Some performance numbers on various machines

For small systems (<100 atoms), the FFT dominates the computational effort. The following table gives the performance in MFLOPS/node of the FFT on several platforms. The FFT includes the multiplication with the potential in realspace. The system has about 20 wave functions and a grid size of 64x64x64.

 
              Nodes   1      2      4       8     16
Platform   \


IBM RS6000 Power 2   80.0   
(civet, old)

SGI PowerChallenge   49.5    50.5  59.9  
R10000 non-dedic (old)

SGI Origin 2000      84.37   79.5  81.6    24.3
R10000 non-dedic (new)

SGI Origin 200 (new) 70.5    65.2  57.2
R10000 180 MHz      

T3E                  69.8    64.3  61.7   -     57.3
(streams off, old code) 

T3E                  85.7    80.5  75.9  74.0   65.9
(streams on,new code)