practiced, you can actually run more than one GPU task at a time on the same GPU. For nvidia, I normally run 2 tasks at once which gives an extra 15% performance of so. For AMD, more than 2 helps (see the performance thread for details, there is a useful PDF there).

Horacio
Horacio
Joined: 3 Oct 11
Posts: 205
Credit: 80557243
RAC: 0

There are several things that

There are several things that affects performance...
But it depends on what application you run...

Eintein's BRPs are hybrid apps which do part of calcs on CPU (due to the CPU beeing better for secuencial operations that cant be parallelized among the multi procs of the GPU)
So, here, in Einstein the PCIe bandwith is more relevant than in other projects... and for faster GPUS is more notorios because they will need to comunicate with the CPU more often.

So if you have a MBD with shared PCI lanes, then when you install only one GPU it will be able to use the x16 lanes, while when you install 2, they will be running at 8x... In this point you will see that with only one GPU the times are a bit shorter than when using 2 GPUs... but with 2 gpus you will be doing twice the work which will give an overall gain in the performance.

About the GPU memory, thats also important, and ussually faster memory (GDDR5) with wider bus (384bits or 256 bits) will make the GPU faster than other models with slower memory (GDDR3) or narrow buses (128 bits or 192 bits)...

About the number of GPU cores, Ive not seen any direct relationship. It is ussually true that GPUs with more cores give more performance, but I guess that this is not due to the number of cores per se, but to the fact that GPUs with more cores are ussually using newer techonologies and then they are better in everything else.

At last the CPU speed and system memory speed also affects the GPU performace, but this effect is ussually negligible. Which really matters is to avoid the GPUs starving due to CPUs not beeing able to feed them, but this is fixed reserving some CPU cores to be allways ready to attend the GPUs.

mikey
mikey
Joined: 22 Jan 05
Posts: 11889
Credit: 1828153831
RAC: 203591

RE: There are several

Quote:

There are several things that affects performance...
But it depends on what application you run...

Eintein's BRPs are hybrid apps which do part of calcs on CPU (due to the CPU beeing better for secuencial operations that cant be parallelized among the multi procs of the GPU)
So, here, in Einstein the PCIe bandwith is more relevant than in other projects... and for faster GPUS is more notorios because they will need to comunicate with the CPU more often.

So if you have a MBD with shared PCI lanes, then when you install only one GPU it will be able to use the x16 lanes, while when you install 2, they will be running at 8x... In this point you will see that with only one GPU the times are a bit shorter than when using 2 GPUs... but with 2 gpus you will be doing twice the work which will give an overall gain in the performance.

About the GPU memory, thats also important, and ussually faster memory (GDDR5) with wider bus (384bits or 256 bits) will make the GPU faster than other models with slower memory (GDDR3) or narrow buses (128 bits or 192 bits)...

About the number of GPU cores, Ive not seen any direct relationship. It is ussually true that GPUs with more cores give more performance, but I guess that this is not due to the number of cores per se, but to the fact that GPUs with more cores are ussually using newer techonologies and then they are better in everything else.

At last the CPU speed and system memory speed also affects the GPU performace, but this effect is ussually negligible. Which really matters is to avoid the GPUs starving due to CPUs not beeing able to feed them, but this is fixed reserving some CPU cores to be allways ready to attend the GPUs.

Most, if not all, Nvidia cards need a cpu core to feed each gpu, while AMD cards sometimes do and sometimes don't, it depends on the project. The project DestRTgen uses 0.959% or higher, of a cpu for my AMD cards AND my Nvidia cards. But Moo only used 0.05% of a cpu core for the AMD gpu's, I don't remember the actual number for the Nvidia gpu's at Moo, but it was nearly a full cpu core. Since a gpu can do around 10 times as much work as a cpu can, generally speaking, it does not hurt your crunching to leave it free as it IS being used.

tolafoph
tolafoph
Joined: 14 Sep 07
Posts: 122
Credit: 74659937
RAC: 0

Last year I bought a GTX580

Last year I bought a GTX580 and put my old GTX260 as the 2nd GPU in my computer.
The performance of the GTX580 dropped over 10% due to the change from 16x to 2 times 8x lanes. Theoretical performance of GTX580 ~50000 credits/ day, GTX260 10000 credits/day.

David Rapalyea
David Rapalyea
Joined: 3 Jan 13
Posts: 79
Credit: 63886821
RAC: 0

Hi Horatio, Within the

Hi Horatio,

Within the week I will be able report on a real world machine that will be delivered next week. It is dual core @2.3ghz, and will include four separate GTX 650 cards. One thing that has not been yet mentioned is Xpress 2.0 v. 3.0.

In this case both the cards and the MB are 3.0. And the lowest slot is 8 lanes [8x]. But if each GPU needs a dedicated core to feed it then none of this will matter. I will only be able to run two of the GPU cards.

drummmmm rolllll with fingers crossed.....

Rapalyea

Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go

mikey
mikey
Joined: 22 Jan 05
Posts: 11889
Credit: 1828153831
RAC: 203591

RE: Hi Horatio, Within the

Quote:

Hi Horatio,

Within the week I will be able report on a real world machine that will be delivered next week. It is dual core @2.3ghz, and will include four separate GTX 650 cards. One thing that has not been yet mentioned is Xpress 2.0 v. 3.0.

In this case both the cards and the MB are 3.0. And the lowest slot is 8 lanes [8x]. But if each GPU needs a dedicated core to feed it then none of this will matter. I will only be able to run two of the GPU cards.

drummmmm rolllll with fingers crossed.....

Rapalyea

It doesn't NEED a cpu core, it can just crunch more units faster if it has a cpu core to keep it fed and happy. Depending on your motherboard you might be able to upgrade to a quad core fairly reasonably, if it isn't built yet call them QUICK! If it has try what you have and then see at what cost a quad core, or more, will cost you.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.