Red Flag announces an Intel-compiled distribution [LWN.net]

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 14:36 UTC (Wed) by pglennon (guest, #649)
In reply to: Red Flag announces an Intel-compiled distribution by dcoutts
Parent article: Red Flag announces an Intel-compiled distribution

Little bit of difference here, in that the intel compiler is *significantly* faster than gcc. I would guess ( without any sort of real evidence or attempt to mathematically justify or investigate ) that they would see a significant performance increase by using the intel compiler.

It is, however, very non-free, non-portable, and i wonder if they had to break very much to get it all working?

Be careful with the rabid cflag-tuning gentoo folks. They will get you, and then they will release images of your crumpled form via portage. You will 'emerge', but not victorious ;)

-P

to post comments

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 14:44 UTC (Wed) by pontus (guest, #3701) [Link] (8 responses)

Do you have any pointers to benchmarks? I keep seeing proprietary compilers advertised as generating ~ 20% faster code than GCC, but in the instance I managed to look up (Green Hills), the comparison was way old, from before the GCC 3.x release. In my experience, 3.x is significantly better than 2.95.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 14:59 UTC (Wed) by pglennon (guest, #649) [Link] (2 responses)

No, as I said, I try to remain unburdened by facts when I make pronouncments. I do recall that the benchmarks I saw were against 2.95, and that the theory was that the intel compilers made heavy heavy use of the P3 and aboves SSE registers ( floating point computations? anyone? ).

not sure if gcc 3.0 delves more into the SSE realm or not, but I remember that 2.95 largely did not.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 15:33 UTC (Wed) by ca9mbu (guest, #11098) [Link] (1 responses)

Richard Guenther is actively tracking performance regressions on mainline and the lno branch. See http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/mon[...] for details. This means that at least performance won't get worse than it is currently - well, at least not without the gcc hackers knowing about it quickly anyway :) As for recent ICC vs. GCC benchmarks, the most recent I could find was http://gcc.gnu.org/ml/gcc/2004-04/msg00913.html.

Cheers,

Matt.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 15:37 UTC (Wed) by ca9mbu (guest, #11098) [Link]

Sorry folks - even after previewing that last comment I managed to screw it up! That last benchmark I found was to do with compilation time, not for the performance of the resulting binaries. Both are of course important, but the latter is arguably more important/more noticeable to end users.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 15:11 UTC (Wed) by B.Tegge (subscriber, #5720) [Link] (4 responses)

German magazine c't published some comparisons using SPEC for x86-64 systems a couple of months ago. IIRC Intel C 8.1 still came out first compared to GCC 3.4, however I don't remember any specifics. If you're interested I'll look them up tonight. Older results (using icc 8.0 and gcc 3.3) are published here: http://www.heise.de/ct/Redaktion/as/spec/ct0404160/

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 15:32 UTC (Wed) by pontus (guest, #3701) [Link] (3 responses)

Thanks, very useful. It seems that the intel compiler is 10-15% faster than GCC in specint after all.

Red Flag announces an Intel-compiled distribution

Posted Aug 5, 2004 2:49 UTC (Thu) by djabsolut (guest, #12799) [Link] (1 responses)

I wonder how often this 10-15% speed increase would show up in real life situations. e.g. how often does one encode three MPEG2 videos and at the same time write a 1000 page report and view a webpage with 100 java applets? The only time this speed increase would approach significantness is when one constantly uses heavy number-crunching (e.g. scientific programs, like Folding@Home). For other applications, is this 10-15% speed gain worth the loss of freedom?

Red Flag announces an Intel-compiled distribution

Posted Aug 5, 2004 18:36 UTC (Thu) by neoprene (guest, #8520) [Link]

"I wonder how often this 10-15% speed increase would show up in real life situations. e.g. how often does one encode three MPEG2.... "

Running the kernel at 100% for hours and hours I often do while using dvd::rip/transcode or running the game AA.

Right, many desktop users have no need for a processor over 1GHz for their e-mail, web sufing and word-processing, but a great many are pushing the hardware to its limits, and upgrade often. A 15% increase in processor GHz can more than double the CPU-price, so its not insignificant. [check a P4 or Athlon64 at 3GHz and price out the 15% increase].

A 15% increase in frame-rate in AA and Transcode translates into a better game experience and quicker production times in transcode [while the machine is useless for anything else]. The game AA is more CPU than GPU bounded. Turning off all detail in textures and reducing screen rez does exactly nothing to FPS. [Running the K7 kernel helps increase the minimum frame-rate, perhaps due to 3D-Now and similar optimisations].

A faster compiled solution for these applications would be most welcome.

Red Flag announces an Intel-compiled distribution

Posted Aug 5, 2004 4:12 UTC (Thu) by JoeBuck (subscriber, #2330) [Link]

Intel tunes their compiler to do well on benchmarks, the GCC team does not. What this means is that, while there are definitely codes where Intel's compiler beats GCC by a good deal, you should not expect to see as much improvement in your own codes as the SPEC tests would suggest.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 15:07 UTC (Wed) by dcoutts (subscriber, #5387) [Link] (2 responses)

It's ok, I'm one of the sensible Gentoo folk. :-)

I can imagine a really good compiler will have an apreciable effect on some C code, but most system software isn't that CPU bound. I imagine there are some inner loops in the kernel that could benefit. But in general I'd be suprised if there were much benefit (just like those Gentoo CFLAGS).

Of course the kernel has no floating point which is where ICC has it's big advantage over GCC at the moment. I believe recent gcc's do make use of SSE floating point but not vectorising/SIMD like ICC can do.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 17:34 UTC (Wed) by stuart (subscriber, #623) [Link] (1 responses)

"It's ok, I'm one of the sensible Gentoo folk. :-)"

Surely no-one who uses Gentoo is sensible by definition. If you spend your life mucking around with CFLAGS, you're mad and if you run gentoo without mucking around with CFLAGS you're mad, because you might as well join the rest of us and use a binary distribution.

I have better things to do than watch things compile / contribute towards the great heat death.

Stu.

Red Flag announces an Intel-compiled distribution

Posted Aug 4, 2004 20:45 UTC (Wed) by flewellyn (subscriber, #5047) [Link]

Obviously, you don't know anything about what Gentoo is actually about. CFLAGS and compile settings are the least of it. Gentoo's system is about flexibility and configurability, through the ability to set compile-time flags on any package to use features you wish. Portage offers the ability to do quick "default" installs using your global settings, or override them for specific packages. You can tune the features of your system, not just the speed. Combined with a very easy to use interface, this makes installing packages set to your specifications quick and painless.

You can find more actual information at the Gentoo website. I suggest having facts before you comment.