PGI vs. Intel

The Intel compiler was interesting because it seemed to promise a significant speed up compared to the PGI Fortran90 compiler. It was able to produce processor-specific instruction sets rather than general i486 instructions. In additon, it was supposed to be better at handling large memory problems than the PGI compiler. At the time we made our decision, we did not have a working version of SEPlib for the Intel compiler, so we decided to dumb down our tests to simple matrix multiplication. Figures 2 and 3 show the result of doing a real and complex matrix multiplication with a matrix of four million elements. Note how the Intel specific code is significantly faster than the corresponding PGF code. The speed advantage offered by the Intel compiler led us to choose the dual P3 option for our cluster.

 

float-comp Figure 2 Speed comparison for matrix multiplication. The horizontal axis is machine (P2-550, P3-800, P4-1700). The different curves represent the PGF compiler, Intel compiler, and Intel with machine-specific instructions. Note the significant advantage of the Intel machine specific code.

 

complex-comp Figure 3 Speed comparison for complex matrix multiplication. The horizontal axis is machine (P2-550, P3-800, P4-1700). The different curves represent the PGF compiler, Intel compiler, and Intel with machine specific instructions. Note the significant advantage of the Intel machine specific code.