Features:

CRAY: WE WILL BE BACK IN 2005
by Tom Tabor, Publisher

Cray Inc. has said it will have the HPC industry's strongest product portfolio in 2005. The company reported disappointing financial results for the second quarter, but expects to return to profitability in early 2005. HPCwire's publisher, Tom Tabor, asked Cray President and CEO Jim Rottsolk and Pete Ungaro, vice president of marketing and sales, for an update on the company's product plans and financial outlook.

Tom Tabor: Cray was rather profitable throughout 2003. Then you reported a 2004 second-quarter loss of about $54 million. Although you said you expect to be profitable again early next year, what happened exactly?

Jim Rottsolk: The main issue has been a substantial slow-down in defense spending that heavily affects 2004 revenue for our current Cray X1 product. The good news is that we've continued to win major contracts for our next- generation products from customers like Sandia, Oak Ridge, the Ohio Supercomputer Center and others. Our order backlog jumped to $126 million at the end of the second quarter, compared to $75 million at the end of the first quarter.

TT: What's the financial picture beyond 2005?

RJ: In the years beyond 2005, we'll see the full effect of our three product lines: the Cray XD1, which targets the small/mid-range HPC market; other Red Storm-based systems; and the Cray X1E. Needless to say, we're pretty excited about our prospects. I won't be more specific than that.

TT: You're reducing your cost structure by about 20 percent, including a workforce reduction. How will that affect product development?

RJ: While workforce reductions are always difficult, we're confident it won't have a major impact. The reality is, we're very far along in the development process for all three next-generation products-the X1E, Red Storm and the XD1. Some very talented people who helped get us to this point will no longer be with the company, and that's something I regret, but we still have the HPC industry's deepest reservoir of talent, including our new colleagues from the former OctigaBay, which is now called Cray Canada. On a one-by-one basis, we're also continuing to hire some of the top talent in the industry.

TT: Can you be more specific about the product development schedule?

RJ: Sure. With our next-generation vector product, the Cray X1E, we've experienced a delay mainly from a supplier, so we'll start shipping a bit later than planned, but we still expect the Cray X1E first customer shipment to happen by the end of 2004. We're now bringing up the first phase of the Red Storm system in our labs. We've sold Red Storm-based systems to five customers, including Oak Ridge and Pittsburgh, and are seeing strong additional demand for that product. The former OctigaBay product, now called the Cray XD1, is ahead of schedule. We now expect to see initial revenue from that product in late 2004, instead of early 2005. We've already started to announce Cray XD1 orders. There's very strong interest in this product, which as you know starts at under $100,000 and is strongly differentiated from the Linux clusters and SMPs that are available in its price range.

TT: Can you say more about the status of the former OctigaBay product and how that acquisition is coming along?

Pete Ungaro: We have brought in the OctigaBay system as a full-fledged member of the Cray product family and we call it the Cray XD1. The integration of the two companies has gone really well and the teams are already collaborating together to make all of our products stronger and integrate the best of each product into our future roadmap. As I said earlier, the XD1 is ahead of its original schedule and we already have beta systems at multiple user sites. We're starting to get good advance order momentum, even before we officially roll out this product. We plan to report some early results for the Cray XD1 system at the September HPC User Forum meeting in Tucson, so I'll hold off on the particulars till then. Suffice it to say we're very excited about the Cray XD1 product, as well as the upcoming Red Storm and Cray X1E products.

TT: So, you still believe Cray will have the strongest product offerings in the HPC industry starting in 2005? What about the competition?

RJ: Almost all vendors today address the HPC market with cluster products that began their lives as collections of general purpose business servers, rather than single, scalable systems purpose-built for HPC workloads. When you try to bundle together separate computers like this, interprocessor communications, system integration and system administration can become really challenging and add a lot to the cost of ownership. So even though the initial pricing of clusters might seem attractive, the amount of real work you can accomplish for the same money is typically much less than with a purpose-built HPC system.

TT: How does Cray's approach compare with this?

PU: We have the exact opposite approach. Instead of trying to scale up commercial systems, we bring the power and technology of purpose-built supercomputers to the broader HPC market, starting from under $100,000 on up to the largest systems in the world. In 2005, we'll have a range of purpose- built HPC products for customers to choose from. We'll have real choices specifically built for HPC customers. Two of the products, the Red Storm- based system and the Cray XD1, will use open and economical Opteron and Linux technologies, combined with Cray high-bandwidth, low-latency interconnects and integrated software (the ingredients of our secret sauce) in purpose-built HPC architectures. The Cray X1E is a substantial upgrade to our current Cray X1 vector MPP system and will push high bandwidth computing to a new price/performance level.

TT: How much faster will the new Cray products be than competing products?

PU: Again, I don't want to get ahead of the early results we'll be talking about in Tucson-and they will only be early results-but I think we're going to see stellar performance on meaningful benchmarks, and even more so on challenging real-world applications. The Cray X1 is the only U.S.-made supercomputer that's proven to be in the same class as the Earth Simulator. We announced a while ago that on a climate application at Oak Ridge, the X1 was actually 50% faster than the Earth Simulator on a per-processor basis. The successor Cray X1E system is going to raise that performance leadership bar considerably. When you design systems with a strong balance of fast processors, fast interconnect and fast I/O, as Cray does, you're going to get industry-leading performance. This is what Cray as a company is all about.

TT: What impact do you think the Cray product line-up will have on the HPC market?

PU: In recent years, makeshift HPC systems-clusters made up of business servers-have taken much of the HPC market share. I think we'll see that powerful trend begin to reverse, now that purpose-built HPC systems are incorporating economical COTS technologies and people are learning how difficult it can be to extract the theoretical processing power out of clusters and make them run efficiently on real workloads. A good example of this reversal is Red Storm, which is going into Sandia, Oak Ridge, the Pittsburgh Supercomputing Center and some other places we haven't named yet. The Cray XD1 is also winning business around the world that in most cases would have gone to standard Linux clusters a year ago, because there were no good alternatives in that price range. The reversal won't happen overnight, but it's interesting to see this shift in HPC market dynamics. HPC customers are really starting to see the value, once again, in systems architected from the ground-up for their demanding problems.

TT: Where does Cray's vector line fit into this picture?

RJ: As you might expect, Cray has a long-term commitment to vector processing. We believe that for an important range of problems there's no substitute for very high bandwidth vector processing. We're targeting increasingly powerful vector capabilities as far out as our product roadmap currently goes. Many of our Cray X1 contracts have upgrade paths to these follow-on systems. We're very excited to see what the Cray X1E and its successors will do for customers.

TT: What about the new HPC Challenge benchmark versus the Top500's Linpack and other tests? How important to Cray is the battle of the benchmarks?

RJ: As a rule, the tougher and more comprehensive the benchmarks, the better Cray systems do. The TOP500 list and Linpack are useful to produce a census of HPC systems based on processor performance, but as Jack Dongarra has repeatedly said, no single test can accurately reflect the overall performance of HPC systems. Looking at processor performance in isolation from other system attributes has its limitations. We think the HPC Challenge suite, which includes Linpack, is far more valuable as a high-level benchmark. So do some of our most important customers. We're pleased that the Cray X1 turned in the strongest overall performance of any HPC system on the HPCC suite. The IDC Balanced Rating is another helpful high-level benchmark suite. Clearly, though, we build our systems for real customer application performance, not kernel-level benchmarks.

TT: Who do you see as your major competitors?

PU: There are a number of vendors who are formidable competitors mainly because of their size and resources, but our biggest competitor is the status quo in the HPC industry. Theoretical peak performance and Linpack-based rankings still carry too much weight in many customers' decision-making models, mostly because they are easy to calculate and understand. We find end-users having to spend a lot of time modifying their applications to fit commodity clusters, instead of pushing the bounds of engineering and science. We are striving to build fundamentally different systems that deliver real value to our customers around the world. That said, I'd add that it is great that programs like the U.S. DARPA High Productivity Computer Systems program are pushing for usability and on areas like more programming model options so the HPC industry can get past some of these current status quo limitations found in the typical cluster. We're really excited to be part of this program with our Cascade project and take usability to heart with all of our products.

TT: What keeps you up at night?

RJ: The thought that the HPC industry might not break out of its current technological stagnation, that the lesson of the Earth Simulator might be forgotten. Cray's reason-for-being is to innovate, to create better and better HPC tools so users can advance their science and engineering work. Clusters are not very innovative.

PU: I'd agree with that. We're already seeing the "machoflops" race heating up again, with vendors touting the peak performance or Linpack performance of large-scale systems, with little or no reference to sustained performance on real applications. At Cray, we're keeping our eyes on the real prize, which is helping customers to do better science and engineering. Our supercomputers are fundamentally different and exciting, purpose-built for the HPC customer. HPC users include some of the world's most creative scientific and engineering minds. It's going to be great seeing the new insights they can achieve, and the rapid progress they can make, with our powerful new tools. That's what we live for at Cray.

TT: Thank you, Jim and Pete. Any final comments?

RJ: It's a very exciting time in the HPC market. We're proud of Cray's heritage as "the supercomputer company", and we're excited about the market prospects ahead of us. Thank you for the opportunity to address the readers of HPCwire.