Matrix Multiplication Results

• Matrix multiplication on larger servers, including large arrays.
  

  Note how a speedup of 3 for 1600x1600 is achieved on the SGI Power Challenge solely by using some basic compiler options:

  "f77 -r8000 -mips4 -64 -O3" ............................ 212 sec
  "f77 -r8000 -mips4 -64 -O3 -LNO:ou=6" .................. 105 sec
  "f77 -r8000 -mips4 -64 -O3 -LNO:ou=8 -pfa -WK,-p=1" .... 71 sec.
  

  On the other hand, the built-in library (f77 ... -lcomplib.sgimath) provides an additional speedup of 2 (routine DGEMM): 34 sec.

  Therefore, between simplistic and good approaches there lies a speed factor of 6.

• Matrix multiplication on workstations.
• Matrix multiplication on PCs (Linux, Pentium).

R. Krivec