MAGMA performance

[llvll0hsen]

Hey all,

I was wondering is there any study regarding MAGMA performance? there are just two plots on the handbook but there is nothing more.

Another thing is I am kind of between choosing appropriate method for choosing my different computation . In one side is CUDA with some libraries like CULA and MAGMA , and the other hand there is MAHOUT library on Hadoop cluster which also seems to perform very good! So I am looking to see which one one theses can over perform the other one. As long as I understood GPU works better for Matrix factorization over mahout(which is think its because of the less communication costs in compare of map reduce ) but for a very large data set GPU can not compete because of the memory limitation.

So can you tell me whether I am wrong or not? and also how can compare them better ?

Thanks

[mgates3]

If you look on the MAGMA site under Publications, there are papers for the different algorithms with performance analysis of each.
http://icl.cs.utk.edu/magma/pubs/

I'm not familiar with the MAHOUT library, so I can't comment on its functionality or performance. However, MAGMA has non-GPU-resident algorithms for LU, Cholesky, and QR, so they are not limited by a GPU's memory.

-mark