Kepler vs Fermi - no performance diff on testing_zgeev

Open discussion for MAGMA

Kepler vs Fermi - no performance diff on testing_zgeev

Postby tpistor » Mon Dec 09, 2013 7:00 pm

I see no performance difference on zgeev Kepler vs. Fermi.
You can see from the output that the run times are the same (even though the magma capability goes from 3 to 2).
Is this simply because no Kepler optimizations were made? If yes, then:
a) Can Kepler optimizations be made? (is it possible)
b) Are there any plans to do so?



tpistor@euclid:~/testing$ ./testing_zgeev
MAGMA 1.4.1 , capability 3.0
device 0: Tesla K20c, 705.5 MHz clock, 4799.6 MB memory, capability 3.5
device 1: GeForce GTX 650 Ti, 928.0 MHz clock, 1023.3 MB memory, capability 3.0
Usage: ./testing_zgeev [options] [-h|--help]

N CPU Time (sec) GPU Time (sec) |W_magma - W_lapack| / |W_lapack|
===========================================================================
1088 --- 3.05
2112 --- 10.12
^C
tpistor@euclid:~/testing$ ../magma-1.4.1/testing/testing_zgeev
MAGMA 1.4.1 , capability 2.0
device 0: Tesla K20c, 705.5 MHz clock, 4799.6 MB memory, capability 3.5
device 1: GeForce GTX 650 Ti, 928.0 MHz clock, 1023.3 MB memory, capability 3.0
Usage: ../magma-1.4.1/testing/testing_zgeev [options] [-h|--help]

N CPU Time (sec) GPU Time (sec) |W_magma - W_lapack| / |W_lapack|
===========================================================================
1088 --- 3.04
2112 --- 10.07
^C
tpistor
 
Posts: 2
Joined: Mon Dec 09, 2013 2:04 am

Re: Kepler vs Fermi - no performance diff on testing_zgeev

Postby mgates3 » Tue Dec 10, 2013 3:59 pm

Most MAGMA functions rely on CUBLAS for BLAS operations. Nvidia has optimized CUBLAS for Kepler, so we automatically get their improvements when running on a Kepler card -- regardless of the GPU_TARGET setting.

The GPU_TARGET setting affects only how MAGMA BLAS functions are compiled. We need some of these to fill in functionality not provided by CUBLAS (e.g., lacpy, lascl, lange), but in most cases it is not performance critical functionality. In the past we relied more on our MAGMA BLAS functions such as our Fermi GEMM, but Nvidia has incorporated that GEMM into CUBLAS.

-mark
mgates3
 
Posts: 416
Joined: Fri Jan 06, 2012 2:13 pm


Return to User discussion

Who is online

Users browsing this forum: Google [Bot], yezhiyun and 2 guests

cron