The slow turning of the tide away from x86 dominance of computing took another step yesterday as longtime second fiddle to Intel, AMD, announced that it would be licensing an ARM core to produce an ARM based Opteron chip for the 2014 market. Even a couple years ago, the idea that a stalwart x86 company like AMD would cross over to work with ARM would have been highly controversial. But let’s take a look at what’s driving AMD’s push into lower power servers.
The big picture
It’s projected that 2012 will see the first decline in PC sales since 2001, except this time around it’s not being fed by a global recession but by a shift in consumer behavior toward mobile computing. AMD along with Intel have surely missed the mobile computing boat (anyone want to play the Where’s Waldo game of mobile computing also known as spot the smartphone with an Intel chip in it).
I’m mentioning these behavioral shifts in computing because AMD’s move to grab a larger portion of the data center market by producing an ARM chip for that market is based on the realization that the one bright spot in the computing market is data center growth. AMD focused on this exact point, exemplified by a slide it showed that projected global data center IP traffic growing at a 31 percent compound annual growth rate (CAGR) through 2016, reaching a whopping 6.6 zettabytes (1 million petabytes is a zettabyte).
The growth of data center traffic is highly levered to the growth of mobile computing so if AMD can’t get its chips into the next Apple or Android phone, it’s next best shot is to go after the data centers that will support all that mobile use. More and more software developers that serve mobile are also wondering if there are benefits to having the back end processors at data centers mirror the ARM processors in the vast majority of mobile devices.
So the question becomes how does AMD differentiate itself in that data center market, now that Intel has a stronghold with its Xeon chip sets and that Intel is gunning for the low power portion of the market with its own low power 64 bit “Centerton” Atom chips due out this year. Perhaps more problematic for AMD is that anyone with some cash can license from ARM and build a chip.
Senior Semiconductor Analyst for Avian Securities Arnab Chanda noted to me, “There are multiple people trying to get into this market. Applied Micro. Marvell. Calxeda. Nvidia. Cavium. We think Qualcomm and Apple are working on 64 bit ARM as well.”
Chanda said he believes that the cloud server market will begin to include ARM based servers, but added, “It’s a zero billion dollar market here…right now it doesn’t really exist. There’s a chance this could be a real bloodbath with so many players. Intel is not just going to cede the market. They have a dual core Atom server chip out there.”
Going beyond the chip
There’s limited value in licensing an ARM core and building a chip around it. The major requirements for putting hundreds of nodes on a server rack and having them work together revolves around the need to have lots of computing tasks running in parallel rather than single threaded computing. Which requires management software and fabric technology to link and control cores.
AMD knew this, which was why it threw down $334 million to acquire low power server maker SeaMicro earlier this year. SeaMicro was actually using Intel Atom chips in its builds, going after the low power server market without asking its customers to recompile any code.
But, more importantly, SeaMicro’s CEO Andrew Feldman was always clear that his fabric technology, which SeaMicro developed with tens of millions of VC, was chip agnostic. What Feldman cared about was having a technology that could network and manage the overall need for massive parallel computing, whether it was done on x86 or ARM or even for a Linux operating system.
ARM server startup Calxeda, which recently raised an additional $55 million knows this as well. It’s VP of Marketing Karl Freund recently told me, “The core is a commodity. We focus on everything around the core, the management, the I/O, the networking and the fabric. We have as many software engineers as we have hardware engineers.”
The software to enable ARM cores to work is the true hurdle to cross both in reducing the barrier to recompiling code, which requires support of the ARM developer community and folks like Redhat which is supportive of the AMD-ARM deal.
And for the companies like AMD which are starting to encroach on the OEM server market by integrating more IP like SeaMicro’s Freedom Fabric into its chip builds, it is all about optimizing the management of hundreds of so called “wimpy cores.” So that at the end of the day those cores are working seamlessly together and hopefully adding up to some not so wimpy power savings.