Taking A Long View On HPC And Beyond
August 5, 2016 Timothy Prickett Morgan
Bad things sometimes happen to good companies, but the great ones are resilient; they ride out the difficulties and keep forging ahead. So it will be with Cray, which does not just make massive-scale machines aimed at supercomputing centers but analytics engines that will see wider adoption among enterprises.
We have said it before and we will say it again: You have to take a long view of the high performance computing business – and we are using that term in the broadest sense – and not look at it on a quarter-by-quarter or even year-by-year basis. And so it makes sense to look at Cray along the timelines of the product cycles of the major compute and networking technologies, which probably run on the order of several years, and the product cycles of its customers, which run three, four, or five years.
The systems that Cray and its few rivals make are unparalleled in scale and complexity, the customers who buy them have the toughest requirements in the world, very formal and stringent acceptance processes, and a tendency to do most of their system approvals at the end of the year or early into the next. Sometimes, there are delays in the supply of key components used in systems, and that can cause massive swings of revenue from one year into the other because systems are architected with specific technologies that are not easily replaceable.
This is the normal condition of the HPC space, and something that bears repeating, and the companies and governments of the world need vendors to step up and make these machines, however tough it may be at times. The same holds true for IBM and its mainframe base or the ODMs and OEMs that are engaged in building rack-scale systems for hyperscalers and cloud builders. The both of these markets were down in the first quarter, by the way, and adversely affected the overall server market, and it seems likely there will be a downdraft in the second quarter when the IDC and Gartner numbers come out for the second quarter, too.
None of this is easy.
HPC customers, like their hyperscale and cloud cousins, were waiting eagerly for the “Broadwell” Xeon E5 v4 processors that Intel launched in March, and many had expected them to be widely available last fall. (A few cloud builders and hyperscalers actually got them a bit early.) Ditto for the “Knights Landing” Xeon Phi many-core processors, which are designed explicitly for simulation, modeling, and machine learning workloads that need lots of compute and memory bandwidth, which shipped initially last fall but only started to ramp early this summer. Customers have also been eagerly awaiting shipments of the “Pascal” P100 Tesla coprocessors from Nvidia, which made their debut earlier this year and which will be ramping up their volumes through 2016 and into early 2017.
To date, Cray has shipped over 100 racks of systems based on Knights Landing, Pete Ungaro, chief executive officer at Cray, said in his conference call with Wall Street analysts to go over the second quarter financials. “On Knights landing, we are continuing to work through technical issues that could impact performance and further impact timing, especially as we scale up to large systems,” Ungaro said. “However, while we continue to work with Intel and their partners to resolve these remaining issues, we began shipping these processors in our XC and CS supercomputers and have been for a couple of months now.”
Ungaro did not elaborate about what the issues were with the Knights Landing chips.
Cray has shipped several Broadwell Xeon E5 systems and has not seen issues with these processors, and has gotten the Pascal Tesla GPU accelerators in house for testing and seeing how they scale in larger systems. “We expect to begin receiving these processors in volume shortly, and we plan to begin shipping them to customers for significant deliveries later in the year,” Ungaro said of the Pascal Teslas. “We are still behind our original integration schedule here. Ultimately, we expect to work through the remaining issues on each of these processors to achieve the anticipated acceptances before the end of the year.”
These transitions as well as some uncertainties in the global economy will cause a slight dip in spending on HPC systems in 2016, and in the high-end supercomputing segment in particular, according to forecasts by IDC – something that Ungaro pointed out.
“In the highest end of the supercomputing market, were we focus, we believe that our addressable market is down year-over-year,” Ungaro explained. “We are not sure how long this slowdown will last, but at this point we believe it will continue through the rest of the year and possibly into 2017. We don’t see this as a longer term trend, but it is clearly impacting our current view.”
Ungaro also explained that delays with third party components are impacting the level and timing of new orders (in contrast with the big systems that were already on order), and this means Cray’s forecasts going forward need to change.
Earlier this year, Cray was forecasting it would bring in around $825 million in revenues in 2016, with about 60 percent of that, or $495 million, coming in the fourth quarter as several large systems were expected to be accepted. Back in April, when going over its first quarter results, Cray had predicted that sales would be about $100 million in the second quarter. The company hit the nail on the head with the Q2 sales, but now believes that sales for all of 2016 will be in the range of $650 million, with a wide range of possibilities depending on the delivery of components and acceptance of the systems.
“In hindsight, we thought what might have been the beginnings of this in the first quarter of this year, where we previously indicated weak order activity,” Ungaro said. “Orders then improved a bit at the beginning of the second quarter only to slow down again in the latter half of the second quarter through today. The exact drivers for this slowdown have been difficult to pin down, but we are hearing a similar perspective from multiple sources in the market. So the market seems to be going through a far slower period this year than anticipated.”
The change in Cray’s 2016 forecasts was not just due to component delays and issues with the systems based on Knights Landing Xeon Phis that Cray is working with Intel and customers to resolve. A “smoke event” caused by the failure of a power component at its smaller factory in Chippewa Falls, Wisconsin happened just as the second quarter was coming to a close, and five smaller systems that are part of some larger projects that Cray is working on and that were stored in the bay where the smoke event happened were damaged. It is important to note that this was not a fire, and it was not in any way the result of a failure in a Cray computer.
Cray expected for some of these now damaged machines to be accepted in the third quarter, others by the end of the year. The majority of the loss will be covered by insurance, according to Ungaro, who added that the impact on 2016 revenues will likely be more than $20 million and as much as $60 million.
You can’t predict a smoke event in a factory, but you can insure against it as Cray has done. And you can’t predict how complex components will behave when they come to market, or when they will get there. In a way, uncertainty is the nature of the high end of the computing market, and always has been. That is what it means to be on the cutting edge.
Dealing with such uncertainty is something that Ungaro and his team have proved adept at. Cray has the broadest mix of supercomputing and analytics technologies and perhaps the closest alliance with Intel, which is obviously very keen on getting its Xeons and Xeon Phis into as many systems, both in HPC centers and enterprise datacenters, as it can. Cray also has very tight alliances with Nvidia and Mellanox Technologies, and is well positioned to be able to provide high performance systems no matter which way the technology rises and falls.
Looking ahead, revenue at Cray is expected to be around $80 million in the third quarter, and that means Cray is expecting about $364 million in sales in the fourth quarter. By the way, over $100 million in revenues that are part of that $650 million projection are tied to big Knights Landing machines being deployed at Argonne National Laboratory, the National Energy Research Scientific Computing Center at Lawrence Berkeley National Laboratory, and a shared system for Los Alamos National Laboratory and Sandia National Laboratory. (Representatives from these and other supercomputing facilities were singing the praises of Intel and Cray as the Knights Landing rollout was commencing in earnest at the ISC16 conference in Germany.)
The thing to remember about revenues being pushed out, if it comes to pass, is that the sales are not lost. All the hard work that Cray and its partners have done still turns into money, even if it takes longer to realize it. We still believe that Cray will be the first supercomputer maker to break through $1 billion, and had projections held, it might have kissed that magic number in 2017. It may take until 2018 or 2019 to do it now. The important thing is that Cray is steady and adamant that its strategy of diversifying its product line to address more enterprise use cases while capturing more of the traditional HPC market is still the correct course for its future.
Cray is right.
The company will compete in HPC and waterfall down into the enterprise and it will eventually get a third of its revenue from enterprise accounts (manufacturing, financial services, life sciences, and energy are the biggies) and not be tied so tightly to traditional HPC shops. Both HPC and enterprise IT spending will bounce back, and Cray will be there to capture its increasing share of both markets.