MAGMA weak performance SGEMM

Open discussion for MAGMA

MAGMA weak performance SGEMM

Postby lcebamanos » Fri Mar 29, 2013 2:50 pm

I was wondering if this result is correct... I see huge difference in performance. MAGMA's been compiled using ACML.

Code: Select all
[color=#BF4040]MAGMA 1.3.0
device 0: Tesla K20c, 705.5 MHz clock, 4799.6 MB memory, capability 3.5

Usage:
  testing_sgemm [-NN|NT|TN|TT] [-N 1024]

Testing transA = N  transB = N
    M    N    K     MAGMA GFLop/s    CUBLAS GFlop/s       error
==================================================================
 1024  1024  1024       617.45           1797.06         3.051758e-05
 1280  1280  1280       712.59           1977.51         3.814697e-05
 1600  1600  1600       789.74           2212.86         4.577637e-05
 2000  2000  2000       833.51           2209.03         6.103516e-05
 2500  2500  2500       795.49           2255.99         7.629395e-05
 3125  3125  3125       840.17           2442.48         9.155273e-05
 3906  3906  3906       855.86           2400.39         9.155273e-05
 4882  4882  4882       870.25           2456.92         1.831055e-04
 6102  6102  6102       867.47           2486.71         2.136230e-04[/color]
Last edited by mgates3 on Mon Apr 01, 2013 3:56 pm, edited 1 time in total.
Reason: add [code] to make output readable
lcebamanos
 
Posts: 4
Joined: Fri Mar 29, 2013 2:46 pm

Re: MAGMA weak performance SGEMM

Postby mgates3 » Mon Apr 01, 2013 3:59 pm

That is correct. CUBLAS has significantly tuned their GEMM for the Kepler (3.x) architecture, while the MAGMA GEMM was only tuned for the Fermi (2.x) architecture.
-mark
mgates3
 
Posts: 442
Joined: Fri Jan 06, 2012 2:13 pm

Re: MAGMA weak performance SGEMM

Postby luiceur » Tue Apr 02, 2013 5:58 am

Hope MAGMA beats that soon with a new release
luiceur
 
Posts: 26
Joined: Tue Jul 10, 2012 4:38 am


Return to User discussion

Who is online

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