New Intel icc compiler

Christophe Devine devine at iie.cnam.fr
Wed Jul 16 12:41:32 PDT 2003


"John Gay" <jgay at celestica.com> wrote:

> The first thing that caught my attention was a comment from CERN stating
> that they saw a 30% increase in run-time efficency over GCC-3.2

That's fairly close to the reality, at least in the field of block ciphers.
A few figures showing the speed of different implementations of Rijndael
(aka. AES), using gcc or icc on the same machine (P3 1.4 GHz) :

    * gcc 3.2 -O2 -march=pentium3

      ks (cycles/key) enc (cycles/byte)  ks (cycles/key) dec (cycles/byte)

devine       803( 1%)         31.7( 2%)         803( 1%)         32.1( 1%)
gladman      363( 2%)         37.0( 2%)         524( 6%)         40.3( 2%)
gnupg       1316( 3%)         43.2( 1%)         459( 1%)         43.3( 4%)
mks         2037( 8%)         29.9( 2%)        2285( 1%)         30.0( 1%)
openssl      319( 0%)         33.2( 3%)         847( 1%)         34.6( 3%)
tsd          697( 2%)         28.9( 4%)         701( 1%)         34.2( 1%)

        ks (keys/sec)   enc (bytes/sec)    ks (keys/sec)   dec (bytes/sec)

devine        1691252          43018503          1718977          43018503
gladman       3615779          36472209          2688656          34239216
gnupg         1059168          32263877          3177503          31655125
mks            718203          47934903           606113          45343827
openssl       4559026          43018503          1664406          39945752
tsd           1978445          47934903          2056031          39945752

    * icc 7.1 -O

      ks (cycles/key) enc (cycles/byte)  ks (cycles/key) dec (cycles/byte)

devine       867( 2%)         25.5( 7%)         866( 4%)         26.3( 0%)
gladman      223( 2%)         26.1( 1%)         359( 2%)         28.0( 2%)
gnupg       1313( 4%)         39.8( 1%)         432( 5%)         38.1( 1%)
mks         2197( 5%)         32.1( 1%)        2514( 5%)         32.1( 1%)
openssl      295( 2%)         29.2( 0%)         851( 3%)         28.4( 2%)
tsd          696( 2%)         28.6( 1%)         690( 2%)         29.1( 0%)

        ks (keys/sec)   enc (bytes/sec)    ks (keys/sec)   dec (bytes/sec)

devine        1542024          54120052          1565039          52428800
gladman       5825422          52428800          3679214          48629612
gnupg         1028016          34239216          3039351          35320455
mks            609637          42473965           539114          41943040
openssl       4462026          46603378          1600879          47259763
tsd           1889326          47259763          1923993          46603378


The percentage numbers in parenthesis are the deviation from
the average. The benchmark tool itself can be downloaded at:

http://www.cr0.net:8040/code/crypto/aesbench.tgz

-- 
Christophe Devine
-- 
Unsubscribe: send email to listar at linuxfromscratch.org
and put 'unsubscribe lfs-chat' in the subject header of the message



More information about the lfs-chat mailing list