Computing experts at Sandia have launched an effort to help discover what computers of the future might look like, from next-generation supercomputers to systems that learn on their own — new machines that do more while using less energy.
“We think that by combining capabilities in microelectronics and computer architecture, Sandia can help initiate the jump to the next technology curve sooner and with less risk,” says Div. 1400 Director Rob Leland. He has outlined a major Research Challenge into next-generation computing called Beyond Moore Computing, part of Sandia’s overall work on future computing.
For decades, the computer industry operated under Moore’s Law, named for Intel Corp. co-founder Gordon Moore, who in 1965 postulated it was economically feasible to improve the density, speed, and power of integrated circuits exponentially over time. But speed has plateaued, the energy required to run systems is rising sharply, and industry can’t indefinitely cram more transistors onto chips.
The plateauing of Moore’s Law is driving up energy costs for modern scientific computers to the point that, if current trends hold, more powerful future supercomputers would become impractical due to enormous energy consumption.
Solving that conundrum will require new computer architecture that reduces energy costs, which are principally associated with moving data, Rob says. Eventually, computing also will need new technology that uses less energy at the transistor device-level, he adds.
Sandia experts expect multiple computing device-level technologies in the future rather than one dominant architecture. About a dozen possible next-generation candidates exist, including tunnel FETs (field effect transistors, in which the output current is controlled by a variable electric field), carbon nanotubes, superconductors, and fundamentally new approaches such as quantum computing and brain-inspired computing.
Sandia’s facilities will play key role in researching future computing technology
Sandia is well-positioned to work on future computing technology due to its broad and long history in supercomputers, from architecture to algorithms to applications. Rob says Sandia can play a key role because of that background and two key facilities: the Microsystems and Engineering Sciences Applications (MESA) complex, which performs multidisciplinary microsystems research and development and fabricates chips to test ideas; and the Center for Integrated Nano-technology (CINT), a DOE Office of Science national user facility operated by Sandia and Los Alamos national laboratories.
No one is sure what tomorrow’s high performance computers will look like. “We have some ideas, of course, and we have different camps of opinion about what it might look like, but we’re really right in the midst of figuring that out,” Rob says.
Erik DeBenedictis (1425) says Sandia can play an important role in creating breakthroughs that are not simply variations of transistors — developments such as computers that learn or technologies that move data from one part of the computer to another more efficiently. Those are crucial for big data problems.
What ultimately prevails might well be something not yet invented, Rob says.
“That’s the first challenge, to figure out what the new device technology is, then work through what the implications of that are, what sort of computer architecture is required to assemble that device into components and subsystems and systems,” he says.
New technology must be broadly adopted to drive improvements
Sandia needs both capability computing, which means finer resolution and more accuracy, and capacity computing, or running many different jobs simultaneously.
“So what does efficiency buy you? It allows you to have a bigger computer or more computers with the same amount of operating expense — paying your power bill,” says manager John Aidun (1425). “There’s no limit to the amount of efficiency we would like to achieve because really there’s no limit to the amount of computing we would like to do.”
Whatever technology comes next must be broadly adopted so it will drive continual improvements, similar to the way the 1947 invention of the transistor transformed society. It’s not enough to have a device that’s fast; it has to be something that can be built into a complete computer system, John says.
Thus, new technology must have commercial uses. “There will have to be some industrial base that supports it and produces it and that can be used to assemble a large number of these into a system that can be deployed for national security,” Rob says. “What we’d really like to do is figure out how to advance the state of the art for national security in a way that is more broadly deployable across society.”
The computer industry is exploring technologies that in essence are drop-in replacements for transistors with improved characteristics: different designs such as the fin FET, a 3-D rather than a flat configuration on a computer chip, John says. While the design would be moderately disruptive for industry, it’s still compatible with standard silicon fab technology and opens the potential for generations of ever-smaller FinFETs on a chip, he says.
While industry views a beyond-transistor technology as something far off, Sandia’s national security interests anticipate bigger changes will be needed sooner than industry would develop them on its own, John says. He estimates Sandia could have a prototype new technology within a decade.
Identifying best computer designs can help accelerate innovation
To accelerate the process, Sandia wants to identify computer designs that could take advantage of new device technologies and demonstrate key components or steps in fabrication that would lower the risk for industry by demonstrating technological feasibility.
“We’d be doing it with an eye toward helping industry give due attention to national security needs in computing,” John says.
The numerical capability developed in computers in World War II remains valuable today for such tasks as nuclear weapons simulations. But the modern era’s largest computing development, the Internet, deals with text and demands computing functions called integer calculation, also used in mobile computing.
Improving mobile computing could allow much more efficient and rapid data processing aboard satellites, so less data would need to be sent to Earth for processing.
“The mobility we see in cell phones and tablets is the closest match for the mobility needs of UAVs and satellites,” Erik says. “The energy and time required to transmit data to the ground, process it there, and send the answer back is a bottleneck, and it can be more resource-intensive than just computing on the device.”
He also suggested turning more programming over to cognitive computers to help programmers manage ever-faster computers. “While computers have gotten millions of times faster, programmers and analysts are pretty much as efficient as they’ve always been,” he says.
Cognitive computing can play role in pattern recognition
Cognitive computers might be able to do more to recognize patterns in satellite imagery, for example. People would still make the judgments, but computers would help by recognizing some lower-level patterns, he says. Up to now, programmers have created ways for computers to recognize images; computers didn’t learn on their own. A cognitive computer, however, would learn to identify patterns, Erik says.
“A computer can learn to recognize images pretty well. Humans assisted by a computer recognizing images could improve the ability significantly,” he says.
Researchers also must determine what hardware and software changes are needed so new devices are possible to manufacture and practical to operate. “You have to design over all those different considerations,” Rob says. “That’s what makes this a particularly challenging problem.”
Today’s computer systems rely on huge, longstanding investments in massive amounts of software.
“So we are strongly motivated to develop computers that will run old software that was optimized for traditional computer architectures that are not used today,” Erik says. “To break out of that, we have to find different architectures that are more energy-efficient at running old code and are more easily programmed for new code, or architectures that can learn some behaviors that once required programming.”
Since the software of today won’t unleash the full capabilities of the hardware of tomorrow, he expects computers in about a decade that can run both today’s software and new software. New software “would learn or would process information in fundamentally different ways, and become the most powerful aspect of the computer over time,” he says.