MAGMA versus CUBLAS performance

[pxavier]

Hello,

I ran some benchmark comparisons between MAGMA BLAS 0.2 sgemm and cublas sgemm and found that magma outperforms cublas by several folds. I am supposing that the improvement has to do with splitting the matrix and the processing in parallel the subtasks using both GPU and CPU cores. Is anyone able to provide some general insights on whether this assumption is correct for the current version of magma ?

Regards

[Stan Tomov]

Hi,
MAGMA BLAS implements only part of the BLAS specification and is meant as a complement to CUBLAS. It improves on certain CUBLAS routines in specific situations (arguments) that are needed in the MAGMA routines. In general the improvements are only up to 2 times only in some cases. If you see improvements of several folds most probably something has gone wrong. Which routine did you test and on what GPU? If you ran something through a magma testing routine, are the errors as expected, e.g., zero as in a

Code: Select all

[tomov@cumin testing]$ ./testing_sgemv

Usage
		 testing_sgemv N

device 0: GeForce GTX 280, 1296.0 MHz clock, 1023.8 MB memory
device 1: Quadro NVS 290, 918.0 MHz clock, 255.3 MB memory

   n   CUBLAS,Gflop/s   MAGMABLAS0.2,Gflop/s   "error"
==============================================================
   64        0.20       0.39 		        0
  128        0.55       1.21 		        0
  192        0.87       2.17 		        0
  256        1.20       3.20 		        0
  320        1.55       4.36 		        0
  384        1.90       5.46 		        0
  448        2.27       6.27 		        0
  512        2.52       7.71 		        0
  576        2.82       9.09 		        0
  704        3.59      11.39 		        0
  832        4.31      13.84 		        0
  960        4.93      16.17 		        0
 1088        5.56      18.50 		        0
 1216        6.29      20.68 		        0
 1408        7.29      24.63 		        0
 1600        8.34      27.83 		        0
 1792        9.40      31.33 		        0
 1984       10.39      34.68 		        0
 2240       11.74      37.87 		        0
 2496       13.13      42.82 		        0
 2816       14.78      47.34 		        0
 3136       16.45      50.56 		        0
 3520       18.49      54.22 		        0
 3904       20.40      55.22 		        0
 4352       22.76      60.70 		        0
 4800       24.81      60.24 		        0
 5312       27.41      61.95 		        0
 5888       30.23      64.80 		        0
 6528       33.20      64.08 		        0
 7232       36.64      65.17 		        0
 8000       40.16      64.58 		        0

Stan