Dual Xeon Linux vs. dual G5 : Where can I find direct speed comparison?-Mac Usenet

digitaleon

06-29-2003, 07:47 AM

To comp.os.linux, comp.sys.mac.advocacy and comp.sys.mac.system
subscribers,

> From: Jim Kroger <jimkkREMOVEME[at]umich.edu>,

> It it hype or has apple blown away the x86 world?
>
> Any pointers most appreciated.

Aside from the hotly-debated SPEC performance comparisons on Apple's
site and the not-so-hotly-debated application performance comparisons
at the same location, nowhere as yet.

Apple has been quoting CPU2000 SPECint and CPU2000 SPECfp benchmark
results on their site [at] www.apple.com/powermac/performance. These
benchmarks have been both criticised as being misleading and hailed as
being _more_ accurate than the benchmarks from Dell, in both cases for
the same reasons and with good points all around.

As yet, Apple's results have not been submitted to SPEC.

During the WWDC keynote, five different applications were demonstrated
on both the PowerMac and the Xeon (which had Windows XP, rather than
RedHat Linux) in which the PowerMac put the boot through the Xeon in
performance, and application-specific benchmarks are also available on
Apple's site. Most of these utilise FP and SIMD processing more than
integer processing, and of course score higher results on the G5.

The performance of these applications was backed up by representatives
from the vendors whose applications these were, which lends some
legitimacy to them, in addition to the fact that SPEC benchmarks have,
at least in my eyes, proven themselves to be not unreliable, but not
reliable either.

When results come in, we will doubtlessly be the first to hear of them.

digitaleon.

Heywood Mogroot

06-29-2003, 12:13 PM

digitaleon <this.is[at]fake.address> wrote in message news:<Vq2dnSG9VNzzFmOjRTvU2Q[at]giganews.com>...
> To comp.os.linux, comp.sys.mac.advocacy and comp.sys.mac.system
> subscribers,
>
> > From: Jim Kroger <jimkkREMOVEME[at]umich.edu>,
>
> > It it hype or has apple blown away the x86 world?
> >
> > Any pointers most appreciated.
>
> Aside from the hotly-debated SPEC performance comparisons on Apple's
> site and the not-so-hotly-debated application performance comparisons
> at the same location, nowhere as yet.

yes, the footnote to the "world's fastest PC" claim is on SPEC2000
*and* "professional apps".

=Heywood=

Heywood Mogroot

06-29-2003, 12:58 PM

Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message news:<jimkkREMOVEME-0B7A01.23203128062003[at]visonmassif.rs.itd.umich.edu>...
> It it hype or has apple blown away the x86 world?

Both!

> Any pointers most appreciated.

What you get with the 1.8Ghz tower:

Features common with Intel's latest & greatest desktop motherboard:
o Dual-channel PC-3200 DRAM (DDR-400)
o 1Gbit onboard ethernet
o 8x AGP
o Dual Serial ATA (not sure what RAID levels are supported)
o Onboard 5.1 audio with optical in/out

Features in common with AMD's Opteron platform:

o 64-bit CPU architecture / expandable over 4GB
o PCI-X

New Features not available on x86 yet:

o Dedicated 900Mhz (6.4GB/s) HyperTransport interface between CPU and
the memory controller (the current leading intel desktop, the 875P,
manages just 4.7GB/s in real bandwidth). For dual G5 systems, *each*
CPU has its own interface to the memory controller, unlike the dual
Xeon systems, which have to share a single 533Mhz FSB interface.

o Onboard support for 802.11g, 1394b

As far as benchmarks outside of the keynote, there was this one:

First 10 frames of Adobe After Effects "Night Flight" script:

Dual 2.66 GHz Pentium Xeon: 1.2 min/frame
Dual 2.0 Ghz G5: 0.6 min/frame

www.macintouch.com/g5reader.html#jun27

While we'll know more about the relative performance of the new G5's,
the one comforting thing is that Intel itself really doesn't have
anything lined up, motherboard-wise, to compete with it until sometime
next year.

(this is something of a rarity to catch Intel so flat-footed;
traditionally they have maintained a 6-12 month lead on Apple).

AMD, on the other hand, will soon have competitive systems in the
workstation market segment, eg. from the NVIDIA nForce3 boards, and
from its "Clawhammer" CPU launch later this summer.

=Heywood=

Seebs

06-29-2003, 05:56 PM

In article <flippo-E843C3.06313729062003[at]news.central.cox.net>,
flip <flippo[at]mac.com> wrote:
>That is, unless you're willing to believe the Presidents of Adobe,
>Wolfram, Luxology, and several other apps who presided over side-by-side
>tests where the G5 kicked a P4's butt halfway to the moon.

Wolfram I would believe. Adobe? You must be kidding me; believing something
because Adobe said it would be just plain stupid.

-s
--
Copyright 2003, all wrongs reversed. Peter Seebach / seebs[at]plethora.net
http://www.seebs.net/log/ - YA blog. http://www.seebs.net/ - homepage.
C/Unix wizard, pro-commerce radical, spam fighter. Boycott Spamazon!
Consulting, computers, web hosting, and shell access: http://www.plethora.net/

flip

06-29-2003, 07:07 PM

In article <seeker1-929275.10415329062003[at]news.comcast.giganews.com>,
Seeker1 <seeker1[at]mac.com> wrote:

> > During the WWDC keynote, five different applications were demonstrated
> > on both the PowerMac and the Xeon (which had Windows XP, rather than
> > RedHat Linux) in which the PowerMac put the boot through the Xeon in
> > performance, and application-specific benchmarks are also available on
> > Apple's site. Most of these utilise FP and SIMD processing more than
> > integer processing, and of course score higher results on the G5.
> >
> > The performance of these applications was backed up by representatives
> > from the vendors whose applications these were, which lends some
> > legitimacy to them, in addition to the fact that SPEC benchmarks have,
> > at least in my eyes, proven themselves to be not unreliable, but not
> > reliable either.
>
> My guess would be that Apple chose five applications that were:
>
> 1. floating-point intensive (since the G5 does much better on that than
> integer)

Mathematica? Yes.

Photoshop? I don't think so. I think it's all integer.

> 2. highly SMP-aware (since this is key for their duals)

True.

> 3. very Altivec-enabled (since this is where the G5 shines)

True.

> 4. very memory-intensive (since now the shoe is on the other foot, and
> Apple is actually AHEAD in memory access speed)

Not necessarily. Mathematica is more computation intensive than memory
intensive, I would suspect.

>
> So, I don't think there was anything misleading about that part of the
> presentation, but one should realize they cherry-picked applications
> where they knew the G5 would shine. I don't view that as a distortion,
> since many people use those applications. However, not all applications
> will show this level of benefit.

That's true. HOWEVER, the items above describe exactly the way that
CPU-intensive apps should be coded. Anyone with an app that could use FP
but who tries to do it in INT should be shot. Anyone with a CPU
intensive app who doesn't make it SMP-aware (multithreaded should be
sufficient) and use Altivec where appropriate doesn't deserve to be in
the business.

Fortunately, the majority of very CPU-intensive apps meet those criteria
- on the Mac, at least.

Heywood Mogroot

06-29-2003, 08:28 PM

flip <flippo[at]mac.com> wrote in message news:<flippo-E843C3.06313729062003[at]news.central.cox.net>...
> In article <CptLa.126$1H5.2814[at]newsfeed.avtel.net>,
> "John" <nospam[at]nospam.com> wrote:
>
> > Jim Kroger wrote:
> > > It it hype or has apple blown away the x86 world?
> >
> > The usual hype. When independent testing is done in a couple of months
> > the world will know that Apple once again pulled one.
> >
> >
> >
>
> That is, unless you're willing to believe the Presidents of Adobe,
> Wolfram, Luxology, and several other apps who presided over side-by-side
> tests where the G5 kicked a P4's butt halfway to the moon.

I'm not entirely convinced that the dual Xeon systems were the fastest
P4 desktop systems available last week for those particular apps.

A 3.0 GHz P4 with its 800Mhz FSB might be somewhat more performant on
those bandwidth-intensive apps (I get 4.7GB/s with the SSE2 tests on
this Canterwood system).

Still, 4.7GB/s is pretty shitty compared to the dual 2.0Ghz's 16.0GB/s
theoretical peak b/w. 3x the b/w for just twice the price!

Plus we Apple proponents shouldn't ignore AMD's new Opteron platform,
even if everybody in x86 land apparently is.

=Heywood=

foo

06-29-2003, 08:44 PM

On 29 Jun 2003 12:28:56 -0700, imouttahere[at]mac.com (Heywood Mogroot)
wrote:

>Still, 4.7GB/s is pretty shitty compared to the dual 2.0Ghz's 16.0GB/s
>theoretical peak b/w. 3x the b/w for just twice the price!

You get 4.7GB/s measured performance on the PC. What's your measured
performance on the G5?

Oh. I thought so. Well, let us know when you can get a number -
until then, "theory" isn't very interesting.

Steve Bellenot

06-29-2003, 09:42 PM

Vareck Bostrom

06-29-2003, 10:28 PM

In article <flippo-5B6E1C.13032629062003[at]news.central.cox.net>,
flip <flippo[at]mac.com> wrote:

> In article <seeker1-929275.10415329062003[at]news.comcast.giganews.com>,
> Seeker1 <seeker1[at]mac.com> wrote:
>
> > > During the WWDC keynote, five different applications were demonstrated
> > > on both the PowerMac and the Xeon (which had Windows XP, rather than
> > > RedHat Linux) in which the PowerMac put the boot through the Xeon in
> > > performance, and application-specific benchmarks are also available on
> > > Apple's site. Most of these utilise FP and SIMD processing more than
> > > integer processing, and of course score higher results on the G5.
> > >
> > > The performance of these applications was backed up by representatives
> > > from the vendors whose applications these were, which lends some
> > > legitimacy to them, in addition to the fact that SPEC benchmarks have,
> > > at least in my eyes, proven themselves to be not unreliable, but not
> > > reliable either.
> >
> > My guess would be that Apple chose five applications that were:
> >
> > 1. floating-point intensive (since the G5 does much better on that than
> > integer)
>
> Mathematica? Yes.

Mathematica tends to need to do computations to arbitrary precision and
as such will most likely be using integer math to do it. The strong
showing by Mathematica is particularly impressive to me.

>
> Photoshop? I don't think so. I think it's all integer.
>
> > 2. highly SMP-aware (since this is key for their duals)
>
> True.
>
> > 3. very Altivec-enabled (since this is where the G5 shines)
>
> True.
>
> > 4. very memory-intensive (since now the shoe is on the other foot, and
> > Apple is actually AHEAD in memory access speed)
>
> Not necessarily. Mathematica is more computation intensive than memory
> intensive, I would suspect.
>
> >
> > So, I don't think there was anything misleading about that part of the
> > presentation, but one should realize they cherry-picked applications
> > where they knew the G5 would shine. I don't view that as a distortion,
> > since many people use those applications. However, not all applications
> > will show this level of benefit.
>
> That's true. HOWEVER, the items above describe exactly the way that
> CPU-intensive apps should be coded. Anyone with an app that could use FP
> but who tries to do it in INT should be shot. Anyone with a CPU
> intensive app who doesn't make it SMP-aware (multithreaded should be
> sufficient) and use Altivec where appropriate doesn't deserve to be in
> the business.
>
> Fortunately, the majority of very CPU-intensive apps meet those criteria
> - on the Mac, at least.

Jerry Gardner

06-29-2003, 10:39 PM

Arturo Pérez wrote:

> Well, one could argue that only the above type of applications need the
> benefit. After all, the current shipping PowerMacs are more than capable
> of supporting applications that (to negate the above).

These types of demos/benchmarks always seem to focus on application
performance, and the applications chosen always seem to be the same.

What I never see is a comparison of the relative performance of the two
operating systems in question here: OS X and Windows XP. After all, we
spend a lot of time interacting with the operating system during the course
of our work with our machines, so why not compare the performance of the OS
as well?

Probably the reason this isn't done is because such comparisons are more
subjective. With something like Photoshop, one can say that the G5 applied
a certain filter in x seconds, while the dual Xeon did it in y seconds.
Hence, a direct comparison is possible. With an OS, this isn't as easy to
measure.

In my experience, OS X is nowhere near as responsive as Windows XP (*). I
have nothing against the Mac, in fact I own an 867 MHz Quicksilver G4. I
wish OS X were as responsive as XP, but just looking at the way OS X is
built, I can see why it's so sluggish.

The way Quartz composites 2D, 3D, and media to a composite buffer, runs it
all through the OpenGL pipeline, flattens it, and finally writes it to the
video card is time consuming. The 2D stuff all seems to be based on a PDF
imaging model, which I'm sure adds to the overhead. Even Quartz Extreme,
which optimizes this by doing some of the work on the graphics card, still
eats up processor time. All this results in slow scrolling windows, slow
window opens and closes, etc.

(*) My WindowsXP system running on a 400MHz P3 is much more responsive than
my 867 MHz G4 running OS X 10.2.6.

--
Jerry Gardner
jg2-usenet[at]gardnerclan.net

Seebs

06-29-2003, 11:21 PM

In article <flippo-5B6E1C.13032629062003[at]news.central.cox.net>,
flip <flippo[at]mac.com> wrote:
>That's true. HOWEVER, the items above describe exactly the way that
>CPU-intensive apps should be coded. Anyone with an app that could use FP
>but who tries to do it in INT should be shot.

How quickly times change!

>Anyone with a CPU
>intensive app who doesn't make it SMP-aware (multithreaded should be
>sufficient) and use Altivec where appropriate doesn't deserve to be in
>the business.

Multithreading doesn't necessarily buy you much for SMP. Many CPU-intensive
apps simply don't parallelize well.

Note also that many CPU-intensive apps can't use Altivec, because Altivec
is low-precision.

-s
--
Copyright 2003, all wrongs reversed. Peter Seebach / seebs[at]plethora.net
http://www.seebs.net/log/ - YA blog. http://www.seebs.net/ - homepage.
C/Unix wizard, pro-commerce radical, spam fighter. Boycott Spamazon!
Consulting, computers, web hosting, and shell access: http://www.plethora.net/

David Wang

06-29-2003, 11:31 PM

In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
> Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message news:<jimkkREMOVEME-0B7A01.23203128062003[at]visonmassif.rs.itd.umich.edu>...
>> It it hype or has apple blown away the x86 world?

> New Features not available on x86 yet:

> o Dedicated 900Mhz (6.4GB/s) HyperTransport interface between CPU and
> the memory controller (the current leading intel desktop, the 875P,
> manages just 4.7GB/s in real bandwidth). For dual G5 systems, *each*
> CPU has its own interface to the memory controller, unlike the dual
> Xeon systems, which have to share a single 533Mhz FSB interface.

Please check again. The hypertransport interface is not between the
CPU and the memory controller. The hypertransport interface connects
between the memory/system controller and a quasi south-bridge I/O
controller.

The 6.4 GB/s is a "peak" data rate, if either the 875P or the PPC970's
companion chip can support 4.7 GB/s sustained bandwidth, then that would
be very impressive. However, I have not seen anything close to that
figure when using McCalpin's STREAM.

--
davewang202(at)yahoo(dot)com

Vareck Bostrom

06-29-2003, 11:34 PM

In article <bdnpbd$cho$1[at]grapevine.wam.umd.edu>,
David Wang <foo[at]bar.invalid> wrote:

> In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
> > Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message
> > news:<jimkkREMOVEME-0B7A01.23203128062003[at]visonmassif.rs.itd.umich.edu>...
> >> It it hype or has apple blown away the x86 world?
>
> > New Features not available on x86 yet:
>
> > o Dedicated 900Mhz (6.4GB/s) HyperTransport interface between CPU and
> > the memory controller (the current leading intel desktop, the 875P,
> > manages just 4.7GB/s in real bandwidth). For dual G5 systems, *each*
> > CPU has its own interface to the memory controller, unlike the dual
> > Xeon systems, which have to share a single 533Mhz FSB interface.
>
> Please check again. The hypertransport interface is not between the
> CPU and the memory controller. The hypertransport interface connects
> between the memory/system controller and a quasi south-bridge I/O
> controller.

What is the bus between the CPU and memory controller?

>
> The 6.4 GB/s is a "peak" data rate, if either the 875P or the PPC970's
> companion chip can support 4.7 GB/s sustained bandwidth, then that would
> be very impressive. However, I have not seen anything close to that
> figure when using McCalpin's STREAM.

Erick Bryce Wong

06-29-2003, 11:51 PM

Steve Bellenot <bellenot[at]math.fsu.edu> wrote:
>flip <flippo[at]mac.com> wrote:
>> Seeker1 <seeker1[at]mac.com> wrote:
>>> My guess would be that Apple chose five applications that were:
>
>>> 4. very memory-intensive (since now the shoe is on the other foot, and
>>> Apple is actually AHEAD in memory access speed)
>>
>>Not necessarily. Mathematica is more computation intensive than memory
>>intensive, I would suspect.
>
>Symbolic computations can/often have immediate calculations which require
>lots of memory. A trivial example would be computing a binomial
>coefficient 200 choose 100 = 200!/(100!)^2 by multiplying the factorials
>out and doing the long division. By the way, this is also an example
>of an integer calculation in Mathematica. The stuff below is maple,
>but the idea is the same:
>> binomial(200,100);
> 90548514656103281165404177077484163874504589675413336841320

Point taken, but kind of a bad example :). Maple, like any sensible system,
doesn't compute the factorials for integer binomial coefficients. Instead,
it computes binomial(200,i) incrementally for i from 1 to 100, each time
multiplying by a single rational (using iquo to avoid unnecessary gcd
computations). So at no stage of binomial(200,100) does the intermediate
result exceed the answer :).

-- Erick

Jim Kroger

06-30-2003, 12:43 AM

In article <bdnqgo$eo4$1[at]morgoth.sfu.ca>,
erick[at]sfu.ca (Erick Bryce Wong) wrote:

> Steve Bellenot <bellenot[at]math.fsu.edu> wrote:
> >flip <flippo[at]mac.com> wrote:
> >> Seeker1 <seeker1[at]mac.com> wrote:
> >>> My guess would be that Apple chose five applications that were:
> >
> >>> 4. very memory-intensive (since now the shoe is on the other foot, and
> >>> Apple is actually AHEAD in memory access speed)
> >>
> >>Not necessarily. Mathematica is more computation intensive than memory
> >>intensive, I would suspect.
> >
> >Symbolic computations can/often have immediate calculations which require
> >lots of memory. A trivial example would be computing a binomial
> >coefficient 200 choose 100 = 200!/(100!)^2 by multiplying the factorials
> >out and doing the long division. By the way, this is also an example
> >of an integer calculation in Mathematica. The stuff below is maple,
> >but the idea is the same:
> >> binomial(200,100);
> > 90548514656103281165404177077484163874504589675413336841320
>
> Point taken, but kind of a bad example :). Maple, like any sensible system,
> doesn't compute the factorials for integer binomial coefficients. Instead,
> it computes binomial(200,i) incrementally for i from 1 to 100, each time
> multiplying by a single rational (using iquo to avoid unnecessary gcd
> computations). So at no stage of binomial(200,100) does the intermediate
> result exceed the answer :).
>
> -- Erick

What?

Jim Kroger

06-30-2003, 01:24 AM

In article <dd5de929.0306291128.4f394084[at]posting.google.com>,
imouttahere[at]mac.com (Heywood Mogroot) wrote:

> Plus we Apple proponents shouldn't ignore AMD's new Opteron platform,
> even if everybody in x86 land apparently is.
>
> =Heywood=

And that's very curious to me.

Jim

Heywood Mogroot

06-30-2003, 02:39 AM

David Wang <foo[at]bar.invalid> wrote in message news:<bdnpbd$cho$1[at]grapevine.wam.umd.edu>...
> In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
> > Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message news:<jimkkREMOVEME-0B7A01.23203128062003[at]visonmassif.rs.itd.umich.edu>...
> >> It it hype or has apple blown away the x86 world?
>
> > New Features not available on x86 yet:
>
> > o Dedicated 900Mhz (6.4GB/s) HyperTransport interface between CPU and
> > the memory controller (the current leading intel desktop, the 875P,
> > manages just 4.7GB/s in real bandwidth). For dual G5 systems, *each*
> > CPU has its own interface to the memory controller, unlike the dual
> > Xeon systems, which have to share a single 533Mhz FSB interface.
>
> Please check again. The hypertransport interface is not between the
> CPU and the memory controller.

yes, this makes more sense. Apple does list a 64-bit "bi-directional"
/ dual 32-bit unidirectional DDR FSB running at 1/4 the CPU speed.

This sounds like IBM's Elastic I/O...

> The hypertransport interface connects
> between the memory/system controller and a quasi south-bridge I/O
> controller.

yes, this makes more sense, very little PPC-specific stuff once you
get to the memory controller.

> The 6.4 GB/s is a "peak" data rate, if either the 875P or the PPC970's
> companion chip can support 4.7 GB/s sustained bandwidth, then that would
> be very impressive. However, I have not seen anything close to that
> figure when using McCalpin's STREAM.

I was using SiSandra's memory bandwidth tests, which uses SSE2
apparently.

=Heywood=

David Wang

06-30-2003, 02:58 AM

In comp.sys.mac.advocacy Vareck Bostrom <bostrov[at]mac.com> wrote:
> David Wang <foo[at]bar.invalid> wrote:

>> In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
>> > Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message
>> > news:<jimkkREMOVEME-0B7A01.23203128062003[at]visonmassif.rs.itd.umich.edu>...
>> >> It it hype or has apple blown away the x86 world?
>>
>> > New Features not available on x86 yet:
>>
>> > o Dedicated 900Mhz (6.4GB/s) HyperTransport interface between CPU and
>> > the memory controller (the current leading intel desktop, the 875P,
>> > manages just 4.7GB/s in real bandwidth). For dual G5 systems, *each*
>> > CPU has its own interface to the memory controller, unlike the dual
>> > Xeon systems, which have to share a single 533Mhz FSB interface.
>>
>> Please check again. The hypertransport interface is not between the
>> CPU and the memory controller. The hypertransport interface connects
>> between the memory/system controller and a quasi south-bridge I/O
>> controller.

> What is the bus between the CPU and memory controller?

Elastic I/O, inherited from the Power4 lineage. A wavepipelined
interface with built in skew cancellation circuitry.

Some information may be found here in an article I wrote right
after Microprocessor Forum. I meant to do a part 2, but the
numerous delays means that I would end up doing a post release
product analysis rather than a pre-release technical analysis
of what the PPC970 sustem architecture must look like.

http://www.realworldtech.com/page.cfm?ArticleID=RWT011303183140

Hypertransport is more of a peripheral to peripheral interconnect,
except that AMD added the cache coherency aspect to it, so Opteron
actually uses ccHT, a cache coherent version of Hypertransport.

Apple has a PowerMac G5 "technology overview" that describes the role
that Hypertransport plays in its system. It seems clear that
Hypretransport is used as the the connection from the northbridge
to the rest of the world.

http://a352.g.akamai.net/7/352/51/e93ca6b90038b4/www.apple.com/powermac/pdf/PowerMacG5_TO_062303.pdf

Page 11.

--
davewang202(at)yahoo(dot)com

David Wang

06-30-2003, 03:30 AM

In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
> David Wang <foo[at]bar.invalid> wrote in message news:<bdnpbd$cho$1[at]grapevine.wam.umd.edu>...
>> In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:

>> > New Features not available on x86 yet:

>> > o Dedicated 900Mhz (6.4GB/s) HyperTransport interface between CPU and
>> > the memory controller (the current leading intel desktop, the 875P,
>> > manages just 4.7GB/s in real bandwidth). For dual G5 systems, *each*
>> > CPU has its own interface to the memory controller, unlike the dual
>> > Xeon systems, which have to share a single 533Mhz FSB interface.
>>
>> Please check again. The hypertransport interface is not between the
>> CPU and the memory controller.

> yes, this makes more sense. Apple does list a 64-bit "bi-directional"
> / dual 32-bit unidirectional DDR FSB running at 1/4 the CPU speed.

> This sounds like IBM's Elastic I/O...

It is Elastic I/O.

http://www.realworldtech.com/page.cfm?ArticleID=RWT011303183140

I did refer to "900 MHz" in the article, which ofcourse should be
900 mbps. 1/9 of the bits goes to the packet protocol overhead.

"64 bit bi-directional" bus? I'm not sure what's 64 bit and
bidirectional in there. The PCI-X bus? It's a bit removed from
the CPU interface...

>> The hypertransport interface connects
>> between the memory/system controller and a quasi south-bridge I/O
>> controller.

> yes, this makes more sense, very little PPC-specific stuff once you
> get to the memory controller.

Apple could *almost* use stock AMD64/Opteron controllers, except
AMD64/Opteron needs an AGP port in its southbridge chipset (the
CPU contains the AGP-less northbridge)

It would be interesting if Apple did use an AMD chipset later.
There'd be 2 AGP ports. One in the northbridge, one in the
southbridge.

>> The 6.4 GB/s is a "peak" data rate, if either the 875P or the PPC970's
>> companion chip can support 4.7 GB/s sustained bandwidth, then that would
>> be very impressive. However, I have not seen anything close to that
>> figure when using McCalpin's STREAM.

> I was using SiSandra's memory bandwidth tests, which uses SSE2
> apparently.

It would be very difficult for a PPC970 system to match that bandwidth,
since much of that "4.7 GB/s" is probably just streaming prefetched
reads. PPC970's read bandwidth is capped at 1/2 of the total bandwidth,
since the ports are unidirectional.

I haven't seen any 875P results for McCalpin's stream. I am guesssing
that it would report somewhere between 3.2 and 3.6 GB/s. If you can
run the stock STREAM, and report the results to McCalpin, you can
be (in)famous.

www.cs.virginia.edu/stream/

--
davewang202(at)yahoo(dot)com

Jim Kroger

06-30-2003, 04:45 AM

In article <bdo5h2$gvk$1[at]grapevine.wam.umd.edu>,
David Wang <foo[at]bar.invalid> wrote:

> Some information may be found here in an article I wrote right
> after Microprocessor Forum. I meant to do a part 2, but the
> numerous delays means that I would end up doing a post release
> product analysis rather than a pre-release technical analysis
> of what the PPC970 sustem architecture must look like.
>
> http://www.realworldtech.com/page.cfm?ArticleID=RWT011303183140
>

Holy cpu!......you know some stuff.

Do you have an opinion on the relative power of the dual G5 and the dual
Opteron?

Thanks
Jim

Heywood Mogroot

06-30-2003, 08:43 AM

David Wang <foo[at]bar.invalid> wrote in message news:<bdo7bf$hld$1[at]grapevine.wam.umd.edu>...
> In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
> > David Wang <foo[at]bar.invalid> wrote in message news:<bdnpbd$cho$1[at]grapevine.wam.umd.edu>...
> >> In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
>
> >> The 6.4 GB/s is a "peak" data rate, if either the 875P or the PPC970's
> >> companion chip can support 4.7 GB/s sustained bandwidth, then that would
> >> be very impressive. However, I have not seen anything close to that
> >> figure when using McCalpin's STREAM.
>
> > I was using SiSandra's memory bandwidth tests, which uses SSE2
> > apparently.
>
> It would be very difficult for a PPC970 system to match that bandwidth,
> since much of that "4.7 GB/s" is probably just streaming prefetched
> reads. PPC970's read bandwidth is capped at 1/2 of the total bandwidth,
> since the ports are unidirectional.
>
> I haven't seen any 875P results for McCalpin's stream. I am guesssing
> that it would report somewhere between 3.2 and 3.6 GB/s. If you can
> run the stock STREAM, and report the results to McCalpin, you can
> be (in)famous.

wstream.exe 2.6C P4 [at] 2.6GHz, 800Mhz FSB DDR400 (1:1) 2225 memory
timings
2.748, 2.726, 2.950, 2.934 GB/s

wstream.exe 2.6C P4 [at] 3.25Ghz, 1000Mhz FSB DDR400 (5:4) 2225 memory
timings
2.923, 2.805, 3.114, 3.126 GB/s

(ran on Abit IC7, Windows XP sp1a)

From the STREAM site, it looks like the closest PC is the 533Mhz P4 [at]
2.8, which got:
1.832, 1.736, 2.249, 2.249 GB/s

=Heywood=

Steve Bellenot

06-30-2003, 01:48 PM

In article <bdnqgo$eo4$1[at]morgoth.sfu.ca>,
Erick Bryce Wong <erick[at]sfu.ca> wrote:
>Steve Bellenot <bellenot[at]math.fsu.edu> wrote:
>>
>>Symbolic computations can/often have immediate calculations which require
>>lots of memory. A trivial example would be computing a binomial
>>coefficient 200 choose 100 = 200!/(100!)^2 by multiplying the factorials
>>out and doing the long division. By the way, this is also an example
>>of an integer calculation in Mathematica. The stuff below is maple,
>>but the idea is the same:
>>> binomial(200,100);
>> 90548514656103281165404177077484163874504589675413336841320
>
>Point taken, but kind of a bad example :). Maple, like any sensible system,
>doesn't compute the factorials for integer binomial coefficients. Instead,
>it computes binomial(200,i) incrementally for i from 1 to 100, each time
>multiplying by a single rational (using iquo to avoid unnecessary gcd
>computations). So at no stage of binomial(200,100) does the intermediate
>result exceed the answer :).

I don't think you read carefully enough. I did NOT say that maple/mathematical
would compute binomial(200,100) by doing
1. computing 200!
2. computing 100! and square it
3. divide the answer in one by the answer in two.

What I did say was that if you did the steps above you would compute
binomial(200,100) which was much smaller than the immediate steps.

While maple does use the special form and knowledge of binomial coefficients
to compute binomial(200,100) by the loop
answer:=1
for i from 2 to 100 do
answer = answer * (200 - i + 1) / i;
enddo
It will not use this routine if you ask it to compute 200!/(100!)^2.

But I get your point too, but can you think of a better example? Basically
one needs something coming from to different calculations where there is
suddenly lots of cancellation. If there is lots of cancellation to see,
it is likely to be a standard case. Here is another one
> sum(f(i),i=1..n) - sum(f(i),i=2..n+1);
/ n \ /n + 1 \
|----- | |----- |
| \ | | \ |
| ) f(i)| - | ) f(i)|
| / | | / |
|----- | |----- |
\i = 1 / \i = 2 /
and a particular case.
> sum(1/i,i=1..n)-sum(1/i,i=2..n+1);
Psi(n + 1) - Psi(n + 2) + 1

clearly maple is not seeing that there is lots of cancellation available
although it is very obvious to many humans that `telescoping' is going on.

--
http://www.math.fsu.edu/~bellenot
bellenot <At/> math.fsu.edu
+1.850.644.7189 (4053fax)

David Wang

06-30-2003, 03:01 PM

In comp.sys.mac.advocacy Heywood Mogroot <imouttahere[at]mac.com> wrote:
> David Wang <foo[at]bar.invalid> wrote in message news:<bdo7bf$hld$1[at]grapevine.wam.umd.edu>...
>> I haven't seen any 875P results for McCalpin's stream. I am guesssing
>> that it would report somewhere between 3.2 and 3.6 GB/s. If you can
>> run the stock STREAM, and report the results to McCalpin, you can
>> be (in)famous.

> wstream.exe 2.6C P4 [at] 2.6GHz, 800Mhz FSB DDR400 (1:1) 2225 memory
> timings
> 2.748, 2.726, 2.950, 2.934 GB/s

> wstream.exe 2.6C P4 [at] 3.25Ghz, 1000Mhz FSB DDR400 (5:4) 2225 memory
> timings
> 2.923, 2.805, 3.114, 3.126 GB/s

Slightly lower than I expected, but data is data.

Send an email to McCalpin, perhaps he'll include it on his web
page.

--
davewang202(at)yahoo(dot)com

Steve Hanson

06-30-2003, 09:05 PM

digitaleon wrote in <Vq2dnSG9VNzzFmOjRTvU2Q[at]giganews.com>:

>To comp.os.linux, comp.sys.mac.advocacy and comp.sys.mac.system
>subscribers,
>
>> From: Jim Kroger <jimkkREMOVEME[at]umich.edu>,
>
>> It it hype or has apple blown away the x86 world?
>>
>> Any pointers most appreciated.
>
>Aside from the hotly-debated SPEC performance comparisons on Apple's
>site and the not-so-hotly-debated application performance comparisons
>at the same location, nowhere as yet.

Apple's strategy with benchmarks has always been to announce the big
lie well in advance of rollout to capitalize on the delay between the
big lie and independent testing. By then the impression left in the
typical user's mind will be "Mac is faster" no matter what the true
story is. Also, the more sand you can throw the better--people will
just remember the complicated argument, not the little pieces of
evidence like an incorrectly compiled graphics app.

Politicians do this all the time. In fact, an Apple board member, Al
Gore, did a variation on this fairly often.

Hmm, I'm beginning to see what he brought to the table. Maybe it
wasn't just the usual post-office payback after all.

Eric Salathe

06-30-2003, 09:07 PM

Jim Kroger:
> It it hype or has apple blown away the x86 world?
> Any pointers most appreciated.

For scientific applications, there are a number of interesting posts
to the Apple SciTech list. However, at this point, comparisons are for
code compiled on a G4 and run on a G5. So, one can only extrapolate to
performance with G5 compilers. Altivec is also a big variable in the
comparison, since Altivectorized code already "blew away the x86
world".

Real-world scalar performance is most certainly somewhere between the
gcc-to-gcc comparison Apple did and the gcc-to-icc comparison taking
P4 numbers from SPEC.org. Since no one uses icc/ifc for scientific and
technical computing on Intel/AMD, but rather gcc and Portland Group
compilers, I tend to lean toward the Apple comparison as more
relevant.

Previously, the G4 was a clear leader only for Altivec-enabled code.
It was close enough for scalar code that considerations other than
strict cpu performance might make it the preferred platform in many
cases.

Benchmarks posted on the SciTech list clearly show that the G5 will
continue to stand out for vectorizable code. Scalar performance is
certain to be on par with P4 once all is said and done. Comparison to
the Opteron is too early and the opteron is pricier, so the
performance per dollar equation needs to be considered.

As before, what is the fastest platform for your dollar will still be
a matter of testing your code on the platform under consideration.

Eric Salathe

Christian Bau

07-01-2003, 12:00 AM

In article
<jimkkREMOVEME-B165BC.20333629062003[at]visonmassif.rs.itd.umich.edu>,
Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote:

> But it doesn't affect at all whether I should buy a G5 or a dual
> Opteron. Or whatever they call it.

I think it is quite pointless to compare G5 and Opteron.

Lets say Steve Jobs goes on a stage and tells the world that 64 bit is
_really_ a must for any decent computer nowadays, and a well-designed 64
bit processor beats the shit out of a P4. What do you think the CEO of
AMD will say about that?

He will agree one hundred percent.

Heywood Mogroot

07-01-2003, 01:47 AM

Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message news:<jimkkREMOVEME-B165BC.20333629062003[at]visonmassif.rs.itd.umich.edu>...
> But it doesn't affect at all whether I should buy a G5 or a dual
> Opteron. Or whatever they call it.

The dual opteron is going to be VERY comparable in performance, at
similar pricepoints, to the dual G5.

Too bad AMD can't get any traction with the OEMs who are in bed with
Intel (and the m/b manufacturers dependent on Intel chipsets)...

Asus is supposed to be announcing an nForce3 Pro board this week
(SK8N).

http://www.xbitlabs.com/news/mainboards/display/20030519042810.html

A reasonably configured desktop (1.4Ghz opteron / 512MB / 120 GB /
9800Pro) is around $1900 from them.

=Heywood=

foo

07-01-2003, 01:53 AM

On 30 Jun 2003 17:47:21 -0700, imouttahere[at]mac.com (Heywood Mogroot)
wrote:

>Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message news:<jimkkREMOVEME-B165BC.20333629062003[at]visonmassif.rs.itd.umich.edu>...
>> But it doesn't affect at all whether I should buy a G5 or a dual
>> Opteron. Or whatever they call it.
>
>The dual opteron is going to be VERY comparable in performance, at
>similar pricepoints, to the dual G5.
>
>Too bad AMD can't get any traction with the OEMs who are in bed with
>Intel (and the m/b manufacturers dependent on Intel chipsets)...

Erm...huh? You've already mentioned Asus, and I'm sure ABIT and
others will announce MBs eventually when the market presents itself.
There's excitement on Madison, too - a lot is happening in x86-land.

>Asus is supposed to be announcing an nForce3 Pro board this week
>(SK8N).
>
>http://www.xbitlabs.com/news/mainboards/display/20030519042810.html
>
>A reasonably configured desktop (1.4Ghz opteron / 512MB / 120 GB /
>9800Pro) is around $1900 from them.
>
>=Heywood=

Heywood Mogroot

07-01-2003, 10:02 AM

foo <foo[at]bar.com> wrote in message news:<tqm1gvgkvh95he5qnli2opblsiv1v2m7ki[at]4ax.com>...
> On 30 Jun 2003 17:47:21 -0700, imouttahere[at]mac.com (Heywood Mogroot)
> wrote:
>
> >Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message news:<jimkkREMOVEME-B165BC.20333629062003[at]visonmassif.rs.itd.umich.edu>...
> >> But it doesn't affect at all whether I should buy a G5 or a dual
> >> Opteron. Or whatever they call it.
> >
> >The dual opteron is going to be VERY comparable in performance, at
> >similar pricepoints, to the dual G5.
> >
> >Too bad AMD can't get any traction with the OEMs who are in bed with
> >Intel (and the m/b manufacturers dependent on Intel chipsets)...
>
> Erm...huh? You've already mentioned Asus, and I'm sure ABIT and
> others will announce MBs eventually when the market presents itself.

For OEM's I was talking about Dell, Gateway etc.

Abit's SK8N with nVidia's nForce3 Pro chip was announced today.

http://www.amdzone.com/releaseview.cfm?ReleaseID=1112

> There's excitement on Madison, too

given the twin turds Intel has laid on the market with Itanium, I
suppose third time's the charm.

$4200 is an... interesting... pricepoint for the 1.5Ghz chip. That'd
nearly buy the CPU's for an 8-way 1.6Ghz Opteron array.

Worst comes to worst, with its 107W TDP the Itanium3 will make a good
waterbed heater...

> - a lot is happening in x86-land.

same circles, faster and faster.

=Heywood=

James Stutts

07-02-2003, 05:55 AM

"Eric Salathe" <esalathe[at]cascade.org> wrote in message
news:e7d80033.0306301207.3f2bbb63[at]posting.google.com...
> Jim Kroger:
> > It it hype or has apple blown away the x86 world?
> > Any pointers most appreciated.
>
> For scientific applications, there are a number of interesting posts
> to the Apple SciTech list. However, at this point, comparisons are for
> code compiled on a G4 and run on a G5. So, one can only extrapolate to
> performance with G5 compilers. Altivec is also a big variable in the
> comparison, since Altivectorized code already "blew away the x86
> world".
>
> Real-world scalar performance is most certainly somewhere between the
> gcc-to-gcc comparison Apple did and the gcc-to-icc comparison taking
> P4 numbers from SPEC.org. Since no one uses icc/ifc for scientific and
> technical computing on Intel/AMD, but rather gcc and Portland Group
> compilers, I tend to lean toward the Apple comparison as more
> relevant.

Except Apple didn't use pgf90 on x86, but Nagware.

JCS

James Stutts

07-02-2003, 05:55 AM

"Heywood Mogroot" <imouttahere[at]mac.com> wrote in message
news:dd5de929.0307010102.2c72b2e8[at]posting.google.com...
> foo <foo[at]bar.com> wrote in message
news:<tqm1gvgkvh95he5qnli2opblsiv1v2m7ki[at]4ax.com>...
> > On 30 Jun 2003 17:47:21 -0700, imouttahere[at]mac.com (Heywood Mogroot)
> > wrote:
> >
> > >Jim Kroger <jimkkREMOVEME[at]umich.edu> wrote in message
news:<jimkkREMOVEME-B165BC.20333629062003[at]visonmassif.rs.itd.umich.edu>...
> > >> But it doesn't affect at all whether I should buy a G5 or a dual
> > >> Opteron. Or whatever they call it.
> > >
> > >The dual opteron is going to be VERY comparable in performance, at
> > >similar pricepoints, to the dual G5.
> > >
> > >Too bad AMD can't get any traction with the OEMs who are in bed with
> > >Intel (and the m/b manufacturers dependent on Intel chipsets)...
> >
> > Erm...huh? You've already mentioned Asus, and I'm sure ABIT and
> > others will announce MBs eventually when the market presents itself.
>
> For OEM's I was talking about Dell, Gateway etc.
>
> Abit's SK8N with nVidia's nForce3 Pro chip was announced today.
>
> http://www.amdzone.com/releaseview.cfm?ReleaseID=1112
>
> > There's excitement on Madison, too
>
> given the twin turds Intel has laid on the market with Itanium, I
> suppose third time's the charm.
>
> $4200 is an... interesting... pricepoint for the 1.5Ghz chip. That'd
> nearly buy the CPU's for an 8-way 1.6Ghz Opteron array.
>
> Worst comes to worst, with its 107W TDP the Itanium3 will make a good
> waterbed heater...
>
> > - a lot is happening in x86-land.
>
> same circles, faster and faster.

Apple spin getting to you?

JCS

07-02-2003, 06:57 AM

"James Stutts" <stuttjc[at]knology.net> wrote in message
news:vg4pdkll1tlc78[at]corp.supernews.com...
>
> "Eric Salathe" <esalathe[at]cascade.org> wrote in message
> news:e7d80033.0306301207.3f2bbb63[at]posting.google.com...
> > Real-world scalar performance is most certainly somewhere between the
> > gcc-to-gcc comparison Apple did and the gcc-to-icc comparison taking
> > P4 numbers from SPEC.org. Since no one uses icc/ifc for scientific and
> > technical computing on Intel/AMD, but rather gcc and Portland Group
> > compilers, I tend to lean toward the Apple comparison as more
> > relevant.
>
> Except Apple didn't use pgf90 on x86, but Nagware.

absoft if also very popular (and available for windows, linux, and os x). i
think the compiler complaints, at least for the fortran side, would have
largely gone away if they had used something like absoft; it's well
respected, well known, and optimizes for the specific chips (athlon, p4,
opteron, g4- including altivec). of course they would have had to get a
port to g5, but i'm sure that had apple asked, they would've done it (since
they've already commited to a g5 port...)

James Stutts

07-02-2003, 07:11 AM

"ed" <news[at]no-atwistedweb-spam.com> wrote in message
news:3vuMa.68$ZQ2.15[at]newssvr16.news.prodigy.com...
> "James Stutts" <stuttjc[at]knology.net> wrote in message
> news:vg4pdkll1tlc78[at]corp.supernews.com...
> >
> > "Eric Salathe" <esalathe[at]cascade.org> wrote in message
> > news:e7d80033.0306301207.3f2bbb63[at]posting.google.com...
> > > Real-world scalar performance is most certainly somewhere between the
> > > gcc-to-gcc comparison Apple did and the gcc-to-icc comparison taking
> > > P4 numbers from SPEC.org. Since no one uses icc/ifc for scientific and
> > > technical computing on Intel/AMD, but rather gcc and Portland Group
> > > compilers, I tend to lean toward the Apple comparison as more
> > > relevant.
> >
> > Except Apple didn't use pgf90 on x86, but Nagware.
>
> absoft if also very popular (and available for windows, linux, and os x).
i

That would've been a better fit then two different versions of gcc, but even
then-
there's no such thing as an actual cross-platform compiler.

> think the compiler complaints, at least for the fortran side, would have
> largely gone away if they had used something like absoft; it's well
> respected, well known, and optimizes for the specific chips (athlon, p4,
> opteron, g4- including altivec). of course they would have had to get a
> port to g5, but i'm sure that had apple asked, they would've done it
(since
> they've already commited to a g5 port...)

Based on their previous benchmark shenanigans, it is quite easy to see what
they
were about.

JCS

Eric Salathe

07-02-2003, 09:49 PM

"ed" <news[at]no-atwistedweb-spam.com> wrote :
> as it stands, who the hell knows anything
> about nagware? who the heck uses the thing? i've never even heard of it,
> and we actually still use fortran quite a bit. =)

I actually use it as the primary compiler on my Mac. The Univ of
Washington has a site license, so I suspect there are quite a few
others using it -- linux and most Unix platforms are supported. NAG
f95 actually translates into very low-level c and then uses the local
c compiler, so for Apple's testing protocol, using NAG actually was
the right choice (whether Apple's protocol is reasonable is outside my
domain of giving a darn).

I have not been able to do a good comparison of fortran compilers on
the Mac, since I do not have the Absoft compiler. I sent some code to
somone with Absoft installed on a 550 MHz PowerBook and compared it to
the same code compiled with NAG runnng on my 800 MHz PowerBook.

Beware: This is pretty sketchy benchmarking! For entertainment
purposes only!

The results were (smaller is better):

System MHz compliler time time*MHz/100
------------------------------------------------------------
Powerbook 550 g77 -O3 120 660
Powerbook 800 g77 -O3 83 664

Powerbook 550 f90 -O (Absoft) 129 710
Powerbook 800 f95 -O (NAG) 84 672

Powerbook 800 f95 -O3 (NAG) 78 624

As you can see from the first two lines, g77 performance scales with
processor speed. So comparing time*MHz is reasonable for this code on
these two powerbooks.

1) NAG appears to outperform Absoft for this particular code with -O2
optimization (-O==-O2 for NAG).

2) Since NAG allows -O3 optimization, but Absoft aparently does not, I
got significantly better performance from NAG with full optimization.

3) Absoft is about comparable to g77, maybe not quite as fast even.

4) NAG significantly outperforms g77 even tho it uses gcc.

g77 is free, but pretty useless for modern code. NAG is fairly cheap
and fast, but includes no extras. Absoft appears to buy little or
nothing in terms of performance, but includes a lot of good tools,
some of which may end up yielding faster computations in the end.

For kicks and grins, same code on two additional platforms:

AthlonMP 1200 ifc -O3 67 804
Alpha 618 f90 -O3 (digital) 44 273

The AthlonMP, dual cpu, cost $1000 in parts off the web (not counting
the salary of the guy who built it and 36 others like it).

-Eric

James Stutts

07-03-2003, 05:33 AM

07-03-2003, 05:41 AM

"Eric Salathe" <esalathe[at]cascade.org> wrote in message
news:e7d80033.0307021249.58f049c6[at]posting.google.com...
> "ed" <news[at]no-atwistedweb-spam.com> wrote :
> > as it stands, who the hell knows anything
> > about nagware? who the heck uses the thing? i've never even heard of
it,
> > and we actually still use fortran quite a bit. =)
>
> I actually use it as the primary compiler on my Mac. The Univ of
> Washington has a site license, so I suspect there are quite a few
> others using it -- linux and most Unix platforms are supported. NAG
> f95 actually translates into very low-level c and then uses the local
> c compiler, so for Apple's testing protocol, using NAG actually was
> the right choice (whether Apple's protocol is reasonable is outside my
> domain of giving a darn).
>
> I have not been able to do a good comparison of fortran compilers on
> the Mac, since I do not have the Absoft compiler. I sent some code to
> somone with Absoft installed on a 550 MHz PowerBook and compared it to
> the same code compiled with NAG runnng on my 800 MHz PowerBook.
>
> Beware: This is pretty sketchy benchmarking! For entertainment
> purposes only!
>
> The results were (smaller is better):
>
> System MHz compliler time time*MHz/100
> ------------------------------------------------------------
> Powerbook 550 g77 -O3 120 660
> Powerbook 800 g77 -O3 83 664
>
> Powerbook 550 f90 -O (Absoft) 129 710
> Powerbook 800 f95 -O (NAG) 84 672
>
> Powerbook 800 f95 -O3 (NAG) 78 624
>
> As you can see from the first two lines, g77 performance scales with
> processor speed. So comparing time*MHz is reasonable for this code on
> these two powerbooks.
>
> 1) NAG appears to outperform Absoft for this particular code with -O2
> optimization (-O==-O2 for NAG).
>
> 2) Since NAG allows -O3 optimization, but Absoft aparently does not, I
> got significantly better performance from NAG with full optimization.
>
> 3) Absoft is about comparable to g77, maybe not quite as fast even.
>
> 4) NAG significantly outperforms g77 even tho it uses gcc.
>
> g77 is free, but pretty useless for modern code. NAG is fairly cheap
> and fast, but includes no extras. Absoft appears to buy little or
> nothing in terms of performance, but includes a lot of good tools,
> some of which may end up yielding faster computations in the end.
>
> For kicks and grins, same code on two additional platforms:
>
> AthlonMP 1200 ifc -O3 67 804
> Alpha 618 f90 -O3 (digital) 44 273
>
> The AthlonMP, dual cpu, cost $1000 in parts off the web (not counting
> the salary of the guy who built it and 36 others like it).

cool, good info. like you said, sketchy, but good info. =)

and for those who still think just because a compiler has the same name on
different platforms, look at the differences in a compiler's (absoft)
inability to reliably compile the same code across platforms, nevermind
assuming equal optimizations:
http://www.cs.rpi.edu/%7Eszymansk/OOF90/eval.html

- ILUVJazz -

07-03-2003, 03:05 PM

On Sat, 28 Jun 2003 23:20:31 -0400, Jim Kroger
<jimkkREMOVEME[at]umich.edu> wrote:

>It it hype or has apple blown away the x86 world?
>
>Any pointers most appreciated.
>
>Thanks
>Jim

http://www.haxial.com/spls-soapbox/apple-powermac-G5/

Daniel Packman

07-03-2003, 04:48 PM

Eric Salathe

07-03-2003, 08:01 PM

Me:
> > I actually use [NAG f95] as the primary compiler on my Mac. The Univ of
> > Washington has a site license, so I suspect there are quite a few
> > others using it

James Stutts
> I have never encountered Nagware in an academic
> or professional setting. I have encountered Absoft.

Well, you just have. I believe NAG f95 is superior to Absoft on the
Mac, and it's significanly less expensive.

I know only one other person using it here (Univ WA), also on a Mac. I
know of no one using Absoft. Since several departments here and at
dozens of other universities <http://www.nag.com/Local/asli.asp> pay
$500 a year for a license presumably one might encounter other NAG
users if you looked harder. These are on systems other than the Mac,
since the Mac version was included in the license only recently.

>Most [users on] UNIX platforms would use the system compiler

Certainly; for commercial Unix, the vendor's compiler should be
fastest. I think NAG has a presence there since it was first with f90
and offers very standard platform independence, due to its use of c
as an intermediary language.

-Eric

James Stutts

07-04-2003, 09:41 AM

"Eric Salathe" <esalathe[at]cascade.org> wrote in message
news:e7d80033.0307031101.321b5ee8[at]posting.google.com...
> Me:
> > > I actually use [NAG f95] as the primary compiler on my Mac. The Univ
of
> > > Washington has a site license, so I suspect there are quite a few
> > > others using it
>
> James Stutts
> > I have never encountered Nagware in an academic
> > or professional setting. I have encountered Absoft.
>
> Well, you just have. I believe NAG f95 is superior to Absoft on the

No, I have not. This is not an academic or commercial setting. :)

> Mac, and it's significanly less expensive.
>
> I know only one other person using it here (Univ WA), also on a Mac. I
> know of no one using Absoft. Since several departments here and at
> dozens of other universities <http://www.nag.com/Local/asli.asp> pay
> $500 a year for a license presumably one might encounter other NAG
> users if you looked harder. These are on systems other than the Mac,
> since the Mac version was included in the license only recently.

Well, we didn't use them in college. We use Portland, CVF, and Absoft at
work - mostly
CVF. I have never seen a NAG product outside of a math library in any
professional
setting. Now, that's anecdotal, but so is your's. NAG is also quite a bit
more expensive for us.

>
> >Most [users on] UNIX platforms would use the system compiler
>
> Certainly; for commercial Unix, the vendor's compiler should be
> fastest. I think NAG has a presence there since it was first with f90
> and offers very standard platform independence, due to its use of c
> as an intermediary language.

Which, as a Fortran user, bothers me. A glorified f2c? <shudder>

JCS

digitaleon

07-04-2003, 01:42 PM

To comp.os.linux, comp.sys.mac.advocacy and comp.sys.mac.system
subscribers,

>> From: Jim Kroger <jimkkREMOVEME[at]umich.edu>,
> From: - ILUVJazz - <ILVJazz[at]NYOB.com>,

>> Is it hype or has apple blown away the x86 world?
>>
>> Any pointers most appreciated.
>>
>> Thanks
>> Jim
>
> http://www.haxial.com/spls-soapbox/apple-powermac-G5/

The information on the above page is both misleading and inaccurate. It
has been debunked many times, in this forum, in other forums, and by
Apple and VeriTest.

The reasons for the benchmark discrepancies are already known: GCC was
used on both platforms, even though it does not produce the most
efficient code on either, and the PowerMac used a less-efficient but
better-performing malloc library.

Other claims on the page, that HyperThreading being disabled decreased
test performance, that compilations were done on the Dell workstation
with optimisations disabled, and that Apple is engaging in misleading
pricing, have already been exposed as blatently dishonest.

Criticisms across the internet have referenced the above article as
evidence in an unprecendented fashion. Some of the articles linked to
from the above page link back _to_ it. A minority of sites that have
published information from the above source have actually displayed
impartiality, whereas a majority of them have wilfully ignored
information from multiple other sources contradicting what is on the
above page, including the vendor.

And of course, this is rounded off by the brazen audacity in claiming
that spl of haxial.net is the only voice for truth out of a chorus of
misguided, zealous believers. It occured to these news sources to
validate the claims of an independant, established benchmarking company
on the basis that they were paid by Apple (it's not like they were
going to do it for free, after all), but the same scrutiny was not paid
to the counter-claims of an anonymous individual on the internet.

Draw your own conclusions.

digitaleon.

Heywood Mogroot

07-04-2003, 07:46 PM

pack[at]eos.ucar.edu (Daniel Packman) wrote in message news:<be1j8c$mdl$1[at]news.ucar.edu>...
> While the scalar code (integer and floating point) speed
> of the G5 is clearly similar to that of p4 machines,
> the big performance gain is still with code that can
> be vectorized.

And broken into multiple worksets for parallel processing.

P4 has its "hyperthreading". Dual G5's, with the supporting memory
controller, are the real thing.

Complicating these comparisons is AMD's Opteron platform, which
appears to offer competitive bang-for-the-buck in multiprocessing
applications.

One thing's for sure, the coming year is going to be a good year for
the high-end PC/workstation user... 3-way competition!

=Heywood=

Michael Doster

07-04-2003, 08:23 PM

in article 33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com, Josiah Fizer at
jfizer[at]classy.com wrote on 7/4/03 2:50 PM:

> On 4 Jul 2003 11:46:50 -0700, imouttahere[at]mac.com (Heywood Mogroot)
> wrote:
>
>> pack[at]eos.ucar.edu (Daniel Packman) wrote in message
>> news:<be1j8c$mdl$1[at]news.ucar.edu>...
>>> While the scalar code (integer and floating point) speed
>>> of the G5 is clearly similar to that of p4 machines,
>>> the big performance gain is still with code that can
>>> be vectorized.
>>
>> And broken into multiple worksets for parallel processing.
>>
>> P4 has its "hyperthreading". Dual G5's, with the supporting memory
>> controller, are the real thing.
>>
>> Complicating these comparisons is AMD's Opteron platform, which
>> appears to offer competitive bang-for-the-buck in multiprocessing
>> applications.
>>
>> One thing's for sure, the coming year is going to be a good year for
>> the high-end PC/workstation user... 3-way competition!
>>
>> =Heywood=
>
> And I'll keep working away on my dual, quad and eight way CPU Sun
> systems and wondering what all the fuss is about.
>
>
> ----== Posted via Newsfeed.Com - Unlimited-Uncensored-Secure Usenet News==----
> http://www.newsfeed.com The #1 Newsgroup Service in the World! >100,000
> Newsgroups
> ---= 19 East/West-Coast Specialized Servers - Total Privacy via Encryption
> =---

1. You know you're sharing your Sun dualie and the eight-headed beast with
everyone and their mother who happens to be on the same LAN and have access
to all your lovely CPUs?

And

2. Exactly how fast are the CPUs on yours, the two CPU Suns I deal with
daily use 360 MHz Ultra SPARC II and our two 8-headed beasts have 400 MHz
Ultra SPARC II?

Michael

Josiah Fizer

07-04-2003, 08:32 PM

On Fri, 04 Jul 2003 19:23:52 GMT, Michael Doster
<michael.doster[at]verizon.net> wrote:

>in article 33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com, Josiah Fizer at
>jfizer[at]classy.com wrote on 7/4/03 2:50 PM:
>
>> On 4 Jul 2003 11:46:50 -0700, imouttahere[at]mac.com (Heywood Mogroot)
>> wrote:
>>
>>> pack[at]eos.ucar.edu (Daniel Packman) wrote in message
>>> news:<be1j8c$mdl$1[at]news.ucar.edu>...
>>>> While the scalar code (integer and floating point) speed
>>>> of the G5 is clearly similar to that of p4 machines,
>>>> the big performance gain is still with code that can
>>>> be vectorized.
>>>
>>> And broken into multiple worksets for parallel processing.
>>>
>>> P4 has its "hyperthreading". Dual G5's, with the supporting memory
>>> controller, are the real thing.
>>>
>>> Complicating these comparisons is AMD's Opteron platform, which
>>> appears to offer competitive bang-for-the-buck in multiprocessing
>>> applications.
>>>
>>> One thing's for sure, the coming year is going to be a good year for
>>> the high-end PC/workstation user... 3-way competition!
>>>
>>> =Heywood=
>>
>> And I'll keep working away on my dual, quad and eight way CPU Sun
>> systems and wondering what all the fuss is about.
>>
>>
>> ----== Posted via Newsfeed.Com - Unlimited-Uncensored-Secure Usenet News==----
>> http://www.newsfeed.com The #1 Newsgroup Service in the World! >100,000
>> Newsgroups
>> ---= 19 East/West-Coast Specialized Servers - Total Privacy via Encryption
>> =---
>
>1. You know you're sharing your Sun dualie and the eight-headed beast with
>everyone and their mother who happens to be on the same LAN and have access
>to all your lovely CPUs?
>

That would be myself and I for the single and dual and myself and the
end users for the 4 way systems. I'm still setting up the 8 CPU system
I administrate for the dev team to test on.

>And
>
>2. Exactly how fast are the CPUs on yours, the two CPU Suns I deal with
>daily use 360 MHz Ultra SPARC II and our two 8-headed beasts have 400 MHz
>Ultra SPARC II?
>
>Michael

The quads are e450s with four 450mhz UltraSparc II CPUs with 4 megs of
cache per CPU. The eight way system is a Sun Fire 3800 with eight
900mhz (?) UltraSparc III CPUs and 16GB of RAM, not sure of the cache.

My home systems are a dual 400mhz Ultra60 and a single CPU CompactPCI
server running at 333mhz. Despite the low speed, these systems take a
much higher load then the PC and Macintosh boxes I use. On a single
task yes the PC or Mac is faster, however under heavy work loads the
Suns just keep on chugging away while the Intel and Apple boxes grind
to a halt. Different tools for different tasks, overall I like the
Suns better.

----== Posted via Newsfeed.Com - Unlimited-Uncensored-Secure Usenet News==----
http://www.newsfeed.com The #1 Newsgroup Service in the World! >100,000 Newsgroups
---= 19 East/West-Coast Specialized Servers - Total Privacy via Encryption =---

Kevin Stevens

07-04-2003, 08:58 PM

In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
Josiah Fizer <jfizer[at]classy.com> wrote:

> And I'll keep working away on my dual, quad and eight way CPU Sun
> systems and wondering what all the fuss is about.

Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
talk about the "first 64-bit personal computer" when the Blade 100/150
has been out for some two years.

KeS

(Will enjoy seeing an Opteron/Sun/G5 comparison in September, should be
interesting.)

Josiah Fizer

07-04-2003, 09:08 PM

On Fri, 04 Jul 2003 12:58:13 -0700, Kevin Stevens
<Kevin_Stevens[at]Hotmail.com> wrote:

>In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
> Josiah Fizer <jfizer[at]classy.com> wrote:
>
>> And I'll keep working away on my dual, quad and eight way CPU Sun
>> systems and wondering what all the fuss is about.
>
>Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
>talk about the "first 64-bit personal computer" when the Blade 100/150
>has been out for some two years.
>
>KeS
>
>(Will enjoy seeing an Opteron/Sun/G5 comparison in September, should be
>interesting.)

Blade 1000 Dual = ~8000$
Blade 150 = ~1200$

Suns have been 64bit since the Ultrasparc series.

----== Posted via Newsfeed.Com - Unlimited-Uncensored-Secure Usenet News==----
http://www.newsfeed.com The #1 Newsgroup Service in the World! >100,000 Newsgroups
---= 19 East/West-Coast Specialized Servers - Total Privacy via Encryption =---

Tom

07-04-2003, 09:56 PM

> And I'll keep working away on my dual, quad and eight way CPU Sun
> systems and wondering what all the fuss is about.

If I could just run Painter, DeepPaint and PhotoShop on there ...

07-04-2003, 10:15 PM

In article <ts2ht-va32.ln1[at]pursued-with.net>,
Kevin Stevens <Kevin_Stevens[at]Hotmail.com> wrote:

> In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
> Josiah Fizer <jfizer[at]classy.com> wrote:
>
> > And I'll keep working away on my dual, quad and eight way CPU Sun
> > systems and wondering what all the fuss is about.
>
> Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
> talk about the "first 64-bit personal computer" when the Blade 100/150
> has been out for some two years.
>
> KeS

I'm a personal computer user and have never heard of the Blade 100/150.
Could you tell more about it please? What's the price?

Josiah Fizer

07-04-2003, 10:19 PM

On Fri, 04 Jul 2003 21:15:03 GMT, <lower[at]slobovvia.com> wrote:

>In article <ts2ht-va32.ln1[at]pursued-with.net>,
> Kevin Stevens <Kevin_Stevens[at]Hotmail.com> wrote:
>
>> In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
>> Josiah Fizer <jfizer[at]classy.com> wrote:
>>
>> > And I'll keep working away on my dual, quad and eight way CPU Sun
>> > systems and wondering what all the fuss is about.
>>
>> Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
>> talk about the "first 64-bit personal computer" when the Blade 100/150
>> has been out for some two years.
>>
>> KeS
>
>I'm a personal computer user and have never heard of the Blade 100/150.
>Could you tell more about it please? What's the price?

The Blade 100 when it was introduced (not sure you can still get it
from Sun) cost 900$ and was a 500mhz UltraSparc IIi based system. Very
nice little case etc. Go check out Suns web page for the current
systems such as the Blade 150.

----== Posted via Newsfeed.Com - Unlimited-Uncensored-Secure Usenet News==----
http://www.newsfeed.com The #1 Newsgroup Service in the World! >100,000 Newsgroups
---= 19 East/West-Coast Specialized Servers - Total Privacy via Encryption =---

flip

07-05-2003, 01:12 PM

In article <mjnbgv09p9d6vuglgvnod1c8au9h4mr39f[at]4ax.com>,
Josiah Fizer <jfizer[at]classy.com> wrote:

> On Fri, 04 Jul 2003 12:58:13 -0700, Kevin Stevens
> <Kevin_Stevens[at]Hotmail.com> wrote:
>
> >In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
> > Josiah Fizer <jfizer[at]classy.com> wrote:
> >
> >> And I'll keep working away on my dual, quad and eight way CPU Sun
> >> systems and wondering what all the fuss is about.
> >
> >Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
> >talk about the "first 64-bit personal computer" when the Blade 100/150
> >has been out for some two years.
> >
> >KeS
> >
> >(Will enjoy seeing an Opteron/Sun/G5 comparison in September, should be
> >interesting.)
>
> Blade 1000 Dual = ~8000$
> Blade 150 = ~1200$
>
> Suns have been 64bit since the Ultrasparc series.

64 bit workstations, not PCs.

http://www.sun.com/desktop/sunblade150/

07-05-2003, 01:59 PM

In article <flippo-63349A.07070705072003[at]news.central.cox.net>,
flip <flippo[at]mac.com> wrote:

> In article <ts2ht-va32.ln1[at]pursued-with.net>,
> Kevin Stevens <Kevin_Stevens[at]Hotmail.com> wrote:
>
> > In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
> > Josiah Fizer <jfizer[at]classy.com> wrote:
> >
> > > And I'll keep working away on my dual, quad and eight way CPU Sun
> > > systems and wondering what all the fuss is about.
> >
> > Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
> > talk about the "first 64-bit personal computer" when the Blade 100/150
> > has been out for some two years.
>
> Which happens to be a workstation, not a PC.
>
> That is, unless you're going to join Stutts and Foo in pretending that
> you know more about Sun's computers than Sun does.

You mean to tell me that we are not compairing a personal computer?
Then what is he talking about comparing a workstation to a Apple 64 bit
computer? It sounds like some more NG BS.

flip

07-05-2003, 02:24 PM

In article <lower-BD83F6.08010105072003[at]zeus-ge0.rdc-kc.rr.com>,
<lower[at]slobovvia.com> wrote:

> In article <flippo-63349A.07070705072003[at]news.central.cox.net>,
> flip <flippo[at]mac.com> wrote:
>
> > In article <ts2ht-va32.ln1[at]pursued-with.net>,
> > Kevin Stevens <Kevin_Stevens[at]Hotmail.com> wrote:
> >
> > > In article <33jbgv8451jjl9umn0c47peprnmlv10sjq[at]4ax.com>,
> > > Josiah Fizer <jfizer[at]classy.com> wrote:
> > >
> > > > And I'll keep working away on my dual, quad and eight way CPU Sun
> > > > systems and wondering what all the fuss is about.
> > >
> > > Go price a Blade 2000 and you'll see. Though it's funny to hear Apple
> > > talk about the "first 64-bit personal computer" when the Blade 100/150
> > > has been out for some two years.
> >
> > Which happens to be a workstation, not a PC.
> >
> > That is, unless you're going to join Stutts and Foo in pretending that
> > you know more about Sun's computers than Sun does.
>
>
> You mean to tell me that we are not compairing a personal computer?
> Then what is he talking about comparing a workstation to a Apple 64 bit
> computer? It sounds like some more NG BS.

Sun calls their Sun Blade a workstation, not a PC.

flip

07-05-2003, 03:17 PM

In article <slrnbgdms0.qjh.matt[at]audrey.boggle.org>,
Matt McLeod <matt[at]boggle.org> wrote:

> In <flippo-FD8FA3.07082405072003[at]news.central.cox.net>, flip wrote:
> > In article <mjnbgv09p9d6vuglgvnod1c8au9h4mr39f[at]4ax.com>,
> > Josiah Fizer <jfizer[at]classy.com> wrote:
> >> Blade 1000 Dual = ~8000$
> >> Blade 150 = ~1200$
> >>
> >> Suns have been 64bit since the Ultrasparc series.
> >
> > 64 bit workstations, not PCs.
> >
> > http://www.sun.com/desktop/sunblade150/
>
> It's not a very useful differentiation, particularly so at the lower
> end (i.e., the Blade 150) where it is effectively a PC with a CPU
> that is neither x86 nor PPC compatible.
>
> "Workstation" these days is a term for the marketing people.

Which is exactly what's relevant to Apple's marketing claim.

David Magda

07-05-2003, 03:35 PM

Steve Lidie <lusol[at]cube0.CC.Lehigh.EDU> writes:

> Now, now, Sun has never been a player in the multi-CPU, high
> performance computing market - their CPUs are too slow and don't
> scale. They can't touch, for instance, an SGI system with hundreds
> of CPUs, gobs of memory bandwidth, etcetera ...

SGI top-of-the-line Origin 3900 can have 128 CPUs per rack; the Sun
Fire 15K can have 106 CPUs per rack. Wow, big difference. The 3900 can
have 256GB of RAM per rack, the 15K can have 576GB.

Though, to be fair, SGI has "NUMAflex" (or whatever) where these
numbers supposedly can be increased for each "image": the website
says 512 CPUs, and 1TB of RAM.

Saying that Sun's CPUs don't scale is a bit much; I don't think there
are theoretical issues with more CPUs on a Sun, just that no one is
really asking for more (at least, not asking Sun).

(I'm assuming that the Origin 3900 runs IRIX.)

--
David Magda <dmagda at ee.ryerson.ca>, http://www.magda.ca/
Because the innovator has for enemies all those who have done well under
the old conditions, and lukewarm defenders in those who may do well
under the new. -- Niccolo Machiavelli, _The Prince_, Chapter VI

Jim Kroger

07-07-2003, 06:44 AM

imouttahere[at]mac.com (Heywood Mogroot) wrote in message news:<dd5de929.0307041046.488feee[at]posting.google.com>...
> pack[at]eos.ucar.edu (Daniel Packman) wrote in message news:<be1j8c$mdl$1[at]news.ucar.edu>...
> > While the scalar code (integer and floating point) speed
> > of the G5 is clearly similar to that of p4 machines,
> > the big performance gain is still with code that can
> > be vectorized.
>
> And broken into multiple worksets for parallel processing.
>
> P4 has its "hyperthreading". Dual G5's, with the supporting memory
> controller, are the real thing.
>
> Complicating these comparisons is AMD's Opteron platform, which
> appears to offer competitive bang-for-the-buck in multiprocessing
> applications.
>
> One thing's for sure, the coming year is going to be a good year for
> the high-end PC/workstation user... 3-way competition!
>
> =Heywood=

I don't know what vectorization is. My interest is in being able to
run (multiple copies of) matlab as fast as possible. I know it does
not use multiple processors, but I can run multiple matlabs, which
helps.

Does Matlab take advantage of vectorization?

Thanks
Jim

Jim Kroger

07-07-2003, 06:51 AM

Steve Lidie <lusol[at]cube0.CC.Lehigh.EDU> wrote in message news:<be5bc8$hm2[at]fidoii.CC.Lehigh.EDU>...
> In comp.sys.mac.system Josiah Fizer <jfizer[at]classy.com> wrote:
> > On 4 Jul 2003 11:46:50 -0700, imouttahere[at]mac.com (Heywood Mogroot)
> > wrote:
> >
> >>pack[at]eos.ucar.edu (Daniel Packman) wrote in message news:<be1j8c$mdl$1[at]news.ucar.edu>...
> >>> While the scalar code (integer and floating point) speed
> >>> of the G5 is clearly similar to that of p4 machines,
> >>> the big performance gain is still with code that can
> >>> be vectorized.
> >>
> >>And broken into multiple worksets for parallel processing.
> >>
> >>P4 has its "hyperthreading". Dual G5's, with the supporting memory
> >>controller, are the real thing.
> >>
> >>Complicating these comparisons is AMD's Opteron platform, which
> >>appears to offer competitive bang-for-the-buck in multiprocessing
> >>applications.
> >>
> >>One thing's for sure, the coming year is going to be a good year for
> >>the high-end PC/workstation user... 3-way competition!
> >>
> >>=Heywood=
> >
> > And I'll keep working away on my dual, quad and eight way CPU Sun
> > systems and wondering what all the fuss is about.
>
> Now, now, Sun has never been a player in the multi-CPU, high
> performance computing market - their CPUs are too slow and don't
> scale. They can't touch, for instance, an SGI system with hundreds of
> CPUs, gobs of memory bandwidth, etcetera ...
>
> Steve

I used a 24 cpu SGI at Princeton. Nothing we ever did could keep more
than one cpu busy, which would get the work done slower than a G3.
Course the prof still got to brag about his super computer that cost
in the upper six figures.
Jim

Buteo Lagopus

07-07-2003, 01:36 PM

imouttahere[at]mac.com (Heywood Mogroot) wrote in
news:dd5de929.0307041046.488feee[at]posting.google.com:

> One thing's for sure, the coming year is going to be a good year for
> the high-end PC/workstation user... 3-way competition!
>

"If your computer had eyes, you'd look like a statue."
--Unknown

Jim Glidewell

07-07-2003, 04:10 PM

In article <12e8gv45er1ub8d3mtfvgstdg40j8onj99[at]4ax.com>, - ILUVJazz - <ILVJazz[at]NYOB.com> wrote:

> On Sat, 28 Jun 2003 23:20:31 -0400, Jim Kroger
> <jimkkREMOVEME[at]umich.edu> wrote:
>
> >It it hype or has apple blown away the x86 world?
> >
> >Any pointers most appreciated.
>
> http://www.haxial.com/spls-soapbox/apple-powermac-G5/

Many of the points this guy made have since been discredited, IMO.

He views $1999 as a "misleading" price. (good grief!)

His attacks on Apple's and (particularly) Veritest's integrity are
unjustified and unreasonable.

He seems to _completely_ accept the (IMHO dubious) concept that:

Highest SPEC score === Fastest computer (always, period)

(Treating SPEC numbers as _the_ single defining metric, rather than
one measure among many)

One question this gentleman never stops to ask is:

Do *other* applications on the Intel platform see the radical
improvement in performance that SPEC sees when compiled using
the Intel compiler instead of gcc?

My understanding is that they do *not*.

Intel appears to have added specific optimizations in their compiler to
maximize its score on the SPEC benchmark. This is common practice these
days, and is done to varying degrees by all vendors offering a compiler
proprietary to their hardware. Tweaking your compiler to get better
benchmark scores is a good move marketing-wise, though it makes the
value of the benchmark itself as a useful comparison tool a bit more
questionable. And at this point in time, SPEC is the one to tweak for.

Using gcc instead of the Intel compiler was a reasonable decision, IMHO.

I neither know nor care enough about compiler options and BIOS settings
to make an informed judgement on those issues, but given his general
ranting and hair-splitting mode, I certainly won't take _his_ word
for 'em.

Finally, it is quite clear that for just about any task you throw at it,
a dual G5 will be at worst a few percent slower, and for a few things
2-3 times faster (see the BLAST results) than the fastest dual Xeon
systems today.

Application performance is what matters (to anyone who uses a computer
as a tool, rather than an ego-stroking device), not SPEC results. I
don't know how anyone who watched the keynote demos could not come
away with a least a kernel of doubt about the usefulness of SPEC in the
face of _large_ performance gaps for apps like Mathematica and
Photoshop. For customers looking to buy a machine for running _those_
apps, it is pretty clear that SPEC is an almost useless predictor
of application speed when comparing different platforms. And that the
G5's are indeed _fast_.

Either a dual Xeon or a dual G5 is *way* faster than anything that you
could get in a PC-priced system two years ago. Given the replacement
lifecycles for most desktop users, it is a very small percentage of
people who have the "fastest" PC on their desks at any given time.
And a claim of "fastest" only has a "shelf life" of 3 months or so
anyway, in a fast-paced market like this.

The obsessing over Apple's claim to have the fastest PC is simply
juvenile - no one machine is "fastest" for all applications, and
Apple provided a lot more solid justification and documentation
for their _marketing_ claim than others have done.

This guy's 15 minutes of fame are just about up.

And I fully expect that when people start running real applications
on these systems, that the relative performance characteristics will
be closer to what the application demos showed than whatever SPEC
might claim.

I can't wait.

--
Jim Glidewell
My opinions only

Arturo Pérez

07-07-2003, 05:19 PM

In article <6dcb1c5e.0307062144.7e17b67[at]posting.google.com>,
kroger[at]princeton.edu (Jim Kroger) wrote:

> I don't know what vectorization is. My interest is in being able to
> run (multiple copies of) matlab as fast as possible. I know it does
> not use multiple processors, but I can run multiple matlabs, which
> helps.
>
> Does Matlab take advantage of vectorization?
>
> Thanks
> Jim

How does matlab compare to Mathematica? If they are similar enough
then the Mathematica demo might be relevant to your expectations
for a G5 running matlab. In that case, I believe the keynote
demo showed Mathematica running 2x the speed of a 3GHz Pentium/Xeon.

Eric Salathe

07-08-2003, 07:04 PM

kroger[at]princeton.edu (Jim Kroger) wrote:
> Does Matlab take advantage of vectorization?

This is a good question and should probably be addressed to Mathworks;
I have not seen a definitive statement anywhere. It is likely that
they use LAPACK to perform matrix operations, in which case it is
likely that they link to veclib, Apple's Altivec implementation of
lapack. In that case, many matrix operations would be vectorized and
would run exceedingly fast.

However, it is unlikely that this is a major consideration. While
operations that are efficiently vectorizable will run faster on the
G4/G5 than anything out there by a considerable margin, such
performance gains are far from universal -- especially in the
mish-mash of Matlab code. If you really want speed, you need to
program in fortran or c.

Furthermore, given that the G5's scalar performance (ie without
Altivec) is equivalent to other similarly priced desktop systems,
processor performance is not particularly interesting or relevant to
selecting a computer.

-Eric

Daniel Packman

07-09-2003, 05:16 PM

In article <6dcb1c5e.0307062151.55ac9b3e[at]posting.google.com>,
Jim Kroger <kroger[at]princeton.edu> wrote:
.....
>I used a 24 cpu SGI at Princeton. Nothing we ever did could keep more
>than one cpu busy, which would get the work done slower than a G3.
>Course the prof still got to brag about his super computer that cost
>in the upper six figures.

This is contrary to our experience. We are regularly able to
get over 95% utilization of all 32 processors in our O3900
machine. We have found IRIX to be the easiest platform to achieve
parallelism in our code. The autoparalleling system works well.

--
Daniel Packman
NCAR/ACD
pack[at]ucar.edu

Steve Lidie

07-09-2003, 06:33 PM

In comp.sys.mac.system Daniel Packman <pack[at]eos.ucar.edu> wrote:
> In article <6dcb1c5e.0307062151.55ac9b3e[at]posting.google.com>,
> Jim Kroger <kroger[at]princeton.edu> wrote:
> ....
>>I used a 24 cpu SGI at Princeton. Nothing we ever did could keep more
>>than one cpu busy, which would get the work done slower than a G3.
>>Course the prof still got to brag about his super computer that cost
>>in the upper six figures.
>
> This is contrary to our experience. We are regularly able to
> get over 95% utilization of all 32 processors in our O3900
> machine. We have found IRIX to be the easiest platform to achieve
> parallelism in our code. The autoparalleling system works well.

Agreed. Our Origin 3800 w/32 CPUs is 100% busy around the clock, we can't
find enough cycles for everyone ...

Jim Glidewell

07-10-2003, 12:27 AM

In article <6dcb1c5e.0307062151.55ac9b3e[at]posting.google.com>, kroger[at]princeton.edu (Jim Kroger) wrote:

> Steve Lidie <lusol[at]cube0.CC.Lehigh.EDU> wrote in message news:<be5bc8$hm2[at]fidoii.CC.Lehigh.EDU>...

> > Now, now, Sun has never been a player in the multi-CPU, high
> > performance computing market - their CPUs are too slow and don't
> > scale. They can't touch, for instance, an SGI system with hundreds of
> > CPUs, gobs of memory bandwidth, etcetera ...

>
> I used a 24 cpu SGI at Princeton. Nothing we ever did could keep more
> than one cpu busy, which would get the work done slower than a G3.

Simply because you personally never used an application that scales
well on a SGI Origin is by no means evidence that such beasts do
not exist.

I help administer a 384 CPU Origin 3800. The bulk of the system is being
used for jobs which range from 64 - 128 CPUs, each averaging over 95%
CPU utilization. According to the users of this particular CFD application,
it scales very nearly linearly up well past 128 CPUs.

While their CPU speeds lag slightly behind the latest and greatest
Intel and/or IBM chips, the ability to run a single copy of the OS,
on a system with 384 CPUs and a single shared 384GB memory space,
provides the ability to get very good performance on a wide variety
of parallel codes.

Which, after many years of promises, have finally begun to arrive.

I personally would love to see a G5-based system using SGI's ccNUMA
shared memory architecture...

Cos' I'm still a vector kind of guy... :-)

> Course the prof still got to brag about his super computer that cost
> in the upper six figures.

Commercial sites don't buy supercomputers for bragging rights.

--
Jim Glidewell
My opinions only

Charles Douglas Wehner

07-26-2003, 06:33 PM

seebs[at]plethora.net (Seebs) wrote in message news:<3eff1a53$0$1097$3c090ad1[at]news.plethora.net>...
> In article <flippo-E843C3.06313729062003[at]news.central.cox.net>,
> flip <flippo[at]mac.com> wrote:
> >That is, unless you're willing to believe the Presidents of Adobe,
> >Wolfram, Luxology, and several other apps who presided over side-by-side
> >tests where the G5 kicked a P4's butt halfway to the moon.
>
> Wolfram I would believe. Adobe? You must be kidding me; believing something
> because Adobe said it would be just plain stupid.
>
> -s

Steve Jobs owns 90% of Adobe.

An independant "third party" would be more believable.

However, Apple design both the hardware and software - and so are in a
unique position to OPTIMISE their systems.

Charles Douglas Wehner

Steve Hix

07-26-2003, 06:46 PM

In article <e349085c.0307260933.168d6354[at]posting.google.com>,
charleswehner[at]hotmail.com (Charles Douglas Wehner) wrote:

> seebs[at]plethora.net (Seebs) wrote in message
> news:<3eff1a53$0$1097$3c090ad1[at]news.plethora.net>...
> > In article <flippo-E843C3.06313729062003[at]news.central.cox.net>,
> > flip <flippo[at]mac.com> wrote:
> > >That is, unless you're willing to believe the Presidents of Adobe,
> > >Wolfram, Luxology, and several other apps who presided over side-by-side
> > >tests where the G5 kicked a P4's butt halfway to the moon.
> >
> > Wolfram I would believe. Adobe? You must be kidding me; believing
> > something because Adobe said it would be just plain stupid.
>
> Steve Jobs owns 90% of Adobe.

This will come as a suprise to Adobe.

jemmy ducks

07-27-2003, 03:40 AM

In article <e349085c.0307260933.168d6354[at]posting.google.com>, Charles
Douglas Wehner <charleswehner[at]hotmail.com> wrote:

> Steve Jobs owns 90% of Adobe.

He is also the leader of an international Jewish conspiracy that aims
to enslave Christiandom. If Job was a Jew, "Jobs" is surely a Jew as
well. Save your women, your children, your self: go far away and hide.

Vladimir Grubor

07-30-2003, 12:51 PM

Someone (third party) at NASA did comparisons. G5 is actually PowerPC970,
which is IBMs cpu which is in turn a poor man's version of Power4 (another
server high end processor by IBM).

G5 is a 64 bit processor, which runs in 32 bit emulation mode (less than
ideal), since apple hasn't compiled MACOSX to support 64bit mode. Less than
ideal situation, so for now, probably Xeon is slightly ahead (as shown on
NASA test), but when apple sets up 64 bit support, I expect G5 to pull away
in performance stakes. Also I did read the log file for independent test
that Apple paid for. Math libraries were which are optimised for SPEC2000
were used to replace normal math libraries. That was not done on a Linux.

Cheers,

Vlad

"Charles Douglas Wehner" <charleswehner[at]hotmail.com> wrote in message
news:e349085c.0307260933.168d6354[at]posting.google.com...
> seebs[at]plethora.net (Seebs) wrote in message
news:<3eff1a53$0$1097$3c090ad1[at]news.plethora.net>...
> > In article <flippo-E843C3.06313729062003[at]news.central.cox.net>,
> > flip <flippo[at]mac.com> wrote:
> > >That is, unless you're willing to believe the Presidents of Adobe,
> > >Wolfram, Luxology, and several other apps who presided over
side-by-side
> > >tests where the G5 kicked a P4's butt halfway to the moon.
> >
> > Wolfram I would believe. Adobe? You must be kidding me; believing
something
> > because Adobe said it would be just plain stupid.
> >
> > -s
>
> Steve Jobs owns 90% of Adobe.
>
> An independant "third party" would be more believable.
>
> However, Apple design both the hardware and software - and so are in a
> unique position to OPTIMISE their systems.
>
> Charles Douglas Wehner

Vladimir Grubor

08-01-2003, 11:00 AM

You are right, I stand corrected. I was under the impression that since it
is a derivative of Power4, which is true 64bit processor, it must run 32 bit
apps in emulation mode, like Itanium?

Cheers,

Vlad

P.S. It might be superior, however, for intense computation (i.e. clusters,
it is still cheaper to get a bunch of P4s and cluster them than a bunch of
PowerPC970s. That is why I prefer PowerPC970 for desktop, but for intense
calculations, cheaper to get P4.
Do you agree with this?

"Arturo Pérez" <arturo[at]bigchalk.com> wrote in message
news:arturo-CC9E45.13460931072003[at]news.comcast.giganews.com...
> In article <3f27b13d$1[at]news.unimelb.edu.au>,
> "Vladimir Grubor" <vladimirgrubor<nospam>[at]hotmail.com> wrote:
>
> > Someone (third party) at NASA did comparisons. G5 is actually
PowerPC970,
> > which is IBMs cpu which is in turn a poor man's version of Power4
(another
> > server high end processor by IBM).
> >
> > G5 is a 64 bit processor, which runs in 32 bit emulation mode (less than
> > ideal), since apple hasn't compiled MACOSX to support 64bit mode. Less
than
> > ideal situation, so for now, probably Xeon is slightly ahead (as shown
on
> > NASA test), but when apple sets up 64 bit support, I expect G5 to pull
away
> > in performance stakes. Also I did read the log file for independent test
> > that Apple paid for. Math libraries were which are optimised for
SPEC2000
> > were used to replace normal math libraries. That was not done on a
Linux.
> >
> > Cheers,
> >
> > Vlad
> >
> >
>
> Some of the above is incorrect.
>
> 1) There is no 32bit emulation on PowerPC. The processor specification
> is both 32 AND 64 bit. IOW, the machine runs at full speed in either
> 32b or 64b mode. You do need a recompile if your application needs
> more than 4GB RAM to work.
>
> 2) The NASA person didn't test Xeons, only P4s. There's a thread in
> the Ars Technica battlefront where he discusses his results.The test he
> ran used machines he had handy. According to that test the G5 running
> unoptimized code (i.e. code compiled for the G4) was as fast as a
> 2.66GHz P4. On the other hand, because the bus speed for the current
> Xeons is only 533MHz anything needing bus bandwidth is going to run
> much (much!) faster on a G5.
>
> 3) The rumour of specialized math libraries is bunkum. The G4 apparently
> has a "math cheat" mode that doesn't fully implement IEEE math. The
> G5 doesn't have this mode. The GCC option that leads people to make
> this claim is non-functional for G5. On the other hand, there was
> that specialized malloc library.
>
> As far as I can tell, the G5 is a bandwidth monster. So, for any
> application needing multiple data streams running simultaneously
> it will be a lot faster than the competition. Think things like
> audio processing, video processing, perhaps databases.
>
> It is also something of a star at floating point performance. I don't
> know if it's better than P4/Xeons though, at minimum, it's probably
> just as good. Although the Mathematica and Luxology demos
> makes me believe it's probably better.
>
> But for a G5 to beat (not be competitive) with a P4 at scalar integer
> math we need to wait for the 2.5GHz G5.
>
> But all bets are off when the next version of the P4, called Prescott,
> is released. Then we'll have to start discussion all over again.

Arturo Pérez

08-01-2003, 03:33 PM

In article <3f2a3a45$1[at]news.unimelb.edu.au>,
"Vladimir Grubor" <vladimirgrubor<nospam>[at]hotmail.com> wrote:

> You are right, I stand corrected. I was under the impression that since it
> is a derivative of Power4, which is true 64bit processor, it must run 32 bit
> apps in emulation mode, like Itanium?

There's no emulation mode. On the Itanium there's essentially an
x86 intepreter built-in. The original specification for the PowerPC
ISA "requires" both 32/64 bit. Though, now that you mention, I wonder
what the processor does when it is in 32b mode and a 64b instruction
comes in?

There is apparently a bit of glue code (called the bridge) that
tells the processor if an executable is 32b or 64b. But the processor
runs full speed either way.

>
> Cheers,
>
> Vlad
>
> P.S. It might be superior, however, for intense computation (i.e. clusters,
> it is still cheaper to get a bunch of P4s and cluster them than a bunch of
> PowerPC970s. That is why I prefer PowerPC970 for desktop, but for intense
> calculations, cheaper to get P4.
> Do you agree with this?
>

That's a very reasonable attitude. What I'm waiting for are the IBM
PPC970 blades. I'd like to see their cost and their performance.

When the 970-based XServes come out you may want to take a look at
them. While "only" competitive with x86 compute farms they do have
some extremely nice management features built-in.

Charles Douglas Wehner

08-01-2003, 05:35 PM

I have a fundamental dislike of bought-in compilers like C. You simply
do not know what the program designer has put into it.

Writing in a macro-assembler enables you to optimise byte by byte.
Then you incorporate your libraries into your OWN in-house compiler.
You gain the benefit of a compiler and that of the fastest kernel.

I feel sure that Apple - who were in at the start of computing - work
this way. I also note - from his utterances - that Steve Jobs is very
aware of Forth. Forth is the ONLY disciplined study of SUBROUTINES.
That is to say, firstly of the machine-code kernel and then of the
groupings of calls. Adobe POSTSCRIPT is Forth.

If a system is tested on higher mathematics, the 64-bit architecture
may make a difference. However, I do not see that for everyday uses
such as Internet browsers there will be much more than a 10% advantage
over a 32-bit system. Indeed, by the use of 64-bit registers to hold
single bytes there may be no gain at all.

The gain comes from visiting and revisiting that kernal, tightening
those pieces of code that are most used.

Charles Douglas Wehner

"Vladimir Grubor" <vladimirgrubor<nospam>[at]hotmail.com> wrote in message news:<3f27b13d$1[at]news.unimelb.edu.au>...
> Someone (third party) at NASA did comparisons. G5 is actually PowerPC970,
> which is IBMs cpu which is in turn a poor man's version of Power4 (another
> server high end processor by IBM).
>
> G5 is a 64 bit processor, which runs in 32 bit emulation mode (less than
> ideal), since apple hasn't compiled MACOSX to support 64bit mode. Less than
> ideal situation, so for now, probably Xeon is slightly ahead (as shown on
> NASA test), but when apple sets up 64 bit support, I expect G5 to pull away
> in performance stakes. Also I did read the log file for independent test
> that Apple paid for. Math libraries were which are optimised for SPEC2000
> were used to replace normal math libraries. That was not done on a Linux.
>
> Cheers,
>
> Vlad
>

Matthew Russotto

08-01-2003, 09:34 PM

In article <e349085c.0308010835.436ae27b[at]posting.google.com>,
Charles Douglas Wehner <charleswehner[at]hotmail.com> wrote:
>I have a fundamental dislike of bought-in compilers like C. You simply
>do not know what the program designer has put into it.
>
>Writing in a macro-assembler enables you to optimise byte by byte.
>Then you incorporate your libraries into your OWN in-house compiler.
>You gain the benefit of a compiler and that of the fastest kernel.

Well, if YOU can keep the scheduling rules for an n-way superscalar
processor in your head and write assembler code to take advantage of
that, more power to you. I think I'll stick with the compiler

I know only one person who can hand-code assembly for a modern
processor and beat the compiler, and even he had to write tools to
help with the scheduling.
--
Matthew T. Russotto mrussotto[at]speakeasy.net
"Extremism in defense of liberty is no vice, and moderation in pursuit
of justice is no virtue." But extreme restriction of liberty in pursuit of
a modicum of security is a very expensive vice.

©z

08-05-2003, 05:25 AM

does anyone here have a Dual Xeon or a Dual G5 ?
--numbers on paper mean nothing--

i would not run out and buy a machine if it were 10% or even 40% faster
for $4000+
then have to deal with a "buggy" first gen 64bit os a few months later

i think i would just add more ram or change my videocard ?
1 Gig DDR 433 MHz, $474.99 us

<G5 vs ?> is really a non issue untill the price comes down

them Emacs [at] 1000mhz look pretty dam cool :)

and in the "real world "
http://www.sgi.com/features/2000/nov/xmen/
-some- of the fx's for x-men were done with a SGI O2
http://www.sgi.com/products/remarketed/o2/

faster is just faster-- not always better :)

"Matthew Russotto" <russotto[at]grace.speakeasy.net> wrote in message
news:XCWdncKWUJcfmLKiXTWc-w[at]speakeasy.net...
> In article <e349085c.0308020542.40550fd[at]posting.google.com>,
> Charles Douglas Wehner <charleswehner[at]hotmail.com> wrote:
> >russotto[at]grace.speakeasy.net (Matthew Russotto) wrote in message
news:<M-KcnYazX6RQU7eiXTWc-w[at]speakeasy.net>...
> >
> >&g