Earlier this month, Facebook announced a new motherboard/daughtercard design it dubbed “Group Hug.” Thanks to innovative work from the Open Compute Platform, the new daughtercard allows CPUs from ARM, Intel, or AMD to be plugged into a single motherboard. At least, that’s the plan – for now, the hardware interconnects are still in the design phase, as is the framework required to manage disparate CPUs from multiple vendors.
There are, in other words, huge questions left to settle. Adding different daughtercards to the same server to optimise CPU usage is a great idea, but writing software that understands how to manage the various assets is still a massive task. In the rush to herald the advent of a bold new era for server commoditisation and the advent of ARM servers, some significant factors are being ignored.
Here’s the biggest: Contrary to what others have said, mobile SoCs will not “remake the server world.” Wired attempts to draw a parallel between Facebook’s experiments with turning off chip cache and the idea of replacing “brawny” CPU cores with “wimpy” ones. But it’s not that simple.
Take a look at ARM’s next-generation IP block for connecting up to 16 processors, their caches, and a variety of additional system devices, as compared to the CCI-400 that’s currently shipping for the Cortex-A15.
The CoreLink 400 runs at half CPU speed. Here’s the server variant:
The CCN-504 doesn’t just connect more chips; it ties in up to 16MB of L3 cache, dual-channel memory (with support for ECC), support for DDR4, 10gigE to 40gige, and a host of other devices. It’s an order of magnitude more complex than anything ARM has shipped previously and there’s a reason it’ll support both the Cortex-A15 and upcoming chips on ARM’s 64-bit extension, ARMv8.
IP blocks like this are as essential to any ARM-based server as the underlying processor and you’ll never see them in a smartphone. Smartphones don’t need 16 processors with 8-16MB of L3 and support for 10gigE, DPI, and SATA. It turns out that once you start adding server-class features to low-end processors, the end products’ power consumption increasingly resemble each other regardless of the CPU architecture you start from.
Intel (and to some extent, AMD) will fight the trend towards commoditised hardware by attempting to compete directly in performance per Watt, and by offering additional features that ARM chips don’t yet have. It’s true that ARM poses a threat to Intel’s server business, but drawing a line from cell phones to servers by way of an impressive box by Facebook ignores the real challenges ARM vendors face in trying to break into the server market.
What’s the future of the data centre look like? Complex and evolving. ARM CPUs are going to have a part to play, but creating a full server ecosystem around these products and achieving mass-market penetration is going to take years. Facebook’s Group Hug platform could kneecap traditional server vendors, but it only threatens Intel if it can’t build cheap processors that offer better performance per Watt than its competition. At the recent Open Compute summit, all of the vendors in question were confident that their own solutions would prove to be the best option for powering next-generation servers.
AMD has the fruits of its SeaMicro acquisition, new 64-bit ARMv8 processors in the works, and next-generation 28nm chips based on its Jaguar core launching this year, though there’s no information on whether or not Kabini and Temash will show up in servers. Intel has its own Atom server products and will refresh those chips with 22nm processors based on the first quad-core, out-of-order Atom that debuts later in 2013. ARM, of course, has server vendors like Calxeda as well as companies like X-Gene, which plans to ship its own 64-bit ARMv8 design by the second half of this year.
The winner will be decided by manufacturing, design, and scalability as much as CPU architecture. Historically, Intel has had a better handle on those issues than any other vendor on the planet. ARM may force Intel to innovate, but the chances of a wholesale takeover are exceedingly small.