large discrepancy in results on IA-64 vs Xeon

[Roberta_G_]

Hi,
One of my users has Fortran code compiled with the Intel Fortran Compilers on both an Itanium2 platform and a Xeon platform. The exact same code base gives different answers on each architecture. This has happened before (difference in the last two digits) and has been accepted due to the difference in architecutre. This time the difference is in the last three digits. What is the best way to diagnose this?
Many thanks,
Roberta

[TimP]

Last 3 digits doesn't tell much. Do you mean 5th through 7th digits for single precision?
If you wished to attempt to get close numerical results between Itanium and Xeon, you would use -fp-model precise -assume protect_parens,minus0 for compilation on both platforms, and assure that you are using an SSE option on Xeon.
With normal optimization (assuming SSE2 code for Xeon), results differ in particular when Itanium gets extra extra precision from fused multiply-add.