Re: Intel to chip away at Itanium prices <- or ... I want my cheap

From: Rick Jones (foo_at_bar.baz.invalid)
Date: 01/22/04 

Date: Thu, 22 Jan 2004 22:42:02 GMT

Bill Todd <billtodd@metrocast.net> wrote:
> "Rick Jones" <foo@bar.baz.invalid> wrote in message
> news:LDVPb.12827$Cg5.6088@news.cpqcorp.net...
>> Any idea where the point of diminishing returns from cache size for
>> SPECint2000 might be?

> Pretty large, assuming that you measure returns geometrically (e.g.,
> by each doubling in size instead of per-MB). Itanic itself provides
> several data points with its different cache sizes (and they clearly
> help right up through 6 MB), and the single-processor IBM POWER4/4+
> results that dedicate the entire 128 MB of off-chip L3 to SPECint
> seem to show some benefit (though that cache is an order of
> magnitude slower than Itanic's on-chip cache, so the effect is far
> less pronounced).

So then it isn't simply just the size of the cache that matters, but
how quickly one can get to it that matters. Shame we don't have a
statistics hobbyist around to compute the correlations between
SPECint2000 score and processor frequency or cache size.

Using the configurable query form on www.spec.org to present the
SPECint2000 baseline scores for processors matching Itanium (sorry, no
matches on "Itanic" :) and then sorting by frequency as the primary,
baseline as the secondary, and then 2nd cache as the tertiary sort
(all descending) is intriguing. Factoring for compiler rev
differences and CECs is a bit beyond my Mk I eyeball at the moment,
but it does seem that the 8.0 vs 7.1 compiler makes a non-trivial
difference.

CINT2000

Company System Baseline MHz Other Ca Compiler Published
SGI SGI Altix 3000 1243 1500 6.0MB 8.0 Sep-2003
Bull NovaScale 4040 1217 1500 6MB 8 Dec-2003
Dell Dell PowerEdge 3 1099 1500 6MB 7.1 Aug-2003
Bull NovaScale 4040 1093 1500 6MB 7.1 Aug-2003
Supermic SS6113M-i 1090 1500 6MB 7.1 Jan-2004
SGI SGI Altix 3000 1077 1500 6.0MB 7.1 Jul-2003
SGI SGI Altix 350 986 1400 1.5MB 8.0 Jan-2004
ION Comp I2X2 (1.4GHz Ita 926 1400 4MB 7.1 Aug-2003
Dell Dell PowerEdge 3 824 1400 1.5MB 7.1 Sep-2003
SGI SGI Altix 3000 1019 1300 3.0MB 8.0 Jan-2004
SGI SGI Altix 3000 875 1300 3.0MB 7.1 Jul-2003
Bull NovaScale 4040 711 1000 3MB 7.0 Mar-2003
SGI SGI Altix 3000 683 1000 3.0MB 7.1 Jun-2003

SPECint being a trademark of SPEC, numbers being as of 1/22/04 from
www.spec.org, etc etc etc...

> My vague recollection is that *all* SPECint2K tests can fit into
> about a 200 MB footprint, so that would place an upper limit on
> effective cache size until a new SPEC suite with a larger footprint
> comes along.

SPECcpu2000 was intended to be runable in a system with at least 256MB
of RAM.

>> Out of curiousity, how many different chipsets and compilers
>> achieve the 3.2 GHz Xeon score?

> I didn't check, but Intel, Dell, and IBM (perhaps others) have Xeon

My check may have been too cursory, but I'm not seeing any SPECint2000
scores from IBM for x86 on www.spec.org - at least none for processors
of frequency > 2800 MHz which is what I used on the configurable query
form. Lots of stuff from Intel. Some stuff from Dell - don't they use
Intel chipsets? Stuff from HP.

They all seem to be using the same compiler, well, not all - more
below.

> SPECint_peak scores in the upper 1200s - only a few percent below
> the one that edged out the HP zx1 Itanic.

But of the systems from those vendors, how many are using different
compilers and chipsets?

>> If the 8.0 compiler also boosts x86, why haven't we seen it used for
>> x86 scores? Is the 8.0 stuff not around at Intel or the x86 vendors
>> for x86 yet?

> Beats me, unless it's simply because it would lift all boats and
> thus give no one any noticeable advantage (hence no great rush to
> adopt it until it's actually available). SGI, by contrast, needs
> all the boost it can get - and has what may be a leading chipset to
> show off to particular advantage (and Intel also benefits from
> anyone who makes Itanic look good, of course, so was presumably
> happy to make the future compiler available to SGI).

As it happens, in looking at www.spec.org to see what compilers were
used by those vendors you mentioned above, there are some SPECint2000
results for x86 using the 8.0 compiler:

http://www.spec.org/osg/cpu2000/results/res2003q4/cpu2000-20030922-02523.html
http://www.spec.org/osg/cpu2000/results/res2003q4/cpu2000-20030922-02521.html
http://www.spec.org/osg/cpu2000/results/res2003q4/cpu2000-20030922-02529.html

are some of them. So, it seems that bullet has been fired for x86.
I've not gone-in far enough to see what the relative delta from 7.1 to
8.0 on the different CPUs.

>> > There's a good reason that no single-processor Itanic TPC-C
>> > submissions exist, I suspect: they wouldn't be any better than
>> > Xeon's.
>>
>> Is that also why there are no single-processor POWER4+ TPC-C
>> submissions?

> IBM just doesn't put that much effort into TPC-C submissions for the
> POWER architecture these days, save at the high end (where they do
> so well that I'd *really* like to see a 4-processor 1.7 GHz system
> tested).

I'm sure there are many who would like to see lower-CPU count POWER4+
TPC-C scores.

> By contrast, Itanic suppliers seem more interested in pushing
> low-end Itanic systems

Could you expand on that a bit? Which Itanium vendors do you see as
being more interested in pushing low-end systems?

> - at least where they can be made to seem competitive (hence my
> suspicion above).

But you do not suspect that IBM may not be putting effort into
lower-end POWER4+ TPC-C because they may not be competitive?

rick jones 

-- 
portable adj, code that compiles under more than one compiler
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...
Loading