[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference turris::fortran

Title:Digital Fortran
Notice:Read notes 1.* for important information
Moderator:QUARK::LIONEL
Created:Thu Jun 01 1995
Last Modified:Fri Jun 06 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:1333
Total number of notes:6734

1266.0. "Double and Signle precision performance issue" by NNTPD::"[email protected]" (poletti sergio) Mon Apr 21 1997 11:08

I'm executing some benchmarks against SGI Origin 2000.
Our system used during the benchmark is an AlphaServer 4100 5.466.

On Single Precision our performance are 20% better than SGI, on Double
Precision our performance drop of 40%.

On Double precision SGI drop only 20%.

Is this an Alpha ferature that differs to R10000 processors ?


 
[Posted by WWW Notes gateway]
T.RTitleUserPersonal
Name
DateLines
1266.1Probably memory bandwidth--not float precisionSTEVEN::hobbsSteven HobbsMon Apr 21 1997 11:3316
With the information given we can only guess at the performance
behavior you are seeing.

On all current Alpha systems, the time for add, sub and mul is the
same for single and double precision.  Single division is faster than
double division and most math library routines are faster for single.

However, the speed of computation is probably not responsible for the
performance differences you are seeing.  Double precision data is
twice as large as single precision data so a program calculating with
double precision data often needs double the memory bandwidth.  The SGI
Origin series has very good memory bandwidth so increasing the data set
size should improve the SGI performance relative to Alpha.

Future Alpha chips will have a big improvement in potential memory
bandwidth over existing chips.