ARM will get its first real crack at that market this year, as shipments of its ARMv8 architecture — ARM’s first chips suitable for data center servers — ramp up. A few designs are already in production. The key is going to lie in making ARM-based processors less of a special case to data center operators.
“One of my goals is to make them incredibly boring,” Masters says.
That was the theme of his talk at the Linley Data Center Conference, held by research firm The Linley Group in Santa Clara, Calif., last week. Masters, who carries the title of chief ARM architect at Red Hat, explained why he’s an ARM believer and how he thinks ARM can make headway in the data center.
The 64-bit design is a milestone for ARM, which is best known for smaller chips that go into devices such as smartphones. With ARMv8, the company is hoping to get into larger systems, with data center servers being a particular target. (ARM isn’t the only x86 alternative — IBM‘s Power architecture is another option, one that Google apparently likes.)
Note that ARM itself doesn’t make or sell chips. ARM creates microprocessor designs that it licenses out to chipmakers — and the semiconductor industry has responded strongly to ARMv8. Chips are out in production from AMD, AppliedMicro, and Cavium.
One factor playing to ARM’s benefit is the rise of microservices, the practice of dividing a service into modular pieces running independently. Microservices, in turn, lend themselves to distributed computing models where individual core performance — Intel’s gold-medal event — doesn’t matter as much.
“It lines up well with having multiple cores and multiple processors,” Linley Group analyst Jag Bolaria said during the conference’s opening keynote on Tuesday. Assuming ARM can provide adequate performance, “it kind of gives these ARM processor guys an inflection point to come in.”
“Adequate” doesn’t mean ARM chips have to keep up with Intel, and, in fact, they probably won’t. The AMD, AppliedMicro, and Cavium chips in production are at the lower end of the performance spectrum so far. But, said Bolaria, they have uses in the data center — in applications where memory bandwidth trumps CPU performance, for instance.
Getting to ‘Incredibly Boring’
ARM’s 64-bit wave has been building for a while, and Masters has used that time to move the architecture toward that “incredibly boring” point.
Since 2011, he’s been working with semiconductor vendors throughout the stages of designing and debugging the pre-commercial versions of the chips, known as A0. “You’re dealing with A0, very early chips — stuff we don’t normally touch,” he said.
This included working with hardware before the actual chips were manufactured. Masters showed a photo of an early AppliedMicro emulator, a huge card carrying seven Virtex-6 chips, the highest-end Xilinx FPGAs available at the time. Chips are expensive to build, so complex designs first get tested out on boards like these, with the programmable FPGAs filling the place of the circuitry that’s still in design.
Because ARM is represented by competing chip vendors, its success in the data center will depend on standardization, according to Masters. That’s why he helped create an Enterprise Group within the Linaro not-for-profit organization. “It brings companies together, like Cavium and Qualcomm and Broadcom, who may not, of their own volition, get in a room together and work on open source technology.”
ARM also has some marketing to do, because its name is associated with small processors. That’s led to the term “microservers,” which Masters doesn’t like.
“ARM isn’t about microservers. That’s a term that marketing organizations have coined to put them [ARM] in a box. Every single one of the ARMv8 designs I’m seeing is highly performant. They may not always win every speed race, but they are good enough.”
ARM’s allies have to change their thinking as well. Masters mentioned talking with Linux developers accustomed to embedded processing, a world where engineers want the operating system to know about tiny chip details such as clock cycles. Servers are different.
“ARM has kind of the inverse problem that Intel has. Intel is trying to go down the stack, from servers. ARM is trying to go up the stack, and when you’re going up the stack, you have to add abstractions,” he said. “Abstract the platform so it looks like what people already know and love.”
Masters will show off some of his work at Mobile World Congress later this month, demonstrating ARM-based chips doing packet processing. The chips will be aided by the Data Plane Development Kit (DPDK), a set of libraries that originated with Intel but is now being adapted to non-x86 chip architectures.