Cray results suffer from AMD delays
History repeats itself, but won't threepeat
Advanced Micro Devices has made business tough for supercomputer maker Cray once again. But it won't get a third chance.
If Cray had its way, Advanced Micro Devices would have been shipping its 16-core "Interlagos" Opteron 6200 processors in volume for a few months already. But the 32 nanometer ramp at AMD wafer baker GlobalFoundries has had a few bumps, and these chips, which are part of the 2011 upgrade cycle for the XE6 parallel supers, had an adverse impact on Cray's business in the third quarter.
In the third quarter ended in September, Cray's product revenues, including midrange and high-end supercomputers, was only $16m, even less than the anemic $23.5m it posted last year. (32 per cent, to be precise.) Services revenues for the company rose by 6.9 per cent, to $20.7m, and so overall sales in the quarter were only down 14.3 per cent. Cray really put the squeeze on services costs in the quarter, and therefore only booked a loss of $12.3m, which was better than the $18.5m loss it had a year ago.
If you listened in on the conference call with Wall Street analysts to go over the Cray numbers, you would have to infer that the delay was due to the late Opteron 6200s. Cray CEO Peter Ungaro never mentioned AMD, Interlagos, Opteron 6200 by name, and only said that a "supplier delay" had "pushed back on our calendar." In the question and answer session, however, Ungaro admitted that the shortage was indeed for the Opteron 6200s processors, which are not only used in the XE6 machines but also in the new XK6 hybrid CPU-GPU supers announced in May.
The delay of the quad-core "Barcelona" Opterons in early 2007 had a ripple effect on AMD, and caused a delay in the dual-core "Budapest" Opterons that Cray used in the XT4 generation of parallel supers. This gave Cray a big haircut.
Since that time, Cray has shifted to the higher-end Opteron parts in its systems and, based on the number of Opteron 6200 chips it has been able to get for big supers that shipped during the quarter (but which were not yet accepted by customers and therefore cannot be counted in the revenue stream), Cray has done a very good job of getting itself to the front of the line.
With the next-generation "Cascade" supers in development at the US Defense Advanced Research Projects Agency, the development lab for the US military, Cray is hooking its proprietary interconnects into processors through their PCI-Express buses, unlike the XT and XE systems, which rely on the HyperTransport interconnect of the Opteron chip. Finally, Cray will be able to have two sources of CPUs in its parallel machines.
By being at the front of the line for the Opteron 6200s, Cray was able to ship both the upgrade to the XE6 and the XK6 on time in the third quarter, as planned – just not at the volume it had hoped. Ungaro said that Cray has sold a multi-cabinet XE6 system to oil giant ExxonMobile in the quarter, and also shipped the 96-cabinet "Cielo" system, which will weigh in at over 1 petaflops of aggregate number-crunching power, to Los Alamos National Laboratory.
Cray also landed the "Titan" upgrade to the "Jaguar" system at Oak Ridge National Laboratory, and has started shipping the machines for the first phase of the contract, which will essentially convert this machine from an Opteron-based XT4 to an XK6 ceepie-geepie.
Even with all of the hustle, Cray is not banking on making the numbers it once hoped for the full 2011 year, although this is technically possible according to Ungaro because the key monster XE6 and XK6 systems needed to make the numbers have either shipped already or will within the next week. It all comes down to when the customers say send the bill.
And because of the uncertainty, Brian Henry, Cray's CFO, told Wall Street to expect full-year sales of between $290m to $320m, with about $60m of that coming from the Titan upgrade at Oak Ridge. That's down from the guidance of sales of between $300m and $340m that Cray was talking about back in July, when it was also saying it would be profitable for the year. The company continues to believe that it can be profitable, even with the lowered guidance for Q4.
By the way, this 10-petaflops Titan upgrade is a bit tricky because it involves swapping hardware in the running Jaguar machine, which will still be supporting workloads as Cray does the upgrade. The company will be metaphorically trying to change the tire as Oak Ridge drives down the highway - which should be possible with a massively parallel machine, after all, since this is one of the inherent benefits of parallel systems.
Ungaro said in the call that Cray had restructured itself into three different groups to better chase opportunities, effectively busting up the Custom Engineering business that was growing very fast in 2010 and which was expected to deliver around $40m in revenues this year. The expectation now is for CE to have revenues "in the high 30s" for 2011.
The new groups are: HPC Systems, which focuses on the midrange and high-end parallel supercomputers; Storage and Data Management, which does exactly what it says; and Knowledge Management, which will make appliances to chew on big data. Ungaro said that it will beta test a new big data system by the end of the year and start shipping it in early 2012.
Henry said that Cray passed another $12m milestone for the DARPA Cascade development contract in the quarter, and that most of this money would be booked in the fourth quarter against R&D costs. (This DARPA contract is not booked as revenue, but as a reimbursement for doing the work to create Cascade.)
If you check out the 8K filing that Cray made concurrent with its financial report today, you'll see that DARPA cut back on the Cascade contract for the second time, this time shaving off $10m to a total of $180m. Cray was a bit vague about exactly what changed in the Cascade system, but in its filing said that the number of cabinets it needs to deliver has been cut as well as tweaks the Cascade delivery date; most importantly for Cray, the effect of the changes will be neutral to its profits. Cray has a remaining $22m in funds coming as part of the DARPA contract, which will be broken up into $12m, $3m, and $7m milestones in 2012.
Back in January 2010, DARPA gave the Cascade project a much shorter $60m haircut on phase two of the High Productivity Computing Systems program, which originally had $250m allocated to it. Phase one, which was awarded way back in 2003, gave Cray $43.1m to develop the hardware and software ideas underpinning the Cascade system.
IBM got $244m for its phase two effort, called PERCS and related to the "Blue Waters" system that is no longer being installed at the National Center for Supercomputing Applications at the University of Illinois. IBM flushed the Blue Waters system deal with NCSA back in August because it was more expensive to build than anticipated. IBM is still shopping around for a buyer for the machine, and rumor has it that a buyer has emerged. ®