Hello,
I was running the Polyhedron 2005 Benchmark Suit with
g95, gfortran 4.1 (of SUSE 10.1), gfortran 4.2, ifort 9.0, ifort 9.1 and
NAGWare f95.
This suit consists of 16 quicker real-world number-crunching programs,
which run
7 to 130 seconds each. To make the data more reliably each test is run
several
times. (The complete set takes a bit more than 2 h per compiler.)
gfortran comes out pretty well. If one sets g95 (gcc 4.0.3) April 23,
2006 as 100%
gfortran 4.1 runs in 79% of that time and gfortran 4.2 (20060504) in 78%.
The ifort 9.0 is faster with 72%. ifort 9.1 speeded up a bit more to 69%.
The NAG f95 (seemingly with gcc 4.1 backend) reaches 84%.
The suit does not only measure the speed but also checks the
correctness. The
calculated values are checked whether they lay within a tolerance around
a reference
value; this is always the case for g95, ifort and NAGf95. However,
gfortran fails for two programs of the 16 programs (in one
of the five (four) subtests). The derivation is not huge, but still this
irritates me a bit.
(It is off by 36% in one case and by 1% in the other.)
I put the results to
http://www.physik.fu-berlin.de/~tburnus/benchmark/polyhedron05.html
In total, I'm rather happy with gfortran and would like to thank you for
creating this nice program.
I have the feeling that it has matured a lot during the last half year,
but maybe this is only because
almost the issues, I had, are now fixed :-)
Tobias