Hello,
I'm building a system using not one but four GPU cards each in a separate PCI-e slot e.g. nVidia 680 GTX. Will Magma automatically load balance across all these four cards? I guess they will show up with different device Id's by calling cudaGetDeviceProperties.
TIA,
Best regards,
Giovanni
how does it work with multiple GPU devices?
Re: how does it work with multiple GPU devices?
Yes, some MAGMA routines are written for multiple GPUs. Specifically,
magma_xgetrf( ... )
magma_xgeqrf( ... )
magma_xpotrf( ... )
where x is the precision. In the forthcoming release, there will also be multi-GPU eigenvalue solvers.
You need to set MAGMA_NUM_GPUS in your environment to the number of GPUs that you want to use. E.g. with csh,
setenv MAGMA_NUM_GPUS 4
Is there particular functionality that you are interested in?
-mark
magma_xgetrf( ... )
magma_xgeqrf( ... )
magma_xpotrf( ... )
where x is the precision. In the forthcoming release, there will also be multi-GPU eigenvalue solvers.
You need to set MAGMA_NUM_GPUS in your environment to the number of GPUs that you want to use. E.g. with csh,
setenv MAGMA_NUM_GPUS 4
Is there particular functionality that you are interested in?
-mark
Re: how does it work with multiple GPU devices?
Hello mark,
Thank you very much! I'm interested in speeding up computation for a machine learning framework, I think the main factorizations which you named are already covered, other than that I am trying to leverage magma into Eigen, since the function signatures are not that different from MKL I think it is very doable to run magma behind the Eigen abstractions where possible.
Best regards,
Giovanni
Thank you very much! I'm interested in speeding up computation for a machine learning framework, I think the main factorizations which you named are already covered, other than that I am trying to leverage magma into Eigen, since the function signatures are not that different from MKL I think it is very doable to run magma behind the Eigen abstractions where possible.
Best regards,
Giovanni