Re: Intel C++ vs. gcc: New Numbers for New Versions

[Jan Hubicka <jh at suse dot cz>]

> Hello,
> 
> I've just posted preliminary numbers for gcc 3.2.1 and Intel C++ 7.0 at:
> 
> http://www.coyotegulch.com/reviews/intel_comp/intel_gcc_bench2.html
> 
> The number of requests for this information has been overwhelming. In spite
> of my desire to wait for more feedback from developers, I decided to post
> the current numbers with an eye toward making another update in a week or
> two.
> 
> Nothing much has changed, really, since the May comparison. People want to
> see the numbers, though -- so here they are!

There are supposed to be no changes in performance, as gcc 3.2.x is the
same as 3.1.x modulo the bugfixes in C++ ABI.
> 
> The lack of change is not surprising. GCC has been undergoing a refinement
> process involving ABI and regression patches; most of the new Intel features
> relate to issues that aren't covered by the benchmarks I can run
> (hyperthreading support, for instance, can only be tested on an HT-enabled
> processor, which I don't have.)
> 
> In many ways, comparing gcc and Intel C++ is an apples-and-oranges affair,
> given that each has strengths in different areas. How to compare Intel's
> support for OpenMP against gcc's portability across platforms, that is the
> question.
> 
> Note that gcc 3.3 is on the horizon (mid-December, I'm told), providing
> another opportunity to keep these numbers updated.

You may want to try current mainline tree.  It should be stable enought
for your testing and now there is chance to fix performance regressions
in the case some appeared.

Also did you experimented with profile feedback?  (ie
-fprofile-arcs/-fbranch-probabilities in GCC and -prof-gen/-prof-use in
ICC?)

Honza
> 
> I intend to write thorough reviews of both products in the next week or two.
> These complex and powerful tools need comprehensive analysis. For now, the
> benchmark numbers will satiate my readers (I hope!). Given the gcc schedule,
> I'm looking to write-up Intel C++ 7.0 sometime in the next week, followed by
> a gcc 3.3 review mid-December.
> 
> As always, I'm looking for any rational feedback on these articles.
> 
> Thanks.
> 
> Scott Robert Ladd
> Coyote Gulch Productions,  http://www.coyotegulch.com
> No ads -- just very free (and somewhat unusual) code.