Cray’s Comeback: CEO Peter Ungaro on Clouds, Exaflops, and the Future of Supercomputing
Where I grew up in the Midwest in the 1970s and early 80s, Cray was synonymous with supercomputing. Back then, a supercomputer was a top-flight machine that could perform a few hundred million floating point operations per second (“flops”). That was good enough to help scientists do intensive calculations in areas like weather forecasting, climate modeling, and nuclear weapons simulations. Cray’s first supercomputer, the famed Cray-1, was bought by Los Alamos National Laboratory for $8.8 million in 1976; eventually, some 80 of the machines were sold, for $5 million to $8 million a pop.
Today, your average desktop computer is far more powerful than a Cray-1, and so the definition of “supercomputer” keeps changing to keep up with the times. But one thing has not changed. Cray (NASDAQ: CRAY) is still a major player in the space, despite a long history of ups and downs. The company, which began in 1972 as Cray Research in Chippewa Falls, WI, was bought by Silicon Graphics in 1996 for $767 million, and then was reborn in Seattle in 2000 following a $50 million merger with Tera Computer (which was renamed Cray). Since then, it has been a long uphill climb to get back near the top of the supercomputing heap against heavyweight competitors like IBM and Hewlett-Packard.
Nobody better to tell that story than Peter Ungaro, the chief executive of Cray. I recently had a chance to speak with Ungaro by phone at his Spokane, WA, office about his company’s strategy and recent history, the technical challenges involved in modern supercomputing, and innovative ways of gaining new customers (how do you sell someone a $10 million machine?). What impressed me was his ability to lay out the financial concerns of his company while also diving deep into the technological aspects of supercomputers—how they will interact with cloud computing, how computational records will continue to be broken, and when computers might exceed all processing capabilities of the human brain.
First off, I wanted to know how Ungaro (left) defines a “supercomputer” these days. Some would say it should be one of the 500 fastest machines in the world. Others would say it’s a machine used for scientific and technical problems that costs more than a certain amount. Ungaro’s definition is simple and focuses on the bottom line. “We like to think of supercomputers as costing more than a million dollars,” he says.
Ungaro, a Washington State University alum, joined Cray in 2003 to run sales and marketing as senior vice president. He had been at IBM for 13 years, most recently running its high performance computing group, a $2 billion business inside Big Blue. Why did he make the jump to Cray? “I really loved the supercomputing space,” Ungaro says. “Customers are doing really interesting things. I really wanted to try and see what a smaller company was like. Even at $2 billion, you’re only 2 percent of IBM’s revenues.” In short, like many entrepreneurs, he wanted to have more impact. “There was no better place to go than Cray. It was a natural move.”
But Cray had its share of problems. The company had struggled to get its next-generation supercomputer product ready, and 2004 was “really rough,” Ungaro says. Cray was losing money and going through a difficult restructuring. In the middle of 2005, Cray’s CEO (and Tera co-founder) Jim Rottsolk retired, and the board asked Ungaro to take on the role of president and then CEO.
Ungaro and his team made some bold moves in an effort to turn the company around. Cray went from just under 1,000 employees to about 750, while retaining most of its engineers. “We resized the company, except for R&D, so we could get more viable,” he says. “At that time, we really started to see a huge shift in the market toward multi-core technology. We saw it as a big thing, and a big place where we could differentiate. Instead of fighting the commodity [microprocessor] wave, we could use it to our advantage. That was a big bet we placed in 2005. Over half the company is R&D, incredibly brilliant engineers, and we wanted to use our traditional strengths.”
That meant focusing on using parallel-processing chips—those with multiple central processing units. Thanks to Moore’s Law, these chips have become quite powerful components for boosting the performance of supercomputers. To support this effort, Ungaro says, Cray invested in custom hardware and software to help pool the processing punch of thousands of machines, software to get as much performance as possible out of each machine, and packaging technology (including advanced liquid cooling systems), all in order to squeeze “more computing power into less space.”
It has taken a few years, but the strategy seems to be paying off. For starters, Ungaro says Cray has become debt-free as of this spring, after paying off $80 million in debt since October. “As the market went down, we started paying back the debt because we could get a great discount,” he says. “It’s a great sign for the company to pay off our debt early. It’s like paying off your home mortgage. It’s part of the growth of our company, becoming more financially stable.”
While paring back debt, Cray has been busy boosting its revenue—a whopping 52 percent increase in 2008 compared with the previous year. And the company’s revenue for the first quarter of 2009 was $74.5 million, compared with $26.1 million in the same period a year earlier. Yet Cray still hasn’t quite pulled itself out of the red, posting a net loss of $4.9 million for the first quarter of this year. So the firm still has a ways to go to get back on top against competitors like IBM (maker of the Roadrunner machine at Los Alamos), Hewlett-Packard (maker of Chinook at Pacific Northwest National Laboratory), Silicon Graphics, Fujitsu, Hitachi, and NEC. (Cray’s second quarter numbers for 2009 will be announced on Tuesday, Aug. 4.)
Nevertheless, the company has built itself back up to about 850 employees, 160 in the Seattle office, which employs a mix of engineers, finance, legal, and marketing people. (Ungaro adds that the Seattle office is hiring.) Cray’s biggest office—which includes employees in development, service, sales, marketing, and administration—is in Minneapolis-St. Paul, MN, and the firm still does all of its manufacturing in Chippewa Falls, WI.
Three business units are at the heart of Cray’s turnaround strategy. The first is “scalable systems,” its cash cow of high-end machines that cost $1 million and up. The second is what Ungaro calls “custom engineering.” This means taking individual supercomputing technologies—things like cooling systems, cabinets, and pieces of hardware or software—and selling them as professional services instead of delivering a complete supercomputer. And the third business unit is what he calls “productivity solutions”—a rather bland name for what amounts to a more mainstream kind of Cray machine, which sounds like an exciting growth area.
Essentially, this unit of Cray concentrates on partnering with companies like Microsoft and Intel to broaden Cray’s market reach. Case in point: teaming up with Microsoft last September to sell the CX1, a $25,000 high-performance desktop that runs on a Windows operating system and uses Intel processors. Suddenly, this takes a portion of Cray out of the realm of U.S. government labs, and closer to mid-sized, data-intensive businesses that need some, but not all, of the features of a traditional supercomputer.
“It’s not a core piece,” Ungaro says, meaning it’s not generating a majority of sales, but the partnerships to offer high-powered, less-expensive computers are exciting because they offer Cray potential to grow with a new type of customer that doesn’t buy conventional supercomputers. “It’s a fun area for us. It’s a lot to grow our brand, and reach out to more people.” He says the CX1 started selling in volume at the beginning of 2009, and Cray has been building a distribution channel of 15 resellers around the world, with plans to expand to 25 or 30 by the end of the year. “It could become a reasonable part of a business, but a smaller piece,” he says. “We don’t want to become a Dell or HP.”
Reflecting a bit more on the “low-end” CX1 (left), Ungaro says, “Most people would never think, ‘Microsoft plus Cray.’ But we both had a similar vision for what’s important in the market. Not just performance, but also productivity—making these things more accessible and usable by people. We’ve been able to work with them to make CX1 an easy system to buy, and easy to install.” (Though he admits he doesn’t have one in his home just yet.)
Since the conversation veered toward mainstream computing, I asked Ungaro about where he thinks supercomputing fits with cloud computing, the phenomenon in which more companies are renting server space to store and process data on computers operated by vendors like Amazon, instead of buying and maintaining their own servers. “We see it as very complementary technology to what we’re doing. People could access Cray supercomputing,” he says. “Companies that can leverage supercomputing to develop their next consumer product—typically they’re not going to need to buy a supercomputer, so the cloud could be a way for people to access that.”
For consumer examples, he cites auto manufacturers running car-crash simulations, aeronautical companies like Boeing modeling air flow over new wing designs, and even golf club designers making new equipment based on advanced supercomputing simulations. One company even announced in Golf Digest that it had designed its latest driver on a Cray supercomputer. “It’s not helping my golf game at all,” Ungaro jokes.
As for the future of supercomputing, it all comes down to the flops—the number of floating point operations per second that a machine can perform. “We just broke the petaflop [1 million-billion flops] barrier. We’re now thinking about what’s next: the exaflop barrier [1 billion-billion flops]. Everybody’s hoping to break it before 2020,” Ungaro says. “But it’s going to take amazing technology, to get much lower-power technology and incredible scale. A year ago, the total number of processors on the Top 500 list was 1 million. An exaflop will have more than a million processors in one machine. That’s what our guys live for, a challenge like that.”
Could this conceivably lead to machines that exceed the capabilities of the human brain? “People say somewhere between 10 to 100 petaflops, you’re approaching the human brain. We’re still at 1.65 petaflops. The next step is 2. Our current technologies can get us to the 10-20 petaflop range. But then to start to think about 100 [petaflops], we really need a major shift in technology.” Such as? Ungaro says it will still have a silicon-based processor—not quantum computing or anything that far out—but somewhere around 2015, he sees all connections and communications within a supercomputer being done with optics instead of copper cables, which could speed up processing dramatically. It’s a hot research area called silicon photonics. “Optical is just starting to come into play,” he says. “We need to get it integrated in motherboards and close to processors and memory.”
Indeed, supercomputers are the stuff of dreams and futures. But closer to the present, Ungaro remains focused on Cray’s three main business units, especially its newer initiatives in custom engineering and mainstream machines. “It’s been a big engine. We have goals to grow the company,” he says. “We’re on a really exciting path.”