SLI or AMD's Crossfire is not used when doing CUDA or OpenCL work. In the early days you had to turn those features of to be able to process the work but newer drivers handle that automatically.
Specialy in E@H and the way it´s works (by ussing a lot of PCIe resources), sure 2x2GPU hosts will perform better against a 1x4GPU host. But take something in mind, even to use a 2xGPU host at 100% you need a "better" MB (means X79 chipset or better for Intel). I have few "less capable" MB with 2xGPU on each, works but can´t get all the GPU potential due the incapacity of the chipset to handle the PCI-e requests.
SLI is a feature, intended almost only for games, in which each GPU of the SLI set just renders a half (or 1/3 or 1/4) of each frame... Doing this the whole frame is drawn faster.
Now, for apps that use the multiprocessors of the GPU just for "maths" SLI is irrelevant, so it doesnt matter if the SLI bridge is installed or not and, with newer drivers, it shouldnt even matter if this feature is enabled or not by software. (On older drivers and on certain GPU models, it was needed to disable SLI to get better performances and in some cases it was needed to disable SLI to avoid faillures in the apps, but it never requiered to remove the SLI bridge).
While for SLI, you need equal GPUs, for number crunching you dont, but BOINC may need aditional steps to work. As you are using equal GPUs, BOINC should see all of them and each one will be running different tasks independently of what the other GPUs were doing.
But, there is trick, BOINC 7.0.28 has a bug and sometimes it doesnt detects all the GPU as equals and then it doesnt use all of them. You are using BOINC 7.0.28 and if this is happening you will see in the beggining of the event log (after starting BOINC) that there are GPUs marked as "not used".
There are 2 ways to fix this, the most recomended is to upgrade to the latest beta version of BOINC (currently 7.0.56, which is a release candidate) or to use the cc_config.xml file to instruct BOINC to use all the GPUs...
There are a lot of threads with instructions on how to do any of this options, but if need more help, just ask ;D
so each gpu will be used independent of each of other..no sli interconnect.
then 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..
I suspect that with each card separate the memory bandwidth in the card gpu will be the limiting factor and not the number of cell processors doing the work or the pci bus it is connected to since each card is independent of each other..
2 , 3 or 4 way sli
)
It's not perfect or complete but have a look here....
http://www.dskag.at/images/Research/EinsteinGPUperformancelist.pdf
shows work unit times...
The speed at which 2 cards will drop the time a bit since the card slots are 16x and 8x and 8x and then 4x
PC setup MSI-970A-G46 AMD FX-8350 8 core OC'd 4.45GHz 16GB ram PC3-10700 Geforce GTX 650Ti Windows 7 x64 Einstein@Home
thanks for the numbers man.
)
thanks for the numbers man. was wondering if the scalling of sli was the same as gaming.. looks like it is.. 2 cards is optimal vs 3 or 4 cards..
meaning 2 pc with 2 gpu can out perform one pc with 4 gpu cards in it..
thanks for the input..
SLI or AMD's Crossfire is not
)
SLI or AMD's Crossfire is not used when doing CUDA or OpenCL work. In the early days you had to turn those features of to be able to process the work but newer drivers handle that automatically.
Specialy in E@H and the way
)
Specialy in E@H and the way it´s works (by ussing a lot of PCIe resources), sure 2x2GPU hosts will perform better against a 1x4GPU host. But take something in mind, even to use a 2xGPU host at 100% you need a "better" MB (means X79 chipset or better for Intel). I have few "less capable" MB with 2xGPU on each, works but can´t get all the GPU potential due the incapacity of the chipset to handle the PCI-e requests.
As the others have said, each
)
As the others have said, each GPU acts as an individual card and tasks are not processed in parallel with other GPU's.
thanks for the info.. so
)
thanks for the info..
so each gpu card is set to run work units independent of each other?
Yes, as long as they are
)
Yes, as long as they are identical (if not I believe only the most powerful GPU will be used).
If you do have non-identical cards, you can create a little text file to use all GPUs - a search for "use_all_gpus" willl explain how to do it.
yes always use the same
)
yes always use the same gpu..
for sli.
my under standing is that the task or work units run separate on each gpu.
no sli bridge connector on them is that correct?
thanks for the input.
SLI is a feature, intended
)
SLI is a feature, intended almost only for games, in which each GPU of the SLI set just renders a half (or 1/3 or 1/4) of each frame... Doing this the whole frame is drawn faster.
Now, for apps that use the multiprocessors of the GPU just for "maths" SLI is irrelevant, so it doesnt matter if the SLI bridge is installed or not and, with newer drivers, it shouldnt even matter if this feature is enabled or not by software. (On older drivers and on certain GPU models, it was needed to disable SLI to get better performances and in some cases it was needed to disable SLI to avoid faillures in the apps, but it never requiered to remove the SLI bridge).
While for SLI, you need equal GPUs, for number crunching you dont, but BOINC may need aditional steps to work. As you are using equal GPUs, BOINC should see all of them and each one will be running different tasks independently of what the other GPUs were doing.
But, there is trick, BOINC 7.0.28 has a bug and sometimes it doesnt detects all the GPU as equals and then it doesnt use all of them. You are using BOINC 7.0.28 and if this is happening you will see in the beggining of the event log (after starting BOINC) that there are GPUs marked as "not used".
There are 2 ways to fix this, the most recomended is to upgrade to the latest beta version of BOINC (currently 7.0.56, which is a release candidate) or to use the cc_config.xml file to instruct BOINC to use all the GPUs...
There are a lot of threads with instructions on how to do any of this options, but if need more help, just ask ;D
as I suspected. so each
)
as I suspected.
so each gpu will be used independent of each of other..no sli interconnect.
then 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..
I suspect that with each card separate the memory bandwidth in the card gpu will be the limiting factor and not the number of cell processors doing the work or the pci bus it is connected to since each card is independent of each other..
thanks for the input.