IBM cluster demolishes HP Superdome

[Yousuf Khan]

http://www.geek.com/news/geeknews/2003Nov/bch20031105022539.htm

[Felger Carbon]

"Yousuf Khan" <removethisspam.bjsk90.removethispam@hotmail.com> wrote in
message
news:Ig2rb.211421$3f.145691@twister01.bloor.is.net.cable.rogers.com... http://www.geek.com/news/geeknews/2003Nov/bch20031105022539.htm

also,
http://www.theinquirer.net/?article=12530

Yousuf, it seems IBM has the answer to your Opteron interconnect
problem, at least for a 48-CPU system.

[Yousuf Khan]

"Felger Carbon" <fmsfnf@jfoops.net> wrote in message
news:qy3rb.2144$nz.1408@newsread2.news.pas.earthlink.net... "Yousuf Khan" <removethisspam.bjsk90.removethispam@hotmail.com> wrote in message news:Ig2rb.211421$3f.145691@twister01.bloor.is.net.cable.rogers.com... http://www.geek.com/news/geeknews/2003Nov/bch20031105022539.htm also, http://www.theinquirer.net/?article=12530 Yousuf, it seems IBM has the answer to your Opteron interconnect problem, at least for a 48-CPU system.

Not really, these aren't really a single monolithic 48-way system, that can
be split into smaller subunits. Instead it's really 24 little 2-way systems,
connected via clustering software. As Robert Myers says, among other people,
there's a difference between supercluster vs. a supercomputer, and therefore
there is also a difference between a cluster vs. an enterprise server. It
all has to do latency and bandwidth in their communications links.

Yousuf Khan

[Robert Myers]

On Sat, 08 Nov 2003 18:06:47 GMT, "Yousuf Khan"
<removethisspam.bjsk90.removethispam@hotmail.com> wrote:
"Felger Carbon" <fmsfnf@jfoops.net> wrote in messagenews:qy3rb.2144$nz.1408@newsread2.news.pas.earthlink.net... "Yousuf Khan" <removethisspam.bjsk90.removethispam@hotmail.com> wrote in message news:Ig2rb.211421$3f.145691@twister01.bloor.is.net.cable.rogers.com... http://www.geek.com/news/geeknews/2003Nov/bch20031105022539.htm also, http://www.theinquirer.net/?article=12530 Yousuf, it seems IBM has the answer to your Opteron interconnect problem, at least for a 48-CPU system.Not really, these aren't really a single monolithic 48-way system, that canbe split into smaller subunits. Instead it's really 24 little 2-way systems,connected via clustering software. As Robert Myers says, among other people,there's a difference between supercluster vs. a supercomputer, and thereforethere is also a difference between a cluster vs. an enterprise server. Itall has to do latency and bandwidth in their communications links.

On the other hand, as a member of the market that one poster here
sneers at as insignificant, I can't help gloating a little. The
technical computing market may not buy alot, but it cares about
performance and it will take risks that the enterprise market won't.

IBM Opteron cluster sales for technical computing will be sales that
Intel would have liked to have had for Itanium. IBM Opteron clusters
will wind up taking customers away from enterprise customers, too, but
in that market they will be competing mostly against Xeon.

You can buy the IBM eSeries 1350 clusters with a high-performance
Myrinet interconnect, but I still don't expect IBM to be offering them
in the "enterprise" server market.

I don't expect any of this to scare Intel off IA-64, but it will get
them to pay more attention to the early-adopter technical computing
market. Naturally, I think that's a good thing. :-).

IBM

[Yousuf Khan]

"Robert Myers" <rmyers@rustuck.com> wrote in message
news:m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com... On the other hand, as a member of the market that one poster here sneers at as insignificant, I can't help gloating a little. The technical computing market may not buy alot, but it cares about performance and it will take risks that the enterprise market won't. IBM Opteron cluster sales for technical computing will be sales that Intel would have liked to have had for Itanium. IBM Opteron clusters will wind up taking customers away from enterprise customers, too, but in that market they will be competing mostly against Xeon.

Until and unless IBM or whoever else releases something with more than 4
Opteron processors in a chassis. At that point this will take it far beyond
the Xeon market and touch parts of the Itanium market.
You can buy the IBM eSeries 1350 clusters with a high-performance Myrinet interconnect, but I still don't expect IBM to be offering them in the "enterprise" server market. I don't expect any of this to scare Intel off IA-64, but it will get them to pay more attention to the early-adopter technical computing market. Naturally, I think that's a good thing. :-).

The specific SPEC benchmarks that they were highlighting were actually
scientific benchmarks, it looks like.

Yousuf Khan

[Rob Stow]

Yousuf Khan wrote:
"Robert Myers" <rmyers@rustuck.com> wrote in message news:m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com...On the other hand, as a member of the market that one poster heresneers at as insignificant, I can't help gloating a little. Thetechnical computing market may not buy alot, but it cares aboutperformance and it will take risks that the enterprise market won't.IBM Opteron cluster sales for technical computing will be sales thatIntel would have liked to have had for Itanium. IBM Opteron clusterswill wind up taking customers away from enterprise customers, too, butin that market they will be competing mostly against Xeon. Until and unless IBM or whoever else releases something with more than 4 Opteron processors in a chassis. At that point this will take it far beyond the Xeon market and touch parts of the Itanium market.

Newisys provides 8-way opteron motherboards to any system
manufacturer who wants them. Haven't heard of anyone
other than RackSaver who has taken them up on it.

You can buy the IBM eSeries 1350 clusters with a high-performanceMyrinet interconnect, but I still don't expect IBM to be offering themin the "enterprise" server market.I don't expect any of this to scare Intel off IA-64, but it will getthem to pay more attention to the early-adopter technical computingmarket. Naturally, I think that's a good thing. :-). The specific SPEC benchmarks that they were highlighting were actually scientific benchmarks, it looks like. Yousuf Khan


--
Reply to newsgroup only please. This e-mail account is real
but effectively abandoned because of excessive spamming.

[Yousuf Khan]

"Rob Stow" <rob.stow@sk.sympatico.ca> wrote in message
news:vqql413p97fm06@corp.supernews.com... Until and unless IBM or whoever else releases something with more than 4 Opteron processors in a chassis. At that point this will take it far
beyond the Xeon market and touch parts of the Itanium market. Newisys provides 8-way opteron motherboards to any system manufacturer who wants them. Haven't heard of anyone other than RackSaver who has taken them up on it.

Of course, I'm talking about the tier-ones. Actually it looks like Sun is
getting ready to release a 4-way server platform from Newisys.

Yousuf Khan

[Dean Kent]

"Robert Myers" <rmyers@rustuck.com> wrote in message
news:m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com... On the other hand, as a member of the market that one poster here sneers at as insignificant, I can't help gloating a little. The technical computing market may not buy alot, but it cares about performance and it will take risks that the enterprise market won't.

Interesting gloat. There are others who don't share your enthusiasm for the
article:

"...a single test (SPECchem) suffices him to entitle his article "IBM
Opteron-based eServer demolishes HP Superdome" He fails to notice the
Superdome is PA-RISC based (not the latest neither the greatest from HP but
he writes bravely "48-way IBM muscle outperforms HP's 64-way finest") and
fails to notice that he is comparing clusters against single image
systems.""
I don't expect any of this to scare Intel off IA-64, but it will get them to pay more attention to the early-adopter technical computing market. Naturally, I think that's a good thing. :-).

Certainly, but I think one has to step back and look at thing a bit more
skepically before declaring any overwhelming victories such as the author
attempts to do...

Regards,
Dean
IBM

[Robert Myers]

On Sat, 08 Nov 2003 23:42:17 GMT, "Dean Kent"
<dkent@realworldtech.com> wrote:
"Robert Myers" <rmyers@rustuck.com> wrote in messagenews:m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com... On the other hand, as a member of the market that one poster here sneers at as insignificant, I can't help gloating a little. The technical computing market may not buy alot, but it cares about performance and it will take risks that the enterprise market won't.Interesting gloat. There are others who don't share your enthusiasm for thearticle:"...a single test (SPECchem) suffices him to entitle his article "IBMOpteron-based eServer demolishes HP Superdome" He fails to notice theSuperdome is PA-RISC based (not the latest neither the greatest from HP buthe writes bravely "48-way IBM muscle outperforms HP's 64-way finest") andfails to notice that he is comparing clusters against single imagesystems.""

That, sir, is marketing.

"The bigger the humbug, the more people like it."

The point here is not whether or not some particular machine is more
suited to some task than some other machine. The point is that IBM is
aggressively marketing the merits of Opteron for
computationally-intensive applications.

http://www.dimemuseum.com/news_article12.htm

The Dime Museum gets its name from the carnival-like museums of the
19th century, where, for a dime, truth was stretched thinner than the
taffy sold on a carnival midway.

"Is it real?" customers often excitedly ask Mr. Horne and Mr. Taylor
as they examine what are described as George Washington's false
eyelashes, the fragile-looking black prostheses propped up on two
small pieces of wood. The curators answer only with smiles.

"People don't mind being humbugged as long as they're being let in on
the joke," said Mr. Horne, paraphrasing 19th-century American showman
P.T. Barnum. "Not knowing how much of what you're seeing is real is
most of the fun. You have to use your sense of humor and sense of
judgment."

Buyers of hardware for computationally-intensive tasks aren't going to
be persuaded by humbug, but they might have their interest aroused.

In going after this market, IBM is going straight at people who buy a
dozen rackmounts (with Intel processors) from Dell and cable them
together with ethernet. People who want a shiny new high-end
Superdome from HP or an Altix from SGI and have the money to pay for
it are probably going to buy one, anyway. Or maybe they won't, if
they think they can get the same ego boost for less money. That's
where the humbug comes in.

RM

[Dean Kent]

"Robert Myers" <rmyers@rustuck.com> wrote in message
news:e22rqvcvmsklku7mrutjm7adto71i503eq@4ax.com... That, sir, is marketing.

Well, not exactly. It is fanboy journalism. Another (different) person
making a similar statement:

"Very stupid comparison - a cluster against a single-image system, optimized
for business applications. Cluster is always better on so called
"embarrassingly-parallel" workloads.
Let's see IBM cluster against HP cluster build of zx2000 workstations or
rx2600 small servers. I have no doubts that HP would will SPEC Chem2002 by
wide margin.

The lack of technical knowledge demonstrated by some the Inquirer writers is
incredible."

Buyers of hardware for computationally-intensive tasks aren't going to be persuaded by humbug, but they might have their interest aroused.

Perhaps, but the other side of it is whether those who actually understand
simply become more immune to the marketing/fanboism, and ignore the part
because of its reputation. Add to that, Newegg.com is selling Opterons -
so, are they HPC parts, IT parts or home-game parts? A reputation is a
two-edged sword...

Regards,
Dean

[Robert Myers]

On Sun, 09 Nov 2003 01:21:26 GMT, "Dean Kent"
<dkent@realworldtech.com> wrote:
"Robert Myers" <rmyers@rustuck.com> wrote in messagenews:e22rqvcvmsklku7mrutjm7adto71i503eq@4ax.com... That, sir, is marketing.Well, not exactly. It is fanboy journalism. Another (different) personmaking a similar statement:"Very stupid comparison - a cluster against a single-image system, optimizedfor business applications. Cluster is always better on so called"embarrassingly-parallel" workloads.Let's see IBM cluster against HP cluster build of zx2000 workstations orrx2600 small servers. I have no doubts that HP would will SPEC Chem2002 bywide margin.The lack of technical knowledge demonstrated by some the Inquirer writers isincredible."
Read laziness. I can't find the actual press release on the IBM web
site, but if you go to

http://www.supercomputingonline.com/article.php?sid=4912

you will probably find nearly the straight text of the press release,
since the text of the web page reads like a press release.

"We are leading the competition by offering customers an expanded set
of building blocks for Deep Computing that provide the power behind
everything from petroleum exploration to digital animation," said
David Turek, vice president, IBM Deep Computing. "Customers in all
segments, including the very high-end, are no longer interested in
building their own systems. The IBM eServer Cluster 1350 ensures
customers a dependable solution with a short deployment time."

Practically the entire page goes on in that vein. The only thing
missing is the "-30-" at the end. Since you are an engineer and not a
journalist or marketeer by profession, it apparently doesn't occur to
you that people with space to fill take stuff pretty much straight out
of press releases. Publicists count on the fact that many people
will. That's marketing.

Just to be sure you *really* wouldn't do the same thing, I took a
quick hop over to realworldtech.com to look at your news articles.
Either you are getting a higher quality of press release than most web
sites, or somebody is putting some thought into what they write.

Add to that the fact that the people writing the press releases are
the ones who buy advertising, and you really don't need The Inquirer
to get your marketing material presented as news.

But why am I telling you this? You run a web site. This stuff is
probably all straight off an RSS feed.
Buyers of hardware for computationally-intensive tasks aren't going to be persuaded by humbug, but they might have their interest aroused.Perhaps, but the other side of it is whether those who actually understandsimply become more immune to the marketing/fanboism, and ignore the partbecause of its reputation. Add to that, Newegg.com is selling Opterons -so, are they HPC parts, IT parts or home-game parts? A reputation is atwo-edged sword...
That's true. Somebody at Intel thought this one through a long time
ago. When that lightbulb clicked on, x86 was doomed at Intel.

RM

[Thomas Edison]

Rob Stow <rob.stow@sk.sympatico.ca> wrote: Yousuf Khan wrote:
Until and unless IBM or whoever else releases something with more than 4 Opteron processors in a chassis. At that point this will take it far beyond the Xeon market and touch parts of the Itanium market.
Newisys provides 8-way opteron motherboards to any system manufacturer who wants them. Haven't heard of anyone other than RackSaver who has taken them up on it.

I couldn't find an 8P box on Racksaver's web site. Looks like they're
all currently 4P. Is this an "upcoming product" or something that is
current? URL?


--
davewang202(at)yahoo(dot)com

[Dean Kent]

"David Wang" <foo@bar.invalid> wrote in message
news:bokgsj$hk0$1@grapevine.wam.umd.edu... I couldn't find an 8P box on Racksaver's web site. Looks like they're all currently 4P. Is this an "upcoming product" or something that is current? URL?

I was going to post the same question. Last time someone told me that
Newisys had something greater than 4-way implementations, I could only find
a mention of 'plans' to do so on their site. Admittedly, that was about 3
or 4 months ago, but is it possible that rumors continue to be used as fact?

Regards,
Dean

-- davewang202(at)yahoo(dot)com

[Felger Carbon]

"David Wang" <foo@bar.invalid> wrote in message
news:bokgsj$hk0$1@grapevine.wam.umd.edu... I couldn't find an 8P box on Racksaver's web site. Looks like they're all currently 4P.

Let me suggest possible reasons why you don't find that 8P box, David.

I have some questions about (hypothetical??) 8P Opteron systems. I
understand that Opterons have circuitry to enable a glueless 8P
implementation. However, I have some concerns over the physical
configuration of such a system.

While it should be possible to physically mount 8 CPUs on one 15" square
standard rack-mount board, it would seem to require heroic measures to
cool such a beast. IMHO this is why you see and hear about lotsa 4P
systems, but not much 8P.

To build an 8P system with standard cooling, it requires _two_ 15"
square standard rack-mount boards. In that case, how does one arrange
for the high-speed HyperTransport links to travel from board to board?

Another option is to use a non-standard sized motherboard. I don't
believe this will happen. The only other option is heroic cooling
measures. I doubt that this is a viable solution.

I will now sit back and let myself be buried under the deluge of
examples of 8P systems that have (really, honestly) already been
delivered to real paying customers. ;-)

[Robert Myers]

On Sun, 09 Nov 2003 09:36:31 GMT, "Felger Carbon" <fmsfnf@jfoops.net>
wrote:

<snip>To build an 8P system with standard cooling, it requires _two_ 15"square standard rack-mount boards. In that case, how does one arrangefor the high-speed HyperTransport links to travel from board to board?

Over a backplane. All it takes is money. Theoretically, you can do
it with a cable. All that takes is money.

PCI Express (which will be available *real soon now*) makes it
unlikely that anybody will ever come up with that money.

RM

[Thomas Edison]

Felger Carbon <fmsfnf@jfoops.net> wrote: "David Wang" <foo@bar.invalid> wrote in message news:bokgsj$hk0$1@grapevine.wam.umd.edu... I couldn't find an 8P box on Racksaver's web site. Looks like they're all currently 4P.
Let me suggest possible reasons why you don't find that 8P box, David.
I have some questions about (hypothetical??) 8P Opteron systems. I understand that Opterons have circuitry to enable a glueless 8P implementation. However, I have some concerns over the physical configuration of such a system.
While it should be possible to physically mount 8 CPUs on one 15" square standard rack-mount board, it would seem to require heroic measures to cool such a beast. IMHO this is why you see and hear about lotsa 4P systems, but not much 8P.

<hypothetical cooling problems compressed>

A little bird that flew by my window had given me the impression
that 8P boxes had been built, tested, and scalability found to
be poor, and that was the reason given for the lack of
commercialization of glueless 8P boxes. The little bird also
left open the possibility that someone would pay to purchase
such a system anyways, even with such poor scalability.

I've been looking for such a system ever since, and every time
someone says XYZ has such an 8P configuration, I have dutifully
chased down the link to null.

Here is another exmaple of someone saying that Racksaver sells
the 8P configuration, and I couldn't find it.

Perhaps it's a combination of Processor cost (AMD charges
extraodinary amounts of money for 8P capable processors),
cooling and poor scalability that we don't see glueless 8P
boxes, but there's always the possibility that someone will
sell such a beast... I'm still looking.

--
davewang202(at)yahoo(dot)com

[Keith R. Williams]

In article <m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com>,
rmyers@rustuck.com says... On Sat, 08 Nov 2003 18:06:47 GMT, "Yousuf Khan" <removethisspam.bjsk90.removethispam@hotmail.com> wrote:"Felger Carbon" <fmsfnf@jfoops.net> wrote in messagenews:qy3rb.2144$nz.1408@newsread2.news.pas.earthlink.net... "Yousuf Khan" <removethisspam.bjsk90.removethispam@hotmail.com> wrote in message news:Ig2rb.211421$3f.145691@twister01.bloor.is.net.cable.rogers.com... > http://www.geek.com/news/geeknews/2003Nov/bch20031105022539.htm also, http://www.theinquirer.net/?article=12530 Yousuf, it seems IBM has the answer to your Opteron interconnect problem, at least for a 48-CPU system.Not really, these aren't really a single monolithic 48-way system, that canbe split into smaller subunits. Instead it's really 24 little 2-way systems,connected via clustering software. As Robert Myers says, among other people,there's a difference between supercluster vs. a supercomputer, and thereforethere is also a difference between a cluster vs. an enterprise server. Itall has to do latency and bandwidth in their communications links. On the other hand, as a member of the market that one poster here sneers at as insignificant, I can't help gloating a little. The technical computing market may not buy alot, but it cares about performance and it will take risks that the enterprise market won't.

If you're talking about me, you're even more clueless than I had
ever guessed. You constantly misrepresent my position.

Yes, there are clusters being made from off-the-shelf processors
that will solve some of your issues. However, your problem (the
one I have been addressing) is that their vector processors don't
do what you want. Sorry, but no one cares about your problem
when designing a CPU. It's up to you to figure out how to use
what is available to the many, for the one.

Good grief, you have a marvelous ability to move goal posts.
IBM Opteron cluster sales for technical computing will be sales that Intel would have liked to have had for Itanium. IBM Opteron clusters will wind up taking customers away from enterprise customers, too, but in that market they will be competing mostly against Xeon.

....and this is somehow badness? Good grief this is the whole
business plan behind Opteron!

OTOH, I haven't seen a cogent plan behind Itanic. ...in six
years.
You can buy the IBM eSeries 1350 clusters with a high-performance Myrinet interconnect, but I still don't expect IBM to be offering them in the "enterprise" server market.

You never know. What makes money, IBM will sell.
I don't expect any of this to scare Intel off IA-64, but it will get them to pay more attention to the early-adopter technical computing market. Naturally, I think that's a good thing. :-)

I don't expect IA64 to go away either. OTOH, I see nothing that
would give it the universal acceptance that many were cheering
for in the mid 90's. In those terms, it's an abysmal failure.

--
Keith

[Dean Kent]

"Keith R. Williams" <krw@attglobal.net> wrote in message
news:MPG.1a18bdffc3bd37bf98a879@enews.newsguy.com... In article <m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com>, rmyers@rustuck.com says... On the other hand, as a member of the market that one poster here sneers at as insignificant, I can't help gloating a little. The technical computing market may not buy alot, but it cares about performance and it will take risks that the enterprise market won't. If you're talking about me, you're even more clueless than I had ever guessed. You constantly misrepresent my position.

Gee. I thought he was talking about me. Quit trying to take credit for
everything - back of the line!!!

Regards,
Dean

[Robert Myers]

On Sun, 9 Nov 2003 21:31:58 -0500, Keith R. Williams
<krw@attglobal.net> wrote:
In article <m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com>,rmyers@rustuck.com says...

<snip>
...your problem (theone I have been addressing) is that their vector processors don'tdo what you want. Sorry, but no one cares about your problemwhen designing a CPU. It's up to you to figure out how to usewhat is available to the many, for the one.

Sales of pharmaceuticals generate about a half a trillion dollars in
revenues. The equations that govern biologically active molecules are
not a mystery, nor are means of solving those equations unknown. What
is mostly lacking is computational muscle.

If anyone *knew* that building a particular computer architecture
would allow the solution of those equations with acceptable accuracy
in an acceptable period of time, the money would condense around the
problem as if out of thin air.

As it is, there is no classical computer architecture that is
available now or in the forseeable future that would do for molecular
biology what computers have done for aerodynamics.

From that point of view, it doesn't matter much what you or I or
anyone else thinks about it, because for all practical purposes, it
can't be done. Throwing money at it would do very little good.

Vector processors, as in the Cray 1, are probably a dead end.
Streaming architectures, which are a generalization of the same idea,
might not be, but people have lost enormous amounts of money trying to
make that idea commercially successful. The Cell processor probably
implements some kind of streaming architecture; it is hard for me to
imagine how else they could get anything close to the claimed
throughput.

People do build specialized processors, such as the Grape procoesor,
which also has some staggering throughput and which is useful mostly
for scientific calculations.

The current fad of thowing together huge numbers of COTS processors to
hit some Linpack benchmark is incredibly discouraging. The
interconnect fabric of RedStorm, which makes RedStorm the pick of the
litter for that class of computer, is very clever, but I have real
concerns about whether the architecture really makes sense for large
classes of important problems. One user of the T3E, of which RedStorm
is a technological descendent, commented that it was practically
useless for time-dependent problems.
Good grief, you have a marvelous ability to move goal posts.

I have no idea where or what you think my goal posts are.
IBM Opteron cluster sales for technical computing will be sales that Intel would have liked to have had for Itanium. IBM Opteron clusters will wind up taking customers away from enterprise customers, too, but in that market they will be competing mostly against Xeon....and this is somehow badness? Good grief this is the wholebusiness plan behind Opteron!
I was commenting on what I perceived to be the hope that Opteron would
become the dominant 64-bit architecture. I see very little chance of
that happening, but you never know.
OTOH, I haven't seen a cogent plan behind Itanic. ...in sixyears. You can buy the IBM eSeries 1350 clusters with a high-performance Myrinet interconnect, but I still don't expect IBM to be offering them in the "enterprise" server market.You never know. What makes money, IBM will sell.

You'll have to do better than Myrinet to turn Opteron into a candidate
for huge shared-data problems. It could be done with hyperlink, but I
don't see anybody moving in that direction.

RM

[Tony Hill]

On Sun, 09 Nov 2003 01:21:26 GMT, "Dean Kent"
<dkent@realworldtech.com> wrote:"Robert Myers" <rmyers@rustuck.com> wrote in messagenews:e22rqvcvmsklku7mrutjm7adto71i503eq@4ax.com... That, sir, is marketing.Well, not exactly. It is fanboy journalism. Another (different) personmaking a similar statement:"Very stupid comparison - a cluster against a single-image system, optimizedfor business applications. Cluster is always better on so called"embarrassingly-parallel" workloads.Let's see IBM cluster against HP cluster build of zx2000 workstations orrx2600 small servers. I have no doubts that HP would will SPEC Chem2002 bywide margin.

I don't think that the above is correct. While the widely touted
number is comparing the 24-node IBM dual-Opteron system to a
64-processor PA-RISC system from HP, IBM did also release another
result.

In the SpecCHEM2002 "Small" test, the Opteron system WAS tested
against an HP Itanium2 system. Ok, the Itanium2s were of the older
"McKinley" variety, running at only 1.0GHz and with 3MB of L3 cache as
opposed to the 1.5GHz clock and 6MB of L3 for current "Madison"
variety, however in this test the Opterons really DID demolish them.
A single 2-processor Opteron server managed to beat out a 4-processor
Itanium2 system by almost 10%. When IBM combined a pair of the
Opteron servers in a cluster they were SIGNIFICANTLY faster, scaling
basically linearly with the number of processors.

This suggests to me that in the "small" test at least, the Opteron
should have little trouble holding it's own against the latest and
greatest Itanium2 chips, and they can often even come out on top.

-------------
Tony Hill
hilla <underscore> 20 <at> yahoo <dot> ca

[chrisv]

On Sat, 08 Nov 2003 09:00:24 GMT, "Yousuf Khan"
<removethisspam.bjsk90.removethispam@hotmail.com> wrote:
http://www.geek.com/news/geeknews/2003Nov/bch20031105022539.htm

Funny how a flame war ensued in the discussion forum... What a shock.

[Robert Myers]

On Mon, 10 Nov 2003 00:00:54 -0500, Robert Myers <rmyers@rustuck.com>
wrote:

<snip>You'll have to do better than Myrinet to turn Opteron into a candidatefor huge shared-data problems. It could be done with hyperlink, but Idon't see anybody moving in that direction.

I meant it could be done with hyper*transport*, of course.

RM

[Rob Stow]

Dean Kent wrote: "David Wang" <foo@bar.invalid> wrote in message news:bokgsj$hk0$1@grapevine.wam.umd.edu...I couldn't find an 8P box on Racksaver's web site. Looks like they'reall currently 4P. Is this an "upcoming product" or something that iscurrent? URL? I was going to post the same question. Last time someone told me that Newisys had something greater than 4-way implementations, I could only find a mention of 'plans' to do so on their site. Admittedly, that was about 3 or 4 months ago, but is it possible that rumors continue to be used as fact?

I talked to someone at RackSaver a couple of hours ago to
ask where the 8P information is. Apparently they have never
received more than a "reference board" for 8P so they are not
going to offer 8P systems after all. He could only suggest
directly contacting AMD and NewiSys for more info about 8P
systems.




--
Reply to newsgroup only please. This e-mail account is real
but effectively abandoned because of excessive spamming.

[Keith R. Williams]

In article <8FErb.25739$M_4.18894@newssvr27.news.prodigy.com>,
dkent@realworldtech.com says... "Keith R. Williams" <krw@attglobal.net> wrote in message news:MPG.1a18bdffc3bd37bf98a879@enews.newsguy.com... In article <m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com>, rmyers@rustuck.com says... > On the other hand, as a member of the market that one poster here sneers at as insignificant, I can't help gloating a little. The technical computing market may not buy alot, but it cares about performance and it will take risks that the enterprise market won't. If you're talking about me, you're even more clueless than I had ever guessed. You constantly misrepresent my position. Gee. I thought he was talking about me. Quit trying to take credit for everything - back of the line!!!

Nah Deano, he has a hard-on for me. I keep trying to tell him
where "it's at", but he won't believe anyone as unplugged as I.

--
Keith

[Keith R. Williams]

In article <ph2uqvks107iqcsivc0gqegde7vbokmh1b@4ax.com>,
rmyers@rustuck.com says... On Sun, 9 Nov 2003 21:31:58 -0500, Keith R. Williams <krw@attglobal.net> wrote:In article <m1eqqvkghe2r6r9jhs3bls7an93vi9cgv1@4ax.com>,rmyers@rustuck.com says... <snip>...your problem (theone I have been addressing) is that their vector processors don'tdo what you want. Sorry, but no one cares about your problemwhen designing a CPU. It's up to you to figure out how to usewhat is available to the many, for the one. Sales of pharmaceuticals generate about a half a trillion dollars in revenues. The equations that govern biologically active molecules are not a mystery, nor are means of solving those equations unknown. What is mostly lacking is computational muscle.

Oh, goodie! Please tell me how that translates into a *BILLION*
dollars (or more) to fund your new wizbang chip, when others glue
together chips made for mere gamers? As I've said, get used to
it and live with the cheap chips you have, or pay up! So far
I've only seen you whine (constantly) about your problem, no
money.
If anyone *knew* that building a particular computer architecture would allow the solution of those equations with acceptable accuracy in an acceptable period of time, the money would condense around the problem as if out of thin air.

Where's the *money*? You don't know (obviously), meanwhile
people are designing chips for those who do. If you can lash
them together to make something interesting for you, go for it.
If not, please stop whining about your "special" problem. It's
getting tiring. That is my point.
As it is, there is no classical computer architecture that is available now or in the forseeable future that would do for molecular biology what computers have done for aerodynamics.

Whine some more. No one cares! We design processors for people
who know what they want and can *pay*. So far you've shown that
you are neither. Suck a lemon.
From that point of view, it doesn't matter much what you or I or anyone else thinks about it, because for all practical purposes, it can't be done. Throwing money at it would do very little good.

Gee, I thought I told you that! (though I thought your real
issue was FP precision, and the ones who have money don't care)
Vector processors, as in the Cray 1, are probably a dead end.

Are? Were! Get over your fetish for the Cray-1. That was a
quarter-centry ago! Get with the real world.

<snip - cannot comment>
People do build specialized processors, such as the Grape procoesor, which also has some staggering throughput and which is useful mostly for scientific calculations.

You do like to ramble, don't you?

Sorry, I don't have the time <snip>

--
Keith

[Robert Myers]

On Tue, 11 Nov 2003 21:54:32 -0500, Keith R. Williams
<krw@attglobal.net> wrote:
In article <ph2uqvks107iqcsivc0gqegde7vbokmh1b@4ax.com>,rmyers@rustuck.com says... On Sun, 9 Nov 2003 21:31:58 -0500, Keith R. Williams <krw@attglobal.net> wrote:

<snip>
...your problem (theone I have been addressing) is that their vector processors don'tdo what you want. Sorry, but no one cares about your problemwhen designing a CPU. It's up to you to figure out how to usewhat is available to the many, for the one. Sales of pharmaceuticals generate about a half a trillion dollars in revenues. The equations that govern biologically active molecules are not a mystery, nor are means of solving those equations unknown. What is mostly lacking is computational muscle.Oh, goodie! Please tell me how that translates into a *BILLION*dollars (or more) to fund your new wizbang chip, when others gluetogether chips made for mere gamers?

I don't have a whizbang chip to sell. If I did, I'd know where to
find the money to build it. I have a pretty good idea of what people
are up to and what's getting funded, but there's nothing on the
horizon I'd want to lose sleep or money over.

I'm with Felger on this one. It's better to let someone else take the
risks.

<snip>
... I thought your realissue was FP precision, and the ones who have money don't care)

Now I understand what you've been going on about. Yes, it would be
very nice if people who build whiz-bang GPU's and DSP chips did them
in 64-bit precision. It would be nice if IBM's embeddable Power4
supported 64-bit precision. It's not something I'm losing sleep over.

By the time game chips are being built with 64-bit precision,
scientific computation will probably be needing 128 bits, anyway.

RM

[Keith R. Williams]

In article <l0a3rvk5dnbbneug8v8tecuvh9rj09n28g@4ax.com>,
rmyers@rustuck.com says... On Tue, 11 Nov 2003 21:54:32 -0500, Keith R. Williams <krw@attglobal.net> wrote:In article <ph2uqvks107iqcsivc0gqegde7vbokmh1b@4ax.com>,rmyers@rustuck.com says... On Sun, 9 Nov 2003 21:31:58 -0500, Keith R. Williams <krw@attglobal.net> wrote: <snip> >...your problem (the >one I have been addressing) is that their vector processors don't >do what you want. Sorry, but no one cares about your problem >when designing a CPU. It's up to you to figure out how to use >what is available to the many, for the one. > Sales of pharmaceuticals generate about a half a trillion dollars in revenues. The equations that govern biologically active molecules are not a mystery, nor are means of solving those equations unknown. What is mostly lacking is computational muscle.Oh, goodie! Please tell me how that translates into a *BILLION*dollars (or more) to fund your new wizbang chip, when others gluetogether chips made for mere gamers? I don't have a whizbang chip to sell. If I did, I'd know where to find the money to build it. I have a pretty good idea of what people are up to and what's getting funded, but there's nothing on the horizon I'd want to lose sleep or money over.

Obviously, though you keep whining about the industry that
doesn't care about your "needs". You pine about the days of your
Cray-1. Give it up! No one cares. If you can build super
computers out of commercial building-blocks, then you'll be
happy. Otherwise, you'll continue to be frustrated. Get over
it! I'm with Felger on this one. It's better to let someone else take the risks.

Yet you continue to whine!... I thought your realissue was FP precision, and the ones who have money don't care) Now I understand what you've been going on about.

Oh no, that's just part of it! You continually whine because
computer architects don't care about you. FP precision is simply
the most obvious stake.
Yes, it would be very nice if people who build whiz-bang GPU's and DSP chips did them in 64-bit precision. It would be nice if IBM's embeddable Power4 supported 64-bit precision. It's not something I'm losing sleep over.

Well, the FPUs do double-precision FP. The vector unit doesn't.
Too bad, you weren't in the target market.
By the time game chips are being built with 64-bit precision, scientific computation will probably be needing 128 bits, anyway.

Exactly my point! It's good that you finally got religion!

--
Keith

[Robert Myers]

On Wed, 12 Nov 2003 21:42:23 -0500, Keith R. Williams
<krw@attglobal.net> wrote:


<snip>Obviously, though you keep whining about the industry thatdoesn't care about your "needs". You pine about the days of yourCray-1. Give it up! No one cares.

That's your (limited) perspective.

http://www.research.ibm.com/resources/news/20030827_trips.shtml

If you can build supercomputers out of commercial building-blocks, then you'll behappy. Otherwise, you'll continue to be frustrated. Get overit!

NEC's Earth Simulator changed that. DARPA got over its woozy 90's
fantasy about COTS, which you apparently are still stuck in.

The COTS fantasy will never completely die because you can get alot of
bang for the buck. I ran across a document

http://www.bayarea.net/~kins/AboutMe/CPUs.html

that gives a really nice summary of what's out there and what you get
for what you pay, at least in very crude terms.

<snip>Exactly my point! It's good that you finally got religion!

You make alot of assumptions about me and about the world around you.
Technology in the US always has been and always will be driven by
considerations you don't see, at least as long as the US remains a
superpower, and I'm not talking about Microsoft-Intel conspiracies.

I won't be able to stop you from speculating publicly about what I
think I know or what my attitude toward the whole business is, so I
won't try. I'll just say that your interpersonal style is
stupefyingly presumptuous.

RM