SiCortex, a development-stage firm whose power- and space-saving Linux cluster architecture drew attention at SC06, recently named HPC industry veteran John Rollwagen as chairman of the board. Rollwagen, the former chairman and CEO of Cray Research, Inc., is now an investor and business advisor specializing in information technology. HPCwire asked Rollwagen what attracted him to SiCortex.
HPCwire: How did you link up with SiCortex?
Rollwagen: John Mucci, SiCortex's founding CEO, called me one day last summer. We'd met when he was at Thinking Machines. He said the board had been “trying to find someone to help lead the company like that John Rollwagen person who used to run Cray Research.” Finally, they decided to approach me directly. I visited the company, which turned out to be in the original Digital Equipment headquarters building in Maynard, Massachusetts. I had interviewed with Ken Olson in that building more than 40 years ago, before deciding to accept a job offer from another small computer company, Control Data. We talked over a period of time; and ultimately, I agreed to join the SiCortex board as chairman.
HPCwire: You were chairman of Cray Research from the time when it truly was a little start-up like SiCortex through the years when the company ruled the HPC market. Do you see any parallels with SiCortex?
Rollwagen: History doesn't repeat itself in exactly the same way because the context changes, but SiCortex does remind me of Cray Research when I first joined the company. Back then, when I read the first description of what the Cray-1 was going to be, I said to myself, “Of course. Computers should have been designed like this from the beginning. If you want to compute fast, keep the wires short and deal with the power and cooling issues.”
The SiCortex machines are designed from the ground up to overcome the limitations of the past 20 years of clusters. Standard microprocessors are physically large and produce a lot of heat. The SiCortex chips reduce the area by about 50 times, cut power consumption by two orders of magnitude, and even include the interconnect circuitry to increase overall performance dramatically.
HPCwire: Any other parallels to Cray Research as a start-up?
Rollwagen: Yes. SiCortex also has a very credible team. These are the guys who designed the Alpha microprocessor. They know what they're doing because they're done this before. I like these people. Like Cray Research when I first got there, SiCortex has the machine assembled but not quite operational yet. It's at the same stage. That also resonated with me.
HPCwire: What are the company's biggest priorities?
Rollwagen: The biggest one is to get the machine up and running. In the first quarter of 2007, we expect to have a pre-serial one machine. In Q2, we'll have the first fully functional machine. With that accomplished, we can get in front of customers.
HPCwire: What can you say about the SiCortex machines?
Rollwagen: They come in two versions, a larger one and a smaller one. The big one packs 5.8 teraflops and 8 terabytes of memory into a single cabinet. It runs on just 20 kilowatts of ordinary wall power. The smaller version provides more than half a teraflop.
HPCwire: How standard are the machines from a programmer's or a user's standpoint?
Rollwagen: They're standard MPI Linux machines.
HPCwire: How is the SiCortex architecture related to current HPC trends?
Rollwagen: These machines are purpose-built for HPC. They respond to issues with standard microprocessors and clusters that have gotten critical. There are growing problems with power consumption and with space, in the sense of both facility space for the machine and machine density for faster computation.
The SiCortex processor packs six MIPS processors on a die. The processors are very small, because these are RISC machines. In Intel processors, more than half the real estate is devoted not to logic, but to predicting the next memory references, and they include a lot of instructions scientists and engineers don't need. The SiCortex processor is almost entirely devoted to logic and was designed expressly for HPC users.
HPCwire: Who do you see as SiCortex's main competitors?
Rollwagen: The status quo, meaning companies that are fielding cluster products built with technologies designed for markets other than HPC. Clusters have done very well in the HPC market by leveraging the economies-of-scale of components sold in larger markets, but it's no secret that clusters have major limitations for HPC. Electricity bills and construction costs are quickly getting out of hand, not to mention the performance constraints that are inherent in machines with long wires. There need to be alternative designs for HPC.
HPCwire: What's the company's market strategy?
Rollwagen: We can enter at a level that's small as a percentage of the market. To start with, we'll aim primarily at the low-hanging fruit out there, customers with budgets in the $2-3 million range.
HPCwire: How confident are you that SiCortex will succeed?
Rollwagen: I'm very optimistic. I understand, though, that we still have important challenges ahead of us. First and foremost, we have to build the product so that it performs as promised.
John Rollwagen serves as a principal of Quatris Fund, an affiliate fund of St. Paul Venture Capital, and on several public and private company boards. He was a founding member of the Computer Systems Policy Project (CSPP), an organization of chief executive officers of the 12 leading computer systems companies in the U.S., created in 1989 to identify and advocate industry positions on trade and technology policy. In 1987, he was appointed by President Reagan as a member of the Advisory Committee for Trade Policy and Negotiations (ACTPN). He was reappointed to this position by the first President Bush.