Cover Story: Inside Intel

  • As the research partnership between Carnegie Mellon and the chipmaker turns 10, it provides useful lessons about collaboration between companies and academics
By Jason Togyer
 
The industrialists of the Gilded Age preferred to do their experimentation on the factory floor and didn't have much time for scientific research--until Andrew and Richard Mellon came along. The millionaire financiers from Pittsburgh had struck it rich by placing big bets on the emerging technological marvels of the late 19th century, such as aluminum smelters and synthetic chemicals.
 
Inspired by a book called The Chemistry of Commerce, in 1913 the brothers founded Pittsburgh's Mellon Institute for Industrial Research, where major corporations funded research by the era's top scientists into plastics, electronics and nuclear physics. As corporations launched in-house laboratories in the 1950s, Mellon Institute turned to pure scientific research, then merged into Carnegie-Mellon University in 1967.

Yet about four blocks from the landmark Mellon Institute building erected in 1937, industry and academia are proving again that they can work together on scientific research--and do it very well. Instead of steel and rubber industry problems, they're solving the riddles of the Digital Age. The Collaborative Innovation Center, opened in 2005, houses both Carnegie Mellon researchers and laboratories for companies that respond to modern consumer needs, such as Apple. (Google was a tenant, too, until moving to larger quarters earlier this year.)

Of the CIC's tenants, perhaps none captures the spirit of Mellon Institute in its heyday quite like Intel's Pittsburgh research lab. The Intel lab, soon to begin its 10th year, works on problems in computer science both large and small. Not all of those experiments will pay off, but those that do will hit big, predicts Dave O'Hallaron, who in July completed a three-year stint as director of Intel Labs Pittsburgh.

"The lab's mission is to do exploratory research in computing and robotics," says O'Hallaron, a professor of computer science and electrical and computer engineering who is still working on several projects with Intel researchers. "Everything's on the table. It's not that different from the research being done elsewhere at Carnegie Mellon, and we don't have to reign ourselves in. In fact, if we're thinking too short-term, we're not doing our jobs."

Intel is now the world's largest semiconductor maker, dwarfing its nearest rival, Advanced Micro Devices. With its chips found in Linux-based, Windows-compatible and Macintosh personal computers, laptops and servers, Intel commands about 80 percent of the world's $37 billion market for microprocessors. But computer prices continue to fall as mobile and tablet devices command more attention from consumers--and as Wall Street analysts have recently pointed out, both of those factors have the potential to eat into Intel's future profit margins.

"Intel is a very interesting company, because they have to look a long way out in order to build and design their products, and they have to have some idea of what computers are going to be used for in 10 years," O'Hallaron says.

The corporation is betting that the next transformative consumer products will involve cloud computing as well as personal robotic devices, computer vision and perception--as a result, the Quality of Life Technology Center, a federally funded research institute run jointly by Pitt and CMU, is a major partner of Intel Labs Pittsburgh. All of the products that could emerge--robots that assist the disabled or telepresence devices that allow people to attend meetings remotely--will require microprocessors and represent new markets for Intel beyond desktop and laptop computers and servers, O'Hallaron says. "Intel is making products that have to work for everything," he says. "It doesn't have any particular axe to grind, so it's a perfect job for a researcher."

The Pittsburgh lab is a fairly small part of Intel's total international research effort, which encompasses everything from semiconductor design and manufacturing to computer architecture and software engineering. Worldwide, the corporation spent more than $5.7 billion in research and development last year, and it employs some 20,000 people doing research at labs and manufacturing centers in North America, Europe and Asia. By comparison, fewer than two-dozen people work at Intel Labs Pittsburgh--indeed, because of its relatively small size, it's sometimes called a "lablet."

Yet what Intel Labs Pittsburgh lacks in size, it makes up for in stature, thanks to its close partnership with Carnegie Mellon University. The Pittsburgh facility is one of three university-based research labs established in 2001 by David Tennenhouse, then Intel Vice President and Director of Research--the others are near the University of Washington and the University of California at Berkeley.

Intel Labs Pittsburgh, however, is the only one located on the partner university's campus, and that makes a huge difference, says Priya Narasimhan, associate professor of Electrical and Computer Engineering. "It really, really helps," says Narasimhan, who just took a leave of absence to begin a three-year stint as the newest director of Intel Labs Pittsburgh. "If you have a problem and you need to work with an Intel researcher, you grab a whiteboard, you grab a researcher, and you say, 'Let's talk.' The proximity makes the partnership organic." About 75 faculty members from the School of Computer Science and the Carnegie Institute of Technology--CMU's engineering school--are currently engaged in research with Intel Labs Pittsburgh, and more than 50 CMU graduate students last year worked on joint projects with Intel.

"That's an outrageously wonderful number," Narasimhan says. "Find me another industry research lab that can boast of those kinds of numbers. That cross-pollination tells you not just about the spirit of collaboration in this lab, but the spirit of collaboration at Carnegie Mellon."

At times, she says, Intel Labs Pittsburgh feels like a start-up company, and she would know. Narasimhan founded three technology companies, including YinzCam Inc., a Pittsburgh-based firm that allows fans sitting in the stands at sporting events to view their own personal instant replays on smart phones and other devices. Its latest clients include the Pittsburgh Steelers and the city's brand-new Consol Energy Center, where the Pittsburgh Penguins began playing hockey this fall.

Start-up companies succeed when they spot an emerging technology and take a chance on it, outmaneuvering established companies, Narasimhan says. "No start-up is risk averse," she says. "Imagine the kind of things you can do with a scrappy little start-up. Well, how much more could you do with Intel's resources behind you?"

The start-up feeling was very much in evidence when Intel Labs Pittsburgh launched in 2001, says Mahadev Satyanarayanan (CS'79,'83), Carnegie Group Professor of Computer Science at SCS and the lab's founding director. "From day one, knowing how proximity was essential to serendipitous interaction, I insisted we find the closest possible location for the lab," Satya says. The nearest office space turned out to be the second floor of a convenience store on South Craig Street.

"To be honest, it was a dump," Satya says. "We had a new lab, run by a company that wasn't then known for its (open) research, and in addition I had to show them this place above a Uni-Mart. The early researchers who joined us took major leaps of faith."

One of those taking the leap was Phil Gibbons, who remembers when Intel Labs Pittsburgh didn't even have the room over the Uni-Mart--it had some borrowed space in a corner of Satya's lab in Wean Hall. Gibbons left a research position at Bell Laboratories to become one of the first Intel employees hired in Pittsburgh. He's still there, nine years later, as principal research scientist.

"We were feeling the corporate pinch at Bell Labs, and part of that was a desire to make us more short-term relevant to the company," says Gibbons, who worked at Bell Labs for 11 years under both AT&T and Lucent ownership. "When I first heard about this Pittsburgh lab starting up, it kind of rang a 'bell,' if you will, because it reminded me of what Bell Labs was like when it first started up--doing more academic-style research." During its first half-century, under the benevolent ownership of America's monopoly long-distance company, Bell Labs conducted far-reaching inquiries into fields that had little to do directly with telephone technology--including motion pictures, stereo broadcasting, electronic music and, most famously, the transistor and the UNIX operating system.

But after the breakup of the Bell System in 1984, Bell Labs' parent company, AT&T, found itself competing for customers and revenue with other service providers. At Bell Labs, researchers now had to focus on protecting profits, and that meant less exploratory science and more concern over protecting intellectual property. By the late 1990s, IP concerns were stifling attempts to collaborate with universities, Gibbons remembers. "Essentially, if you could convince a faculty member to sign over all of their interests in the intellectual property, you could work with them," he says.

No such red tape hampers collaboration between CMU and Intel, thanks to something called the Open Collaborative Research, or OCR, agreement. The OCR specifies that both parties jointly own any research conducted by Carnegie Mellon faculty with Intel. It was devised by Tennenhouse and Satya, who remembered the headaches CMU encountered while trying to commercialize the Andrew File System it developed jointly with IBM in the 1980s. Although Carnegie Mellon wanted to license AFS to third-party vendors, IBM executives were afraid to give any competitors access to the technology, which hampered its acceptance, Satya says.

"I learned that giving intellectual property ownership to a company was totally deadly from our point of view," he says. When Tennenhouse approached CMU about a research partnership, Satya and others drew a line in the sand. "I told David that unless you're willing to come to some kind of an arrangement (about sharing IP), it's not worth continuing the discussion," Satya says.

To his surprise, Tennenhouse told him Intel executives were having the same kinds of conversations. And the corporation wasn't interested in having its university labs work on product development, which eased fears about letting corporate secrets slip. "David said, 'We're interested in door-opening research,'" Satya says. "He was specifically brought in because Intel wanted to take those kinds of risks."

Although the other university labs also use the OCR agreement, the CMU partnership is "absolutely special," says Limor Fix, a 15-year veteran Intel employee who has been in Pittsburgh for five years. She's now served as associate director with Narasimhan, O'Hallaron, and Todd Mowry, professor of computer science, who directed the Pittsburgh lab from 2004-2007.

"This is a dramatically different model of collaboration, and moreover, 10 years later both sides are very happy with it--which is very unusual!" Fix says. "What we do here is create new things--possible future directions for Intel." Working on existing product lines "is not what we do," she says. "We're looking at new markets, investigating new opportunities."

Narasimhan is the first director of Intel Labs Pittsburgh whose primary CMU affiliation wasn't in SCS. She says the lablet's researchers are supposed to be looking at cutting edge problems that are "five, 10, 15 years out." Among those looking into the future is Michael Kaminsky, senior research scientist at Intel Labs Pittsburgh, who joined the team in 2004 after earning a doctorate in computer science at Massachusetts Institute of Technology. Among his current research interests is the "fast array of wimpy nodes" being tested by David Andersen, assistant professor of computer science (see "Technology for a Greener Earth," The Link, Spring 2010). "Wimpy nodes" are less-powerful processors running in parallel that can perform the same tasks as newer rack-mounted servers, but use about one-sixth of the electricity.

One such inquiry--being conducted with Andersen and Intel's Dina Papagiannaki, Michael Kozuch and Madeleine Glick, along with a grad student from Rice University--involves direct optical switching in data center networks. Though data can be transmitted at high speeds and at great volumes over fiber-optic networks, when it goes from one group of machines to another, it eventually has to be converted back to electrical signals in a copper wire and move more slowly through a conventional switch. Traditional data-center networks use a "tree" of switches, with racks of computers living along the "leaves," but each switch is a bottleneck, and in cloud-computing applications, the inefficiencies quickly add up, Kaminsky says.

In an ideal world, high volumes of data would bypass this "tree" of switches, moving over faster, direct optical links instead. Kaminsky and other researchers are working on the algorithms that would be necessary to route the traffic. "A lot of the efficiency gains depend on what the traffic looks like," he says. "If you have a very uniform traffic pattern between racks, it's harder to gain efficiency than if you have traffic skew, with higher bandwidth demand between some pairs of racks."

Another of Kaminsky's projects, being conducted with co-principal investigator Dina Papagiannaki of Intel along with Andersen, Srini Seshan and Peter Steenkiste of SCS, is called Neighborhood Aware Networking. NaN imagines a day when local network traffic--say, data from one house to another in the same housing plan--is passed directly along from computer to computer using 802.11 technology, rather than being sent out of the neighborhood to the public Internet, and then back into the same exact neighborhood. Such an innovation could dramatically speed up data transfer while easing the burden on Internet Service Providers, Kaminsky says. Based on work with the NaN team, one Carnegie Mellon graduate student is currently writing a thesis on the use of directional antennas in indoor wireless networks, Kaminsky says.

In fact, more than one CMU student has generated a master's or doctoral thesis from work done at Intel Labs Pittsburgh. Gibbons says that over the past eight years, he's worked directly with 15 CMU students who--as a result of those collaborations--wound up as primary authors on papers submitted to conferences or peer-reviewed journals. If Intel benefits from access to Carnegie Mellon, the university is reaping benefits as well.

In some ways, Satya says, it's as if CMU has added an entirely new academic department at no cost to the university. "The caliber of the people in that lab are the same caliber of the people we would consider as faculty candidates," he says. "They have the ability to serve on thesis committees or mentor grad students. If you have a new idea you want to work on, you go down to the Intel Lab and have access to all of these high-quality researchers. It's as if the Computer Science Department was expanded by the size of the lab."

Kaminsky, Gibbons and several of their colleagues hold adjunct faculty appointments in SCS or CIT and mentor both undergraduate and graduate students, and say they see their role as being something like that of a professorship. "Carnegie Mellon faculty are very much involved with their students, but sometimes we as researchers have more time to devote to a particular topic," says Gibbons, who adds that working with students is perhaps the thing he likes best about being with Intel rather than Bell Labs. "Grad students bring tremendous energy, enthusiasm and eagerness to their work," he says.

Intel also hires students as "summer fellows"--not interns--to work in the lablet. Since its arrival on campus, Intel Labs Pittsburgh has each year supported the entire education of several CMU grad students, Satya says, and provided supplemental funding for others. "We're talking maybe $500,000 to $600,000 per year for 10 years," he says. "That's a tremendous benefit." Add to that CMU's access to Intel Lab Pittsburgh's Open Cirrus Cluster, which offers 1,300 processing cores and 610 terabytes of disk storage. The cluster, housed in the CIC, currently supports 20 different projects involving 80 researchers from Intel and CMU as well as other universities.

There are intangible benefits as well, Narasimhan says. Working closely with Intel gives CMU faculty insight into emerging issues facing industry, she says. "I get passionate about solving real problems in the real world," Narasimhan says. "In academia, we sometimes trick ourselves into working on problems only because they haven't yet been published. At Intel, we're not going to waste our time chasing fictional problems. It makes our academic research more relevant, and that, to me, makes sense. You're learning way more than you can with a 'toy' lab environment."

As a result, Narasimhan didn't think twice when she was asked to assume the directorship for three years. "When I had a chance to come to work here, I thought, 'Oh my goodness, there isn't a meeting I've gone to that I didn't say, I wish I had time to make that field my career,'" she says.

The rotating directorship was designed into the university partnerships from the very beginning. Fix calls it a "major component" in making Intel Labs Pittsburgh successful. "Whenever a new director comes in, he or she brings a flow of new ideas and additional connections with Carnegie Mellon," she says. "The energy level goes up. And it really has established a deep and trusting relationship between Intel and CMU. Any problem, any issue that comes to the table can be solved if you've worked together for three years and developed a deep trust."

There are lessons from the Intel-CMU partnership that should be studied by any university and corporation considering a similar effort, Narasimhan says. Don't erect legal barriers to sharing intellectual property, make sure that you're close to campus so that collaboration between the lab and the faculty is organic, and "tap into the power locked in those young student minds," she says.

As a result, she contends, Intel Labs Pittsburgh is poised to benefit the entire Pittsburgh region. "Our faculty are becoming known inside the highest ranks of Intel," Narasimhan says. "It gives CMU a bigger megaphone, both nationally and internationally. In everything we care about in computer science and electrical engineering, we couldn't ask for a better partner."
For More Information: 

Jason Togyer | 412-268-8721 | jt3y@cs.cmu.edu