Server Processor War Heats Up

By Kurt Shuler
Yesterday’s announcement that Intel will acquire Cray’s interconnect hardware program, including IP and 74 employees, is the latest salvo in the race to develop commercially viable massively multicore server processors.

On the surface, this acquisition seems like another instance of Intel beefing up its board-level interconnect technology, after having already acquired Fulcrum Systems (40Gb Ethernet switches and chips) and Qlogic’s InfiniBand IP and development team. eWeek states that the acquisition includes Cray’s “Gemini” interconnect ASIC IP, which can allow scaling beyond 1 million processor cores, and Cray’s “Aries” IP, which is the next generation interconnect after “Gemini.”

This may also be a strategic response to AMD’s acquisition of SeaMicro, a server maker whose core competency is creating the interconnect chips and IP that connect the CPUs and memory.

The real reason: ARM
But this was not a competitive response to AMD.

On the contrary, it was a response to CPU IP provider ARM, which is attacking Intel by enabling massively multicore server CPUs using its CPU IP. Cray VP Barry Bolding’s statement with EE Times journalist Sylvie Barak gives us the real reason:

“‘This is a good story for Cray and Intel,’ said Barry Bolding, Cray’s vice president of storage and data management. Bolding said the deal was motivated largely by the realization that as process technology continues to scale, chip makers had more space and were planning to integrate interconnects on-chip eventually.”

So the real reason is to implement Cray’s technology on-chip. As semiconductor critical dimensions shrink, what used to be on a board can now be shrunk onto a single chip. ARM has been very successful creating IP and an ecosystem to enable this. And I’m willing to bet “Gemini” and “Aries” will someday end up on Intel silicon.

Exascale computing = massively multicore SoCs
Much of this business activity in HPC server interconnects is probably being driven by the Exascale computing initiative, sponsored by the U.S. Department of Energy and its National Nuclear Security Administration.

Why DoE? In the old days, if we wanted to know how well our nukes worked as their initiators and fissile materials aged, we’d just blow them up in Nevada. But the Comprehensive Test Ban Treaty of 1996 ended that party. So now we must simulate the ageing and decaying of our atomic and thermonuclear bombs using supercomputers. And Uncle Sam is willing to pay billions of dollars to do this.

A limiting factor to the required Exascale computing is not just performance, but also power consumption and heat dissipation. Putting as much as possible on-chip, rather than on a board, helps both. And as we go down to 22nm and below, the opportunity is there for some massive integration.

Shrinking to success
What Intel’s acquisition of Cray’s interconnect assets means is that it realizes that Exascale computing will require not just massive numbers of server boards consisting of massive numbers of chips, but also a huge number of cores on each chip.

And just as today’s mobile phone application processor SoC is similar to a miniaturized PC motherboard 5 or 10 years ago, tomorrow’s server SoCs will be similar to complete server boards today.

Makes me excited to be in the semiconductor IP biz!

Sources:
• EE Times, “Cray sells interconnect hardware unit to Intel.”
• EWeek, “Intel to Buy HPC Interconnect Assets of Cray for $140M.”
• Cray (Gemini).
• Cray (Aries).

—Kurt Shuler is vice president of marketing at Arteris.