Supercomputer maker Cray is methodically and inevitably shifting its technology focus from hardware to software. Another step in that direction played itself out this week in the company’s sale of its highly treasured supercomputing interconnect technology. On Tuesday evening, Cray and Intel announced that they signed a “definitive agreement” that would transfer the interconnect program and expertise to the x86 chipmaker.
The sale would send $140 million in cash to Cray, and would transfer the majority of the supercomputer maker’s interconnect IP and corresponding 74-person development team to Intel. The acquisition includes 34 patents (or applications), representing 15 percent of Cray’s patent portfolio. The transaction is expected to close by the end of the quarter.
The deal is important because it gives Intel a high performance, scalable interconnect technology to potentially integrate into its high-end server processors. At the same time it gives Cray a nice influx of cash and relieves the supercomputer maker of the significant R&D expense of maintaining an ongoing interconnect technology program.
On the downside, Cray is sending a critical piece of its technology into the general circulation. As pointed out in a commentary by Wolfgang Gentzsch, there are reasons for concern:
I can imagine that selling these interconnect assets could make Cray more vulnerable against their competitors, in the mid-term, because the company is losing a great differentiating asset (and patents), and that would be a pity. Cray now has to compete on the integration and services levels, which is more difficult to do, because every competitor tries this.
For the time being though, it’s going to be business as usual at Cray. The acquisition doesn’t effect the current Gemini interconnect in the existing XE6/XK6 supercomputers, only the technology tied to the forthcoming Aries interconnect that is slated to appear in Cray’s next-generation “Cascade” system. The deal will also encompass Pisces, a future Cray interconnect that is in the early stage of development.
As Cray CEO Peter Ungaro explained in a conference call with investors and analysts on Wednesday, the motivation behind the technology sale has to do with the technology shift he and others see coming for these system networks. Over the next several years, interconnect processing will likely be moving from standalone chips onto the central processor. This is apt to be especially true for CPUs destined for large-scale clusters, and most especially for supercomputers. “It’s been clear that the important technology trends are starting to intersect, enabling more powerful and energy-efficient systems interconnects by more closely integrating this technology with the processor,” explained Ungaro.
The agreement excludes Intel from reselling standalone Aries chips to anyone but Cray, which puts a protective bubble around the Cascade supercomputer over its projected lifetime (2013 to 2016). After that though, the supercomputer maker is effectively out of the interconnect business. “What will be changing in 2017, and beyond, is that we won’t continue to build system interconnect hardware,” said Ungaro.
What that leaves for Cray is its other differentiating technology assets, namely HPC system software, compilers and tools as well as its custom designs for cooling and packaging supercomputers. It also allows Cray to focus on its newfound ventures in big data and storage, which they intend to integrate more intimately with the rest of their supercomputing offerings over the next several years. The majority of these technologies revolve around software. As commodity hardware has worked its way up the food chain into HPC, its the software that, more and more, has become the differentiating element for the kinds of high-end systems that Cray sells.
Over the years Cray has slowly weaned itself away from relying on its own custom hardware, as general-purpose technology has become powerful enough to support supercomputers. Witness the fact that Cray no longer uses its home-grown vector processors or other CPUs of its on making for its flagship supercomputers, relying instead on industry standard x86 chips and GPUs. The move to jettison the interconnect IP is another step down that road.
From Intel’s point of view, they get am additional interconnect technology in their collection (alongside QLogic’s InfiniBand plus Fulcrum’s and NetEffect’s Ethernet technology), as well as the human expertise to drive the technology forward. Intel has yet to outline its plans for the Cray IP, but since the company intends to be the chipmaker-of-choice for the largest machines in the world — both supercomputers and ultra-scale enterprise clusters — it’s certainly conceivable that Aries technology will find its way onto future Xeon and/or MIC chips before the end of the decade.
The acquisition also serves to counter AMD’s recent buy of SeaMicro, in which the Intel rival took ownership of SeaMicro’s supercomputing fabric ASIC. Like Intel, AMD is also likely to integrate that interconnect technology on-chip.
Assuming Intel, AMD, or both, are able to successfully integrate these high-performance interconnects onto commodity parts, that’s good news for HPC. Getting this critical element of supercomputing technology onto a tick-tock processor cycle is preferable to the slower pace of development possible under a small vendor like a Cray or SeaMicro.
This would put pressure on companies like Mellanox and some Ethernet NIC vendors who could see a good chunk of their market disappear onto the processor. And it could also persuade vendors like Fujitsu and IBM, who build supercomputers with their own custom-built interconnects, to rethink their strategy. But that’s a story for another day.