by Alan Beck, editor in chief
San Diego, CALIF. — Seymour Cray dominated the field of high performance computing for more than 40 years, and despite his death in an October 1996 car accident, he’s still at it. With less than two weeks to go before the IEEE/ACM SC2000 show, there’s plenty of buzz about what SRC Computers, the Colorado-based company Cray founded just three months before his death, is calling “Cray’s final design.” SRC will preview Serial Number 1 of its SRC-6 system in Booth #278, and expects to formally announce the machine and begin shipments in 2001.
To get the skinny on the forthcoming system, HPCwire talked to Chuck Breckinridge, co-founder and director of SRC Computers and a colleague of Cray’s since the days of Control Data Corp’s 1604 computer in 1959.
HPCwire: Cray died four years ago. What is the extent of his involvement in this design?
BRECKENRIDGE: The SRC-6 design is largely Seymour’s. The changes from his original design came as a natural evolution of the development process. They’re the work of an engineering team Seymour selected. So, it really is accurate to say, “The legend continues.”
HPCwire: SRC Computers was just 3 months old at his death. Did you consider disbanding the company?
BRECKENRIDGE: We made the decision almost immediately to continue despite the loss of Seymour. We had identified the opportunity for the company, and that hadn’t changed. The essential design of the system was complete. The engineering team that Seymour had put together was in place. In view of those factors, it made perfect sense to continue.
HPCwire: The HPC marketplace has changed pretty dramatically since Cray’s death in 1996. How have those changes affected the approach you’re taking with the SRC-6?
BRECKENRIDGE: The market has moved strongly in the direction Seymour anticipated. He felt that other companies would be reluctant or unable to scale their shared-memory architectures beyond tens of processors, and believed that was something SRC could improve upon – which we have. He also felt that the general-purpose computing market would never emphasize high-bandwidth, low-latency interconnects as aggressively as high-end users, both commercial and scientific, would really like. These things have played out in the market just as Seymour expected, which leaves a real opportunity for SRC.
HPCwire: Let’s talk about the new machine. What’s innovative about it?
BRECKENRIDGE: The SRC-6 continues the trend of leveraging Moore’s Law with microprocessors, but also unleashes over five billion gates of programmable logic in each system. It also aggregates up to 512 Intel processors on a common memory switch, to provide a high-bandwidth, low latency, shared-memory model and communicating on a uniform memory- access basis. We believe the performance implications for both peak and sustained measurements are potentially staggering.
HPCwire: So you’re combining commodity microprocessors with reconfigurable logic?
BRECKENRIDGE: Exactly. The performance improvement curve of field-programmable gate arrays (FPGAs) is extremely steep. With the increasing density of these components over the next few years, there is the very real and new potential to capture meaningful pieces of algorithms directly into hardware. This promises huge gains in both performance and efficiency for real applications, and as programming tools are developed, will provide a means of adapting the hardware to the software rather than the reverse.
HPCwire: Cray had a long career in high performance computing. Did he have a consistent vision for what supercomputing could be? How does the SRC-6 fit with that vision?
BRECKENRIDGE: Seymour’s vision for the high end of computing was remarkably consistent in many ways throughout his career. One of his great talents was in optimizing the performance of discrete components, whether they were vacuum tubes, transistors or integrated circuits. In a way, utilizing microprocessors is no different from his earlier work – it’s just that now we have a higher level of component integration. Seymour saw the microprocessor as another commodity component that he could buy off the shelf and find creative ways to optimize its performance.
HPCwire: What applications does the SRC-6 architecture best fit?
BRECKENRIDGE: This is a question that won’t be fully answered until we deploy the systems and work closely with our early adopter customers. During 2001 most SRC-6 systems will very likely end up in the world’s leading government and university research centers. We intend to work in partnership with this inspired group to answer the question definitively. The fact that most of these sites have multi-disciplinary applications expertise is very attractive to us because we’ll undoubtedly find some “surprise” star performers.
Now, having said all that, we do have some applications areas that we feel have strong potential for this architecture. Examples would be signal and image processing, bioinformatics, business intelligence and data mining, scientific and technical, and some Internet infrastructure applications. The SRC-6 will perform very well for applications that are well adapted for performance on the microprocessor.
HPCwire: What are the common elements to look for if you’re evaluating your applications for the SRC-6?
BRECKENRIDGE: If your applications and algorithms have intensive floating-point vector and matrix operations; if they perform bit operations on long vectors or matrices; if they have multiple floating point, integer and/or bit operations that can be concurrently scheduled for execution – then you’ll definitely want to take a look at the SRC-6.
HPCwire: When you delivered Cray’s memorial address at SC96, you said Seymour had a gift for separating the important from the unimportant and focusing on the important. How did that principal play out in the design of this system?
BRECKENRIDGE: It was probably that single principle that drove us to use Intel microprocessors, to focus on the memory bandwidth and latency, and to use reconfigurable technologies.
HPCwire: There’s been a significant trend in the industry over the last several years toward standards and compatibility, via either hardware or software. It sounds like you’re breaking with the pack – is that right?
BRECKENRIDGE: Well, it wouldn’t be a Seymour Cray design if it didn’t!
You’re right that the pendulum has swung very far toward standards. But bear in mind that many users are wringing their hands because they’ve spent the last 5-10 years pushing to find the edge of the envelope for scalable parallel architectures. Those efforts have produced some extremely impressive results, but have also made the limitations very visible – I’m talking about the limitations in inter-processor latency, bandwidth, and overall low sustained-to-peak performance. It was Seymour’s intent with the SRC-6, as always, to address these inherent weaknesses. Success in this endeavor will benefit users in bioinformatics, aerospace, petroleum, automotive industry, leading edge research – those users are crying for performance.
Most HPC vendors today are going after incremental performance increases. SRC is one of a very few companies trying to create enough innovation to take the industry a giant step forward. Seymour always looked for potential architectural changes that provided so much greater performance that you couldn’t afford not to consider them. I think the SRC-6 architecture will be comparable to the Cray-1 architecture – it may require some changes and software development to take advantage of it, but the potential performance improvements are so significant that users will find, once again, that the extra effort will pay huge dividends for many applications.