tef (06/09/82)
After waiting 4 months for delivery, and then another 45 days for the necessary ECO's to be installed, I now have an FPA installed and working on my VAX-750. Accordingly, I ran the benchmark which I have been using as a standard for several years now. The benchmark does a 2-dimensional matrix diagonalization in order to compute eigenvalue vectors and is very floating point intensive with few procedure calls. The results are as follows (all times in seconds; "xf77" is UCB's new experimental fortran compiler): 11/70 w/FPU 780 w/FPA 750 w/o FPA 750 w/FPA f77 w/ real*4 62.5 26.1 105.4 50.8 f77 w/ real*8 71.1 39.4 171.7 59.9 xf77 w/ real*4 --- 20.8 79.1 32.7 xf77 w/ real*8 --- 31.5 146.9 42.4 Thus, adding the $8,000 FPA to your 750 can give a 60% (xf77 single precision) to 70% (xf77 double precision) improvement in execution time. Also note that for this compute bound floating point program, a 750 w/FPA gives about 65% (xf77 single precision) to 75% (xf77 double precision) of the performance of a 780 w/FPA. This is consistent with DEC's marketing claims. Tom Ferrin UCSF Computer Graphics Laboratory