Sonny,
Yes Intel's into alot more things than I think we realized.
Check this article out. Sorry I couldn't get the link right. So I posted what I could.
computerworld.com then go down to In Depth (if you want the pictures)
In Depth - Nuclear Aging
The nation's atom bombs are long in the tooth. A crash supercomputing project intended to safeguard them is yielding some down-to-earth benefits.
By Gary H. Anthes
(In Depth, March 23, 1998) They used to make nuclear weapons at these New Mexico labs. Now they're bombing Moore's Law.
Computer scientists at Sandia National Laboratories and Los Alamos National Laboratory have mounted an all-out assault on the frontiers of computing. The goal is to develop by 2004 supercomputers that can perform 100 trillion computations per second - 100 times faster than today's speediest machines.
Along the way, those scientists are developing software that is fundamentally changing the way science and engineering are done. And they are inventing techniques for software development, modeling, data mining, data storage, networking and other disciplines that are finding nontechnical, commercial applications.
The effort is called the Accelerated Strategic Computing Initiative (ASCI) and also involves Lawrence Livermore National Laboratory in Livermore, Calif. ASCI was created to replace nuclear weapons testing, which the U.S. ended in 1992, with computer simulation. The initiative's aim is to ensure the safety and reliability of the nation's aging atomic arsenal as part of the U.S. Department of Energy's Stockpile Stewardship and Management program.
Propelled by an annual budget that next year will be $517 million, ASCI is infused with a sense of urgency. By 2004, half the nuclear weapons designers in the U.S. will have left their jobs, and some weapons will be beyond their intended life spans, which generally don't exceed 12 years. The U.S. is the only nuclear power that isn't building new weapons, so scientists must find ways to extend the life of existing weapons and must greatly improve their ability to predict the performance and potential failures of those aging weapons.
Advanced Computing Lab's Andrew White.
Andrew White, director of the Advanced Computing Laboratory at Los Alamos, says one weapon simulation program there takes 500 hours to run on a supercomputer. The model would be satisfactory if its results could be verified by occasionally exploding a bomb underneath the Nevada desert. But without that validation, he says, the program will require such extensive enhancements that it will need 10,000 times more computer power, or close to 200 trillion floating-point operations per second (TFLOPS). "This is deadly serious - a matter of national security," White says. "We need to accelerate what's going on."
More than Moore
For the labs to meet their schedule, they must triple the real performance of their computers every 18 months - 50% faster than is suggested by Moore's Law, which says the power of computer chips at a constant price doubles every 18 months. The acceleration is possible only by applying more processors to a problem, speeding interprocessor communications and using more efficient software, says Charles Slocomb, director of computing, information and communications at Los Alamos National Lab.
The labs' fastest supercomputers today are good for a bit more than 1 TFLOPS. The plan is to scale up to 10 TFLOPS by 2000, to 30 TFLOPS by 2001 and to 100 TFLOPS by 2004. The existing ASCI supercomputers are supplied by Intel Corp. at Sandia; Silicon Graphics, Inc./Cray Research at Los Alamos; and IBM at Lawrence Livermore.
To keep costs down and ensure continued availability, ASCI is buying mainstream, commercially available hardware and software when possible and is sponsoring development of items that are likely to find broad use in industry. "The paradigm is to develop with industry, then transfer to industry and buy back from industry," Slocomb says.
Progress on the application front is aided by collaboration with computer users in industry. The labs use a model that offers U.S. corporations and government agencies expertise in advanced computing disciplines, and the companies and agencies contribute funds and interesting problems to work on.
THE RUBBER MEETS THE ROAD:
A Sandia National Labs simulation of an auto tire's contact patch. Goodyear's Loren Miller calls the company's cooperative work with Sandia a "breakthrough."
Indeed, one can learn about weapons by studying tires, say scientists at Sandia in Albuquerque. Goodyear Tire & Rubber Co. in Akron, Ohio, has contributed $46 million to Sandia in a technology sharing partnership that developed advanced software for modeling new tire designs. Sandia has learned how to better model the performance of polymers used in weapons, and Goodyear has learned how to make better tires, the parties say.
Loren Miller, director of tire performance modeling at Goodyear, says the high-performance modeling software couldn't have been obtained at any price from commercial sources. "It's been a significant breakthrough for us," he says. "We are talking about reducing [computer run] time from years to days."
And the Stockpile Stewardship program's dictate - that the performance of objects must be reliably predicted without physically testing them - is revolutionizing the way Goodyear develops tires, Miller says. The company previously built tire prototypes, tested them on a track, modified the design and repeated the process - often many times. Now computer models allow designers to bypass much of that physical prototyping and testing. Miller won't say how much time or cost has been cut from the tire development cycle. But he says major components of that effort - such as minimizing the rolling resistance of a tire - have been slashed by "substantially more than half."
National labs and companies are learning how to make computer models more realistic and more predictive by moving to "full physics" simulations. Those detailed models use enormous amounts of data and would have been impractical to run before the development of superefficient algorithms and faster hardware.
"We are accelerating the development of a discipline called simulation," says ASCI program head Gil Weigand, the Energy Department's deputy assistant secretary for strategic computing and simulation. "We believe this discipline will ultimately prevail in science on a par with experiment and theory."
Work with the power
The labs are pursuing several supercomputer architectures, but they all involve harnessing the power of many processors in parallel. The 100 TFLOPS behemoth is likely to have 10,000 processors, draw 10 to 20 megawatts of power and take up most of a new $125 million building, Slocomb says.
For years, the real performance of computers with many processors has lagged behind their theoretical performance. That's because they can be hard to program - application code has to be laboriously "parallelized" - and because of "latency," the nonproductive time it takes instructions and data to move between processors and memory.
The labs are working hard on both problems as well as other software challenges. "It's the software, stupid - that's our mantra," says John Reynders, a senior project leader at Los Alamos, where 400 people are dedicated to computer science research.
BRIDGING THE GAP:
A worker at Sandia National Laboratories inspects cables on an Intel teraflops supercomputer's "disconnect cabinet."
Los Alamos is adding parallel constructs to the Common Object Request Broker Architecture - a standard for exchanging software objects.
It also has invented a framework for developing parallel applications called Parallel Object-Oriented Methods and Applications (POOMA).
POOMA is a C++ class library that offers a high-level programming interface and application-specific objects.
Hidden underneath is hardware-independent software that structures code for parallel computing by taking care of load balancing, message passing and multithreading.
Less time, less effort
"The key to POOMA is it uses objects that are 'parallel aware,' " Reynders says. Programmers who have no parallel computing knowledge can write serial code on a Linux-based laptop and easily port the application to the Intel, SGI/Cray or IBM parallel supercomputers, he says. A three-dimensional simulation program, written using POOMA in just six weeks, would have required four to five months without POOMA, Reynders says.
POOMA was developed to handle large-scale scientific problems but its concepts are suited for other complex applications such as those used on Wall Street to evaluate and price investments, Reynders says. One could just substitute different object classes on top of POOMA's lower layers, he says.
Center for Adaptive Systems Applications, Inc. (CASA), a company in Los Alamos started by refugees from the weapons lab, is applying techniques developed at the lab to mine multiterabyte databases.
The company has done fraud detection and consumer buying habit analysis for Citibank and other companies that use algorithms adapted from lab work in chemicals and astrophysics, says John Davies, president of CASA.
Indeed, the weapons-oriented research has surprising applicability in the non-defense world. The following applications were completed recently or are under development:
For the U.S. Health Care Financing Administration, Los Alamos developed a system to spot fraud in 2T bytes of claims data. Based on neural networks and other techniques, it can find new patterns of fraud not predefined by users.
Working with several government agencies, including the Los Angeles County Fire Department, Los Alamos developed a model to predict the spread of wildfires and to evaluate firefighting actions during a blaze.
Sandia developed VRaptor, a virtual reality-based simulation tool for training security and law enforcement officers on how to respond to terrorist attacks.
With other agencies, Los Alamos is developing climate prediction models to help understand the costs and benefits of complying with the recent Kyoto, Japan, carbon emissions accords.
Los Alamos has joined a "pandemic planning task force" and is working on a model to predict the spread of disease and to evaluate prevention strategies.
A nuclear weapon can have as many as 6,000 components. Using computer simulation to ensure the reliability and safety of these devices isn't a slam dunk, lab officials say. In fact, it may not work - period.
The task is extraordinarily hard because unlike more conventional simulations that take small steps into the unknown based on relatively small amounts of data, the stewardship program must be able to predict the future of immensely complex systems.
And the simulations no longer can be completely verified with experiments, yet they must be made ultratrustworthy.
"The year 2004 is a key date in the Stockpile Stewardship program," White says. "It's a watershed for whether you really can use modeling and simulation in a predictive sense."
--------------------------------------------------------------------------------
Anthes is Computerworld's editor at large. His Internet address is gary_anthes@cw.com.
Andrew White photo: Chip Simons
Sandia Labs photos: Randy Montoya
--------------------------------------------------------------------------------
Copyright c 1998 @Computerworld. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of @Computerworld is prohibited. Computerworld and @Computerworld and the respective logos are trademarks of Computerworld, Inc. |
