🚀 go-pugleaf

Oh, geez, here we go again.

----- Original Message -----
From: "Hugh" <hugh@nospam.com>
Newsgroups:
comp.os.linux.hardware,gnu.gcc,comp.sys.ibm.pc.hardware.chips,comp.text.tex
Sent: Saturday, June 15, 2002 11:05 AM
Subject: P4 vs Athlon for Linux latexing, is gcc-2.95 responsible?


> The following is my general finding so far:
>
> Special features of Pentium 4 cpu is not generally utilized
> by gcc-2.95.** which is used for most Linux distributions rpms.
> Latexing a big document with Zeon Pentium 4 at 1.7GHz (52 sec) is actually
> even slower than Pentium III running at 700MHz giving 35 sec.
>
> On Athlon at 1.6GHz, it gives 16 sec.  The best.
>
> Is my understanding for the reason of slow speed with P4 correct?
> gcc-2.95.  If I do obtain te-latex.src.rpm and then compile it
> with gcc-3.1, would I get much faster binary under Pentium 4?
>
An even more interesting question would be whether Intel's own compiler,
icc, which for Linux is available free for personal use, would give better
results.  If anybody should be able to write a compiler to exploit the
strengths of the P4 and to minimize its weaknesses, it should be Intel.  At
least up until gcc 3.0, icc generally beat the pants off gcc.  Even RedHat,
which isn't known for being conservative about new releases, has yet to
bless and release a distribution of gcc 3.0, and if RedHat isn't ready to go
there, neither am I.

> The problem with Pentium 4 is not just limited with Latex.
> I in fact hate Pentium 4, because I simply believed in
> their high SPEC rating, and bought a bunch of Pentium 4 Xeon
> duals only to find out that the even the Fortran speed
> (with their Intel compiler) is in fact dismal.
>
In my still very brief study of the situation, I have found that Intel's own
compiler makes very conservative assumptions about possible vector (array)
dependence in loops and will often fail to vectorize loops that can, in
fact, be vectorized if you modify the code to make it explicit that a vector
dependency is not possible.

Without SSE2, which requires that loops be vectorized, a 1.7GHz P2 is a very
weak foating point performer, as many have noted, and as you have confirmed
for yourself.  If I take naive code compiled with icc, my overclocked
Celeron running at 1.5MHz will beat a 1.7GHz P4 every time.

If you're using a Fortran compiler, it almost certainly means you're doing
scientific computation, and that almost certainly means you have code full
of loops using pointers (arrays) that are just dying to be vectorized.  If
you're using single precision, SSE2 is worth a factor of four in throughput;
for single precision it's worth a factor of two, generally with lower
instruction latencies.

The fact that you have dual Xeon's is not likely to help much unless you
have specifically written software to exploit the availability of more than
one processor.  You may find it worthwhile to explore Intel's math kernel,
also free for personal use.

The design decisions behind the P4 seem mostly to have been driven by the
need to get multi-media performance, at which the processor excels.  It is
probably possible to turn the P4 into a decent number-cruncher, but not
without some work.  A processor that requires special code to get decent
performance is not especially attractive, and if the P4 weren't an Intel
product, no one would be using it.  As it is, the future of high-performance
personal computing looks like it belongs to the P4 for the forseeable
future, so we might as well get used to it.