Muti GPU SGEMM implementation

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
dwaras3284
Posts: 2
Joined: Wed Jul 09, 2014 3:16 pm

Muti GPU SGEMM implementation

Post by dwaras3284 » Wed Jul 09, 2014 3:21 pm

Hi,

I'm new to MAGMA library. I"m unable to find any documentation or source code for multi GPU implementation for SGEMM. Can you please let me know if it has been implemented on multi GPU? I could only find single GPU implementation. Am I missing something?

Thanks for your help.

mgates3
Posts: 918
Joined: Fri Jan 06, 2012 2:13 pm

Re: Muti GPU SGEMM implementation

Post by mgates3 » Thu Jul 10, 2014 10:42 am

MAGMA does not implement a multi-GPU gemm. It is easy to fashion a multi-GPU gemm out of single-GPU gemm, but it depends on how the matrices are distributed across the GPUs. See this previous message:
viewtopic.php?f=2&t=569&p=1784#p1784

If you have a specific data distribution in mind for A, B, C, I can help you figure out what to do.

CUBLAS has a CPU interface that uses multi-GPUs.
https://developer.nvidia.com/cublasxt

-mark

Post Reply