MAGMA 0.3 for Fermi and CUDA 3.2 RC2 results

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)

MAGMA 0.3 for Fermi and CUDA 3.2 RC2 results

Postby Allan Menezes » Tue Sep 14, 2010 12:07 am

Here are the results of Magma 0.3 for Fermi versus Cublas 3.2 RC2 on a GTX-480:

Code: Select all
This are MAGMA 0.3 DGEMM and SGEMM Routines for Fermi GPUs.
In this version matrix sizes have to be divisible by 64


Usage:
      ./testing_dgemm N

    N   magmablas0.3 GFLops/s   cudablas-3.2 GFlops/s    error
==========================================================================
  512        130.18208              127.70478         0.000000e+00
 1088        153.85420              153.24094         0.000000e+00
 1664        157.40103              156.64116         0.000000e+00
 2240        160.20845              159.43350         0.000000e+00
 2816        159.83791              159.06029         0.000000e+00
 3392        163.98298              164.37494         0.000000e+00
 3968        164.20745              164.60266         0.000000e+00
 4544        164.12606              164.75651         0.000000e+00
 5120        164.55581              165.02481         0.000000e+00
 5696        164.41005              164.91638         0.000000e+00
 6272        164.49734              165.06085         0.000000e+00
 6848        164.47954              165.15991         0.000000e+00
 7424        164.53056              165.10375         0.000000e+00

This is a MAGMA 0.3 SGEMM Routine for Fermi GPUs.
In this version sizes have to be divisible by 96.


Usage:
      ./testing_sgemm N

    N   magmablas0.3 GFLops/s   cudablas-3.2 GFlops/s    error
==================================================================
  864        665.26307              702.20201         0.000000e+00
 1440        801.71406              824.51581         0.000000e+00
 2016        794.79485              814.87221         0.000000e+00
 2592        817.72439              832.54093         0.000000e+00
 3168        824.51008              836.61608         0.000000e+00
 3744        825.33249              837.03466         0.000000e+00
 4320        832.99221              842.95151         0.000000e+00
 4896        830.81927              840.00669         0.000000e+00
 5472        833.72471              842.49372         0.000000e+00
 6048        834.53324              842.74165         0.000000e+00
 6624        834.80414              842.60437         0.000000e+00
 7200        837.26658              844.44771         0.000000e+00
 7776        836.65414              843.91701         0.000000e+00
 8352        837.34876              844.12340         0.000000e+00



So for sgemm the perfomance of the gtx-480 is now better for culas-3.2 rc2.
Allan
Allan Menezes
 
Posts: 14
Joined: Wed Aug 05, 2009 10:01 pm

Re: MAGMA 0.3 for Fermi and CUDA 3.2 RC2 results

Postby Stan Tomov » Tue Sep 14, 2010 1:04 am

Alan,
Thanks for posting these results. The gemms in CUBLAS 3.2 for Fermi are
based on the MAGMA 0.3 Fermi gemms. That's excellent it is out now.
Stan
Stan Tomov
 
Posts: 254
Joined: Fri Aug 21, 2009 10:39 pm


Return to User discussion

Who is online

Users browsing this forum: Bing [Bot] and 2 guests