October 27, 2005 1:57 PM PDT

Blue Gene/L tops its own supercomputer record

Lawrence Livermore National Laboratory and IBM unveiled the Blue Gene/L supercomputer Thursday and announced it's broken its own record again for the world's fastest supercomputer.

The 65,536-processor machine can sustain 280.6 trillion calculations per second, called 280.6 teraflops, IBM said Thursday. That's the top end of the range IBM forecast and more than twice the previous Blue Gene/L record of 136.8 teraflops, set when only half the machine was installed.

In addition, the lab unveiled a lesser known but also powerful machine with a speed up to 100 teraflops. The ASC Purple is built from more conventional IBM server products. Together, ASC Purple and Blue Gene/L cost $290 million. Both will be used for nuclear weapons simulations and other computationally demanding tasks.

News.context

What's new:
Lawrence Livermore Laboratory and IBM unveiled the Blue Gene/L supercomputer on Thursday and another extremely powerful machine, ASC Purple. Blue Gene/L broke its own record for calculation speed, sustaining 280.6 trillion calculations per second.

Bottom line:
These supercomputers are capable of handling a range of extremely demanding computational tasks, from data management for businesses to simulations of nuclear explosions and complex biochemical reactions.

More stories on this topic

Broadly speaking, both machines use a similar approach: A powerful network connects huge numbers of processors. But there are differences. ASC Purple's 12,544 Power5 processors are individually more powerful than the specialized Power chip variants used in Blue Gene/L. And, each ASC Purple processor has access to more memory so that more complicated simulations can run. However, Blue Gene/L consumes less power, and instead of using a massive, complex central switch to connect processors, it uses a collection of five separate networks.

Supercomputing was once a domain limited to companies, such as Cray, Hewlett-Packard and Silicon Graphics, with deep research programs. But the steadily increasing power of mainstream processors and networks has meant that even companies without major engineering programs, such as Dell, have entered the market.

Blue Gene/L first became the top-ranked supercomputer a year ago on the Top500 supercomputing list with a 70.7-teraflop speed. The speed is measured with a mathematical test called Linpack, but Top500 organizers acknowledge it has been an incomplete measurement of a computer's abilities.

When Blue Gene/L runs Linpack, each processor performs mathematical calculations. But for many jobs, half the system's processors are used for networking tasks.

Blue Gene/L and ASC Purple

ASC Purple, originally called ASCI Purple, got the first part of its name from the Department of Energy's Advanced Simulation and Computing program, which was designed to guarantee that U.S. nuclear weapons would continue to work even as they age well past their intended life spans. The "Purple" comes from a mixture of red, white and blue: ASC Purple was intended to be the culmination of the series of ASCI Red, White and Blue supercomputers built at Sandia, Lawrence Livermore and Los Alamos national laboratories.

ASC Purple had been due at the end of 2004, but the schedule was extended because of a move to a less expensive design.

Among the dignitaries attending the unveiling Thursday are Linton Brooks, administrator of the Energy Department's National Nuclear Security Administration; Ray Orbach, director of the DOE's Office of Science; Nicholas Donofrio, IBM's executive vice president for innovation and technology; and LLNL Director Michael Anastasio.

Blue Gene/L and ASC Purple together are expected to consume 10 megawatts of the 45-megawatt capacity that LLNL's Terascale Simulation Facility can supply for computing and cooling.

Protein research and beyond
The Blue Gene design began as a research project within IBM to simulate fundamental biochemical processes involving proteins. Lawrence Livermore National Laboratory became its first customer, but the machines now are used for several research projects around the world as IBM commercializes the technology.

IBM sells the Blue Gene machines for about $2 million per 1,024-processor rack, but it also rents access to Blue Gene and other supercomputers for those who don't want to buy a full machine. On Friday, IBM is expected to announce four new customers for the rental service, called Deep Computing Capacity on Demand. Two of the new customers will be using Blue Gene.

One customer is SmartOps, which helps companies such as Caterpillar and Cardinal Health to manage inventory and distribution. Another is QuantumBio, which plans to sell processing power to pharmaceutical companies that perform drug simulation and other computationally intense tasks.

Nuclear weapons remain an active area of computing research in the U.S. as the DOE tries to ensure their reliability without using actual nuclear tests. ASC Purple, still undergoing final acceptance tests, will be able to fully simulate a nuclear explosion in three dimensions.

But the labs won't be able to rely on another potential source of funding that DOE has sought for years: development of a controversial "bunker-buster" warhead called the Robust Nuclear Earth Penetrator, which can destroy deeply buried sites.

The Bush administration has sought funding for the project in recent years, but now has agreed to focus instead on conventional explosives that can do the job. The change of plans--and the disappearance of $4 million in funding for national laboratories that the Senate had earmarked for the current fiscal year--was announced Tuesday by Sen. Pete Domenici, a Republican from New Mexico.

But there is plenty of other supercomputing work. LANL's ASC Q Machine has been used for biological research into the creation of proteins with a cellular factor called a ribosome.

12 comments

Join the conversation!
Add your comment
Wrong processor count ...
I believe this article is in error if the system has reached it's
complete 64 compute rack size.

BlueGene/L has 1024 dual processor nodes per rack, or 2048
processors per rack. The article states this incorrectly as '1024
processors per rack'. The total processor count should be
131,072, if this is the full 64 rack system.

Please feel free to verify this with IBM.

Jason Lockhart
Assoc. Director,
Virginia Tech Terascale Computing Facility
Posted by multimediavt (5 comments )
Reply Link Flag
Wrong Processor Count
The processor count IS correct - it's the number of cores per die that is in question.

Each of the PowerPC 400 700 MHz processors are dual-core, with a shared 4MB L3 cache architecture, so it is 65,536 processors, and 131,072 cores.

The dual-core architecture also implies that single core PowerPC 400 processors are NOT available.
Posted by alpha754293 (2 comments )
Link Flag
Reporter responds: it's a definitional issue
Most server makers these days define a processor as that which plugs into a socket, not as a processor core. There are significant exceptions, and IBM is one of them, though it's not consistent.

IBM, which led the server world to the dual-core world with its Power4, says a dual-core chip has two processors. Sun, Hewlett-Packard, Intel, AMD, Microsoft, BEA Systems, VMware and others say a dual-core chip is one processor. News.com goes with the socket definition (though we, too, were in flux for some time). So if you were to buy a new PC with a dual-core Pentium or Athlon, for example, you wouldn't call it a dual-procesor system even though of course it behaves as such in many ways.

The issue gets particularly complicated with some more radical multicore approaches. For example, Sun's Niagara chip has eight cores, and each core can execute four threads, so from a software point of view it looks like 32 processors. Do you call Niagara a single processor, or eight, or 32? And how about Azul Systems' Vega, which has 24 cores but runs Java code rather than a conventional operating system--is it a single processor or 24?

Worse, many software companies charge per processor, but the definitions vary. IBM itself is conflicted; when describing the dual-core Power4 and Power5 chips, it calls each chip two processor--but it labels the dual-core Xeon and Opteron one processor (see <a class="jive-link-external" href="http://news.cbsi.com/IBM+shifts+software+price+for+dual-core+x86+chips/2100-1006_3-5679679.html" target="_newWindow">http://news.cbsi.com/IBM+shifts+software+price+for+dual-core+x86+chips/2100-1006_3-5679679.html</a>). Oracle counts each core on a multicore chip as three quarters of a processor for softwre licensing purposes, so software for a dual-core chip would cost 1.5 times that for a single-core chip.

Regarding Blue Gene, the computer has 65,536 nodes, each node with a single chip. Each chip has two PowerPC 440 processor cores. So if you're counting by chips, it's 65,536, by cores, twice that. IBM tends to use the "processor = core" definition in its literature, which accounts for the discrepancy with the story.
Posted by Shankland (1858 comments )
Link Flag
Wrong processor count ...
I believe this article is in error if the system has reached it's
complete 64 compute rack size.

BlueGene/L has 1024 dual processor nodes per rack, or 2048
processors per rack. The article states this incorrectly as '1024
processors per rack'. The total processor count should be
131,072, if this is the full 64 rack system.

Please feel free to verify this with IBM.

Jason Lockhart
Assoc. Director,
Virginia Tech Terascale Computing Facility
Posted by multimediavt (5 comments )
Reply Link Flag
Wrong Processor Count
The processor count IS correct - it's the number of cores per die that is in question.

Each of the PowerPC 400 700 MHz processors are dual-core, with a shared 4MB L3 cache architecture, so it is 65,536 processors, and 131,072 cores.

The dual-core architecture also implies that single core PowerPC 400 processors are NOT available.
Posted by alpha754293 (2 comments )
Link Flag
Reporter responds: it's a definitional issue
Most server makers these days define a processor as that which plugs into a socket, not as a processor core. There are significant exceptions, and IBM is one of them, though it's not consistent.

IBM, which led the server world to the dual-core world with its Power4, says a dual-core chip has two processors. Sun, Hewlett-Packard, Intel, AMD, Microsoft, BEA Systems, VMware and others say a dual-core chip is one processor. News.com goes with the socket definition (though we, too, were in flux for some time). So if you were to buy a new PC with a dual-core Pentium or Athlon, for example, you wouldn't call it a dual-procesor system even though of course it behaves as such in many ways.

The issue gets particularly complicated with some more radical multicore approaches. For example, Sun's Niagara chip has eight cores, and each core can execute four threads, so from a software point of view it looks like 32 processors. Do you call Niagara a single processor, or eight, or 32? And how about Azul Systems' Vega, which has 24 cores but runs Java code rather than a conventional operating system--is it a single processor or 24?

Worse, many software companies charge per processor, but the definitions vary. IBM itself is conflicted; when describing the dual-core Power4 and Power5 chips, it calls each chip two processor--but it labels the dual-core Xeon and Opteron one processor (see <a class="jive-link-external" href="http://news.cbsi.com/IBM+shifts+software+price+for+dual-core+x86+chips/2100-1006_3-5679679.html" target="_newWindow">http://news.cbsi.com/IBM+shifts+software+price+for+dual-core+x86+chips/2100-1006_3-5679679.html</a>). Oracle counts each core on a multicore chip as three quarters of a processor for softwre licensing purposes, so software for a dual-core chip would cost 1.5 times that for a single-core chip.

Regarding Blue Gene, the computer has 65,536 nodes, each node with a single chip. Each chip has two PowerPC 440 processor cores. So if you're counting by chips, it's 65,536, by cores, twice that. IBM tends to use the "processor = core" definition in its literature, which accounts for the discrepancy with the story.
Posted by Shankland (1858 comments )
Link Flag
my abacus has 128K beads ...
go figure :)
Posted by Lolo Gecko (131 comments )
Reply Link Flag
my abacus has 128K beads ...
go figure :)
Posted by Lolo Gecko (131 comments )
Reply Link Flag
Blue Gene/L ASIC
Why don't they upgrade the L1 D-cache &#38; L1 I-cache of the 440 core ?

or "Can they upgrade the... ?"
Posted by 56455734 (2 comments )
Reply Link Flag
Blue Gene/L ASIC
Why don't they upgrade the L1 D-cache &#38; L1 I-cache of the 440 core ?

or "Can they upgrade the... ?"
Posted by 56455734 (2 comments )
Reply Link Flag
 

Join the conversation

Add your comment

The posting of advertisements, profanity, or personal attacks is prohibited. Click here to review our Terms of Use.

What's Hot

Discussions

Shared

RSS Feeds

Add headlines from CNET News to your homepage or feedreader.