CUBLAS performance under toolkit 4.1

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everybody, does any of you know whether the cublas GEMM performance increases in the new toolkit (4.1)?
I can access a server equipped with a GTX580 card. But unfortunately, the OS Fedora 13 is no longer supported under 4.1 toolkit.

So, if you can access a GTX580 or similar Fermi card, can you help to test the SGEMM and DGEMM’s performance to see whether the 4.1 is better than 4.0?

Thanks in advance. External Image

2

The fact that a Linux version is not supported usually means that it is not tested.
For a Linux version the important things are the gcc and libc versions. Just look for a package for another distro that has the same version of gcc/libc and
it will work.

There were some improvements in GEMM in 4.1.

3

Thanks mfatica,

I tried the one for 10.4 ubuntu. It works.

The large sgemm performance increases about 4% over 4.0 on gtx580 (from ~1000 to ~1040 with 4800x4800 matrices).

actually, i am working on the sgemm_N_N. by far i have achieved ~1130 with 4800x4800 matrices. (7% to 8% over 4.1 cublas)