Intel compilers for Linux 5.0 beta
Ole W. Saastad
ole at scali.no
Wed Jul 4 23:58:17 PDT 2001
Intel has released the c and fortran compilers version 5.0 in
beta tests for Linux. At present they are out in 5.01 versions.
We have run some tests with the compilers and so far the overall
performance is very good. The fortran compiler is outstanding and
outperforms the others we have tested (Fujitsu, GNU and Portland).
The intel c compiler is less impressive than the fortran compiler.
The Intel compilers are lacking a few extensions that limit its
usefulness. It lacks the use of Cray pointers and byteswap io.
Both needed in some applications like the weather program mm5.
The speed by the fortran compiled code is not only due to efficient
use of SIMD (or vector processor codes), since I run the tests
without including the option to generate SSE code (it might be that
the Intel compiler does this anyway.).
I have have run some test to check how these compilers tackle
the test for vectorising compilers since the SIMD SSE unit is
a vector unit. The vector test shown that none of the compilers
gain much performance by selecting the vector code generator,
either for Portland nor Intel's fortran compiler.
Some results are published on our simple unofficial web site
computational-battery.org where we welcome contributions.
Fortran compiler benchmarks:
C compiler: (not yet on web site)
(Rate constants fitting Kinetic simulation and fit of rate constants
by means of "Simplex" minimization algorithm. Fourth order Runge-Kutta
and Simplex minimization (both from Numerical Recipes.
Tested on 800 MHz PIII 256 k cache.)
Gnu (egcs-2.91.66): 14.5 sec.
Gnu (gcc 2.95.2): 7.2 sec.
Portland (pgcc 3.1): 7.5 sec.
Fujitsu fcc (1.0): 10.9 sec.
Intel icc 5.0: 8.0 sec.
Ole W. Saastad, Dr.Scient.
Scali AS P.O.Box 70 Bogerud 0621 Oslo NORWAY
Tel:+47 22 62 89 68(dir) mailto:ole at scali.no http://www.scali.com
ScaMPI: bandwidth .gt. 220 MB/sec. latency .lt. 4 us.
More information about the Beowulf