New IBM Super Computer

[Bob Ayers]

Amazing performance, yet no details on the IC technology involved. Liquid/gas cooled?

Ironically, it's probably crunching these huge datasets using FORTRAN compilers and IMSL libraries with roots going back over 40 years... ye olde S/W tools are just so thoroughly debugged, i.e. Verified and Validated (V&V), and so focused on quad-precision floating-point and vector processing.

Sleep well America, secure in the knowledge that NO MicroSquish products are used when it really, REALLY counts.

IBM Zurich just announced this new chip cooling technology:

International Business Machines Corp. said scientists in its Zurich laboratories have demonstrated a way to cool microprocessor chips by pumping water through hair-thin channels in a development that may pave the way for speeding computation by stacking chips atop each other.

Stacked, or three-dimensional, microchips have been regarded as one of the most promising approaches to enhancing chip performance. Stacked chips could shorten the distance that electrons need to travel to just 1/1,000th of the distance they would typically travel in side-by-side chips, moving data faster. But heat buildup between stacked chips would quickly render them useless.

Thomas Brunschwiler, project leader at IBM's Zurich Research Laboratory, said conventional cooling mechanisms can't dissipate the heat. IBM researchers, working in collaboration with scientists at the Fraunhofer Institute in Berlin, developed a method to use cooling water, in a new approach, and to keep it from disrupting the electrical work of the chips.

"By implementing the idea of taking water to the chip, they have taken stacked chips closer to commercialization," said Yogendra Joshi, chair of the school of mechanical engineering at Georgia Institute of Technology, who chaired a conference where the IBM work was reported.

Mr. Brunschwiler that said he anticipates the stacked chips will start showing up in high-performance scientific computers in five to 10 years. Prof. Joshi said he thinks the technology will eventually have broad applications in corporate data centers.

"The challenge was to insulate water from interfering with the electrical signals," Mr. Brunschwiler said. "It's like the human brain, where you mix electrical signals from the neurons with blood vessels that provide cooling and oxygen."

Mr. Brunschwiler said researchers piped water through a horizontal test structure, consisting of a cooling layer between two heat sources. The cooling layer measures only about 100 microns in height and is packed with 10,000 vertical wires known as "vias" that connect the chips above and below. A micron is 1/1,000th of a millimeter. The scientists hermetically sealed the vias by leaving a silicon wall around each and adding a layer of silicon oxide to insulate it to prevent water from causing electrical shorts

Mr. Brunschwiler said his team is working on cooling systems for even smaller chip dimensions and is trying to figure out ways to bring added coolant to exceptionally hot spots in the chip. He said the team had investigated other fluids but concluded water was best because it flowed most easily and didn't cause environmental problems.

 

[MAKG]

Jeez, liquid cooling has been around for DECADES. IBM just discovered it?

There are a bunch of old Cray X-MPs around the country (one is in the NAS lobby here, functioning as a bench). Those had liquid cooling. The T3E I cut my teeth on (also a Cray) had liquid cooling and it would crash in seconds if it was interrupted (FN building contractors did that more than once).

And 3D CPU/memory arrays were an SGI idea, more than 10 years ago (before they bought Cray).

From my own experience, yes, people still run massive Fortran programs on the superdupercomputers, but some do use C. OO doesn't work very well for the heavily optimized performance these guys are looking for, so you see very little C++ and no Java. My thesis code was 13,000 lines of Fortran-77, using MPI (still a standard).

One has to be aware of what is going on simultaneously on all the processors or it won't scale well. That's hard to do with OO data hiding. Not impossible though; there have been a couple of OO framework projects.

 

[Bob Ayers]

Jeez, liquid cooling has been around for DECADES. IBM just discovered it?

.

Not at the chip level, the source of the heat!! IBM's new idea eliminates package thermal resistance.....

Still waiting on you to produce that 950 teraflop Columbia Supercomputer that held the record before IBM's Roadrunner!!

 

[MAKG]

What I said was that there was no significant difference between 950 teraflops and 1 petaflop. NOT that Columbia produced that performance. It didn't.

 

[Bob Ayers]

What I said was that there was no significant difference between 950 teraflops and 1 petaflop. NOT that Columbia produced that performance. It didn't.

This is what you posted:

And if you think it's important, why is it more important than the Columbia supercomputer? Never heard of it? It's the previous record holder, and it also does climate change work, plus a few other things like galaxy formation, solar models, reentry dynamics, and other nasty fluids calculations.

But it doesn't have the magic odometer rollover number.

THAT'S why it's a big yawn. The difference between a petaflop and 950 teraflops (for instancE) is not significant, but it WILL make the difference over whether AP pays attention.



If you weren't talking about the Columbia, what computer were you talking about then? You also said the Columbia was the previous record holder, which it wasn't, it was IBM's Blue Gene!

 

[MAKG]

This is what you posted:

And if you think it's important, why is it more important than the Columbia supercomputer? Never heard of it? It's the previous record holder, and it also does climate change work, plus a few other things like galaxy formation, solar models, reentry dynamics, and other nasty fluids calculations.

But it doesn't have the magic odometer rollover number.

THAT'S why it's a big yawn. The difference between a petaflop and 950 teraflops (for instancE) is not significant, but it WILL make the difference over whether AP pays attention.



If you weren't talking about the Columbia, what computer were you talking about then? You also said the Columbia was the previous record holder, which it wasn't, it was IBM's Blue Gene!

What do you think "(for instance)" means?

 

[Bob Ayers]

What I said was that there was no significant difference between 950 teraflops and 1 petaflop. NOT that Columbia produced that performance. It didn't.

The Columbia wasn't the previous record holder as YOU claimed either!!!!

 

[MAKG]

The Columbia wasn't the previous record holder as YOU claimed either!!!!

You're right. It was A previous record holder. They change quickly, and it's not important which is the fastest. Computing is not NASCAR, and it is ENTIRELY possible to get a faster calculation out of a slower computer. CPU speed isn't even close to the only variable. Algorithmic synchronizations are FAR more important, as is cache behavior. And when it takes a year to engineer away a factor of two for a calculation that would have taken a few months, it's stupid to waste the time.

Why don't we ooh and aah over its color. It's only slightly less relevant than a small increase in aggregate speed.

An old saying we used to use at NASA HPC was that the only meaningful statistic was the mean time to publication.

 

[Bob Ayers]

You're right. It was A previous record holder. They change quickly, and it's not important which is the fastest. Computing is not NASCAR, and it is ENTIRELY possible to get a faster calculation out of a slower computer. CPU speed isn't even close to the only variable. Algorithmic synchronizations are FAR more important, as is cache behavior. And when it takes a year to engineer away a factor of two for a calculation that would have taken a few months, it's stupid to waste the time.

Why don't we ooh and aah over its color. It's only slightly less relevant than a small increase in aggregate speed.

An old saying we used to use at NASA HPC was that the only meaningful statistic was the mean time to publication.

At 88.88 teraflops, the Columbia isn't even close!!!!!!

 

[MAKG]

So, you think it's BS.

What's your PERSONAL computation record? Have you ever designed a computer that made it onto TOP500? You simply have no idea.

FYI, Columbia is several years old. The speed record goes up FAST. 10 years ago, we were really excited about 1 TERAflop. By your stat, Columbia is 90 times faster. And it was #1 in 2004.

Speed records are silly. When the Beowulf cluster I designed made it to #163, I really didn't care; it was designed to solve specific problems in astrophysics, not break silly speed records.

 

[Bob Ayers]

So, you think it's BS.

What's your PERSONAL computation record? Have you ever designed a computer that made it onto TOP500? You simply have no idea.

FYI, Columbia is several years old. The speed record goes up FAST. 10 years ago, we were really excited about 1 TERAflop. By your stat, Columbia is 90 times faster. And it was #1 in 2004.

Speed records are silly. When the Beowulf cluster I designed made it to #163, I really didn't care; it was designed to solve specific problems in astrophysics, not break silly speed records.

"Speed records are silly"



The "flop" is the industry standard for computer speed! Too bad you can't accept it!!!

It sounds like you haven't kept up with the latest technology at all!! You thought that Columbia was number one before Roadrunner!!!

 

[Evan]

"Speed records are silly"

Try telling that to a business who's trying to crunch a batch of 1,000,000 financial transactions before the next business day starts.

They sure care about speed. 80% of the world's business logic still runs on COBOL. It's my job to lower that number.

 

[MAKG]

So, Bob, why do YOU care how many in-cache multiplications or additions some computer can do? What's the significance?

No, I've been out of the HPC field for several years. And as I said, it's silly and I didn't care who was the "best" even at the time.

HPC is not NASCAR. It's not penis comparison. The import of a computer is what it can calculate, and that is NOT measured in flops. There isn't any perfect measure, but a good one is literature citations.

 

[MAKG]

Try telling that to a business who's trying to crunch a batch of 1,000,000 financial transactions before the next business day starts.

They sure care about speed. 80% of the world's business logic still runs on COBOL. It's my job to lower that number.

Speed. Not speed records. That computer would be useless for financial transactions written in COBOL. A traditional ethernet-linked weakly bound cluster would make more sense for that.

But it ain't gonna win records.

And I don't think you're understanding the scale. A million transactions per day can be done on MUCH slower computers. That thing calculates around a trillion fluid cells per day (based on the order of 1000 flops per grid cell). With a 1000^3
grid and 1000 timesteps, that gets you in the ballpark for some types of nuclear explosions being calculated in a few weeks.

 

[Bob Ayers]

So, Bob, why do YOU care how many in-cache multiplications or additions some computer can do? What's the significance?

No, I've been out of the HPC field for several years. And as I said, it's silly and I didn't care who was the "best" even at the time.

HPC is not NASCAR. It's not penis comparison. The import of a computer is what it can calculate, and that is NOT measured in flops. There isn't any perfect measure, but a good one is literature citations.

Why don't you give it up!!! You act like the rest of us don't know what a cache is, and all your doing is throwing around buzz words you have heard!!

. A traditional ethernet-linked weakly bound cluster would make more sense for that.

.

WOW, you are really showing how behind the times you are!!! For speed, you don't want to use copper, period!!!


I'm done arguing with you!!