Features:

CRAY TO BUY OCTIGABAY
by Alan Beck, Editor in Chief

After U.S. stock markets closed today, Cray announced plans to acquire OctigaBay, the Vancouver, Canada-based company that is completing development of an AMD/Linux system with high bandwidth, low latency and prices starting at under $100,000. HPCwire Editor-in-Chief Alan Beck had an exclusive interview with Cray Chairman and CEO Jim Rottsolk and OctigaBay President and CEO John Seminerio about this important development.

HPCwire: Jim Rottsolk, what is Cray's rationale for this acquisition?

JIM ROTTSOLK: There are two main justifications. First, joining forces with OctigaBay will enable Cray to address the entire HPC market, rather than just the high-end capability segment we address today. Secondly, both companies share a common product design philosophy. Both companies strongly believe in the importance of balanced architectures, where fast processors are matched by fast internal communications networks and I/O. Almost all HPC systems available today are unbalanced and therefore not very efficient or cost- effective for challenging HPC codes.

HPCwire: Why is it important for Cray to expand beyond the high end of the market?

JR: We want to be big enough and strong enough financially to continue for many years as the leading innovator for high-end customer requirements. The new Cray has been gaining market share and momentum every year in the high-end capability segment, and this will remain our flagship market. We'll continue to grow substantially in 2004, but it's no secret that innovation is expensive. That's why most HPC vendors leverage business computers into this market, rather than providing purpose-built HPC systems. We also believe that purpose-built HPC systems can have enormous value beyond just the high end. In conjunction with OctigaBay, we plan to provide systems at a wide range of price points that allow customers to get more work done at a lower cost than they can with mainstream HPC systems today. Keep in mind that this acquisition is not the first step Cray has taken to expand beyond the capability segment. The Red Storm-based product we are developing is aimed at what IDC calls the enterprise segment as well as the capability segment. The OctigaBay product is for the lower-price point departmental and divisional segments.

HPCwire: John, the OctigaBay booth was one of the busiest at SC2003, where you previewed the OctigaBay 12K product. Was the kind of innovation Jim referred to the cause of all that excitement?

JOHN SEMINERIO: Absolutely. In recent years the HPC market has been starved for innovation, especially at the entry-level and midrange, the range from under $100,000 up to about $2 million. Vendors have ignored this part of the market. That's why the only choices are either seriously unbalanced clusters, which are an exercise in system integration as opposed to innovation, or seriously expensive SMP systems. We're bringing radical innovation to this part of the market. We're breaking the communications bottleneck with a much faster interconnect than anything else available today in this part of the market. We're adding commercial-grade reliability and manageability. We're making it possible to scale without a penalty, and a lot more. From the user's standpoint, all this extra capability will be transparent. The OctigaBay 12K uses standard AMD Opteron processors and will run any x86 Linux application without modification.

HPCwire: How will OctigaBay benefit from the union with Cray?

JS: Cray is the right company for OctigaBay to join with. Cray has a phenomenal reputation, brand name and market strength. Also, Cray is the right size. OctigaBay won't get lost within Cray. We can truly contribute to the company's success. Our dream has been to achieve large market acceptance for the 12K and its successors. Cray shares the same vision for high performance computing: balanced architectures. Our products are complementary. We address different market segments and different price points. The Cray sales force can sell our products without special training. Cray's very familiar with Opteron and the HyperTransport technology. Also, we can learn from each other about Opteron compilers, ASIC development and other things. There's a long list of benefits.

HPCwire: Without getting too technical, can you summarize the 12K's innovations?

JS: I'll be a little technical. The OctigaBay 12K is designed around the Direct Connected Processor architecture, an innovative design that directly links together processors to remove the memory contention and interconnect bottlenecks found in cluster and SMP systems. This is very similar to the design philosophy Cray uses in the high-end HPC market. Our interconnect uses dedicated communications processors and an embedded switch fabric designed with 30 times higher bandwidth and 30 times lower latency than typical cluster systems available today. Innovations in management include self-monitoring and self-healing features for high availability, and single system command and control for simplified system administration.

HPCwire: When you say the acquisition will allow Cray to address the entire HPC market, do you mean every opportunity in the market?

JR: No. The combined company will be strongly positioned to capture communications-intensive work in all segments of the HPC market. This essentially means the most challenging problems and workloads in every segment. This is what Cray already does in the high-end capability segment, and what the OctigaBay 12K is designed to do in the entry-level and midrange segments.

HPCwire: Does the Cray sales force have the experience to sell the 12K into the low end?

JS: Our go-to-market strategy is aimed at government labs, large academic centers and other early adopters the Cray sales force knows very well. These high-end early adopters don't just buy high-end systems. Our plan was to build credibility by selling to these customers first, then use that credibility to sell into the non-high-end segments. But at SC2003 and ever since then, we've also had interest from customers who aren't at the high end. They're talking to us now, even before we deliver anything to high-end customers. Another important point is that Pete Ungaro, who joined Cray from IBM recently as head of worldwide sales and marketing, knows a lot about selling into lower-price point HPC markets. He's not the only one. Ulla Thiel, who was head of IBM's HPC sales in Europe, just became Cray's European sales leader.

HPCwire: What is the current status of the OctigaBay 12K?

JS: It's in the final development stage. The system has passed lab tests and is up and running at OctigaBay. It's scheduled to begin customer trials shortly with general availability in early 2005.

HPCwire: Can you be more specific about your definition of "balance"?

JS: Sure. We're talking about the balance between the processor speed in a system, and the bandwidth available to keep the processors busy. When that balance is off, the computer is inefficient, especially for challenging problems. Balance is expressed as the ratio between flops and bytes. A ratio of 1.0 or greater means you have a balanced architecture. Anything less than 1 and you have a system where the processors are in danger of being starved for work. The OctigaBay 12K system has a ratio of 1.0. The Red Storm system is at 1.6, and the X1 is at 2.0. These are balanced systems. Contrast those numbers against current clusters and SMP systems from the large U.S. vendors, which have ratios of less than 0.2. Even the upcoming IBM Blue Gene/Light system has a ratio of less than 0.4. The commodity HPC systems out there are much worse.

HPCwire: So, what are clusters good for, in your opinion?

JS: We think of clusters as "good enough" computing. They're good enough for embarrassingly parallel applications and other codes that don't need much interprocessor communication. They're good enough for doing current science a little better, a little faster, as long as the problems aren't complex. They're not good enough for new science and engineering, for achieving new insights and competitive advances.

HPCwire: Is this acquisition, along with the Red Storm product plans, a signal that Cray plans to reduce its reliance on vector supercomputers?

JR: Absolutely not. We're fully committed to our product roadmap, which includes the Cray X1E and multiple generations of successor "BlackWidow" vector supercomputers. Accelerating our growth through this acquisition will help ensure that Cray can fortify its position as the leading innovator in vector technologies.

HPCwire: Cray is already doing many things. Does Cray have the people and resources to take on the OctigaBay 12K, too?

JR: We do. Remember, Cray isn't just acquiring the OctigaBay 12K, we're also acquiring the OctigaBay workforce that created the OctigaBay 12K and that has advanced it to the final development stage. OctigaBay will become a Cray division dedicated to bringing the 12K to market and developing its successors. The annual cost of operating this 66-person division will be modest.