[RC5] P4 Speed Question

Colin L. Hildinger colin at ionet.net
Thu Jun 14 21:26:26 EDT 2001


Another way of looking at this is that it's very likely that:

1) Intel knows that it takes a long time for compiler optimizations for a
given processor to make it into the mainstream.
2) Intel knows that if software doesn't become P4 optimized soon they are
going to continue losing market share (and worse than that, profit margins).
3) Intel has agreed to help Microsoft in every possible way with P4
optimizations, probably including paying for new programmers.

The fact is, even if the PIV optimizations in the latest compiler were
*superb*, most applications would still focus on P6 generation optimizations
for the next couple of years because companies want their software to run
well on the installed base of computers, not the installed base that will
exist in 3 years.  Some people do optimizations for multiple platforms.
These people include people like d.net, graphics card vendors, game
companies, multimedia API vendors, etc.  People writing productivity
software just aren't going to do it, and they're not going to use PIV
optimizations if it means that the majority of users with 500MHz PIII's will
actually suffer for it.

----- Original Message -----
From: "Kevin O'Gorman" <kogorman at pacbell.net>
To: <rc5 at lists.distributed.net>
Sent: Thursday, June 14, 2001 10:31 AM
Subject: Re: [RC5] P4 Speed Question


> Appropos of the P4 speed question, I got a call out of the blue today,
> from a guy saying he was a recruiter for Microsoft and Intel (they recruit
> jointly??? news to me!).  They're having trouble with the compiler
> optimization for P4 it seems (Intel's concern) and MS is coming out with
> a language called C# (C-sharp) which is supposed to be their answer to
> Java.  They both need help from folks with compiler smarts.  To hear this
> guy tell it, they're looking to hire _lots_ of help.
>
> I'm not available until I finish this bloody PhD, but I found this
> interesting, and worth passing on insofar as it reflects on the likelihood
> of things happening quickly to your local friendly neighborhood compiler.
> My guess is that this means that optimizing for the P4 is somewhat harder
> than Intel had supposed.  And that there's competition for the relevant
> talent.  I'm glad I found this out; I was about to go shopping for a
> server, and I'm now really likely to stay with a dual Celeron-A.  Or maybe
> P-III.  But the P4 can definitely wait.
>
> There's a whole lot more I would say about the idea of C#, but it's not
> relevant here, and I'll spare you.
>
> ++ kevin
>
>
> Ryan Malayter wrote:
> >
> > www.news.com reports a forthcoming DDR chipset from Intel for the P4,
code
> > named "Brookdale," Expected "early next year."
> >
> > But as you say, so the f*ck what... P4 performance sucks with
P3-optimized
> > code. But the P4 isn't all that bad of a CPU core... it's just a bad
choice
> > for the current Windows software market.
> >
> > The SPEC CPU2000 bechmarks, which are designed to resist hand-tuning,
but
> > allow for compiler instruction-order optimizations, show us some of the
P4
> > core's potential. (I know that there are still loopholes in SPEC's hand-
or
> > feedback- optimization rules. Manufacturers can exploit these to a
degree,
> > but CPU2000 is much stricter than CPU95 in this regard, and a new enough
> > benchmark that Intel/AMD probably haven't had time to really juice up
their
> > compilers with CPU2000-specific optimizations. In any case, benchmarks
like
> > SPEC are really the only way to compare two different CPU architectures
> > equitably).
> >
> >                                                 Cint200 Cfp2000
> > Athalon 1.4/PC2100 DDR SDRAM            554             458
> > Intel P4 1.4 Ghz/PC800 RDRAM            529             538
> >
> > As you can see, the P4 core fares well running P4-optimized code
> > clock-for-clock with the Athalon. (AMD chose to use Intel's p3-optimized
5.0
> > compilers, Intel used the same compiler with P4 flags set). These
benchmarks
> > were submitted by AMD and Intel themselves to SPEC, so you can bet they
> > represent the best both companies could muster for their platforms.
> >
> > Still, since most binary code today is P3-optimized, and the Athalon is
> > designed for this code base, that's what I'd buy next week. But if/when
MS
> > and other vendors start optimizing instruction order for the P4, you're
> > going to see the Athalon's current application-level speed advantage
> > disappear. Intel is still, after all, the bully in the market, and
they're
> > going to make P4 optimization pretty cheap and for most x86 shrink-wrap
> > developers.
> >
> > Somewhat ironically, it's the Linux folks - who can recompile most of
their
> > software to their liking - that might find they actually like
performance
> > (maybe not price) of the P4 right now. Assuming, of course, that these
Lunix
> > folks have access to good P4-optimized compilers and script
interpreters...
> >
> >         -ryan-
> >
> > -----Original Message-----
> > From: Andrzej Hamka³o [mailto:niktu at dcc.pl]
> > Sent: Wednesday, June 13, 2001 12:00 PM
> > To: rc5 at lists.distributed.net
> > Subject: RE: [RC5] P4 Speed Question
> >
> > > It is, the nForce chipset from nvidia introduces dual DDR sdram banks
with
> > > a combined bandwidth of more then the dual channel rdram that intel is
> > > pushing for the p4 and with much lower latency.
> > >
> > > On Wed, 13 Jun 2001, dan the person wrote:
> > >
> > > > Peter Cordes wrote:
> > > >
> > > > > On Tue, Jun 12, 2001 at 12:31:49AM +1200, dan the person wrote:
> > > > > > Only thing super about a p4 is it's super high price.
> > > > >
> > > > >  In their defence, they do have more memory bandwidth than
> > > > > any other PC
> > > > > hardware, with dual rambus channels feeding a 400MHz bus.
> > > > >
> > > >
> > > > Is this not more a feature of the motherboard chipset than the
> > > CPU design?
> >
> > Well, time to spam out of list subject a bit ....
> >
> > Processor-memory bandwith depends on motherboard (chipset, type of ram)
> > or processor (type of processor bus)
> > - every time you get LOWER value of both, slower componnent chokes other
...
> > (extra ram bandwith could be used ie. by graphic pocessor, or in propely
> > designed
> > dual system - Intel designs do not apply, there all procs share same bus
> > ...)
> >
> > Latencies (of memory, chipset and proccesor bus are additive - and
> > bigger latency is bad, FYI ;)
> >
> > and now short look back:
> > <SLB>
> > RDRAM (big latency, big bandwith) +
> > first incarnation Pentium III (100Mhz single pumed bus - rather slooow,
> > even newer 133 version don't catch up even close with RDRAM)
> > =
> > what you get?
> > you right!, weak points of both (big latency, low bandwith system)
> > - but you are proud owner of most expensive system in shop :))))))
> > - congrats to INTEL for that ingenious idea
> > </SLB>
> >
> > and now present:
> > best of best list:
> > DDR RAM - highest badwith, lowest latency, sane price
> > nForce chipset - make that bandwith twice :), one but:
> > you can use it up for bulitin geforce MX ...
> >
> > P4 100*4 bus - nice thingie, but besides bus there are other thingie
> > in processor to watch not to screw up ...
> >
> > AMD's fastest Proc bus is 133*2
> > (if only AMD would launch 266*2 procs :)
> >
> > For accuracy: I know, width of bus are important too (took that under
> > consideration, thats why i won't say that RDRAM's  are fastest :)
> >
> > <rant>
> >
> > But Mhz are THE MOST IMPORTANT, right?
> > Then people buy 1,7GHz P4 system and are suprised ...
> >
> > I bet, the same money could buy you dual Athlon DDR system ...
> > (if not 1,4Ghz kind, certainly 1,33, but considering prices of RDRAM,
> > and i850 motherboard i think you could even save on dual 1,4Ghz  :)
> > you would be suprised, this time rather positively ...
> >
> > Most people don't get it that even single 1,4Ghz Athlon performs better
than
> > 1,7GHz P4 ...
> > especially on code unoptimized for P4 (currently? - about 98%)
> >
> > Then, most programs that ARE optimized for P4, (or for some strange
reason,
> > run better :) can take advantage of second processor (2 in price of 1,
> > remember?)
> >
> > I will not compare cow speed on these systems ... (quick guess: 4:1 in
favor
> > AMD,
> > somebodybdy check that and precise it out), because rarely somebody buys
> > system only to run dnetc.
> >
> > I can't think of application that would justify P4 existance in current
form
> > ...
> >       ... then again you could run Quake III under win ME  :)))
(scratch
> > one Athlon :))))
> >
> > Now you know why Intel launched 1,3GHz cheepo P4. If somebody wants
> > develop and optimize his apps for P4, you can't force poor guy to waste
too
> > much
> > money :)
> >
> > </rant>
> >
> > Always wanted to write that:
> > Dear Intel, please make good motherboard chipset at last (DDR?, hoping),
if
> > Via won't have good
> > concurency, they won't improve their chipsets as much... (that nasty
latency
> > thing ...)
> > BX owner
> >
> > PS. I didn't bored you to death?, did you really read all of this?, wow
:)
> >
> > PPS. excuse my keyboard, it sometimes can't spell propely ;) and
> > excuse me for being blunt, just got fed up with today exam, and had to
let
> > off steam ...
> >
> > PPPS. Weird, all my essays always came out shorther than this ...
> >
> > --
> > To unsubscribe, send 'unsubscribe rc5' to
majordomo at lists.distributed.net
> > rc5-digest subscribers replace rc5 with rc5-digest
> > --
> > To unsubscribe, send 'unsubscribe rc5' to
majordomo at lists.distributed.net
> > rc5-digest subscribers replace rc5 with rc5-digest
>
> --
> Kevin O'Gorman  (805) 650-6274  mailto:kevin at kosmanor.com
> Permanent e-mail forwarder:  mailto:Kevin.O'Gorman.64 at Alum.Dartmouth.org
> At school: mailto:kogorman at cs.ucsb.edu
> Web: http://www.cs.ucsb.edu/~kogorman/index.html
> Web: http://kosmanor.com/~kevin/index.html
>
> "In theory, there's no difference between theory and practice
>     in practice, there nearly always is."
> --
> To unsubscribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
> rc5-digest subscribers replace rc5 with rc5-digest
>
>

--
To unsubscribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
rc5-digest subscribers replace rc5 with rc5-digest



More information about the rc5 mailing list