AMD today rolled the dice on a risky proposition: enthusiasm for ARM-powered servers in the data center.
The announcement fleshes out what the vendor outlined last June.
It's an idea that looks high-risk from several angles: the number of server vendors in the word has just shrunk again with Big Blue exiting the x86 server market with its sale to Lenovo, while efficiencies derived from virtualisation keeps the likes of Dell and HP awake at night.
Which is not to say that any putative AMD ARM-based server business will be virtualised out of existence in its first year: as its announcement makes clear, today's reveal of some new chip designs is a first step. Getting the ecosystem needed for a full-frontal assault on the data centre is way off for now.
Here's the potted view of what AMD has launched:
- A development platform for an upcoming ARM chip;
- The Opteron A1100, a 64-bit ARM processor built on a 28nm process, which it says will go into sampling soon;
- A microserver design that will be handed over to the Open Compute Project, as part of the “Group Hug”* common slot motherboard architecture.
The processors will come in four or eight core ARM Cortex-A57 variants, supporting 4 MB of L2 and 8 MB of L3 shared cache, memory channels that can be configured to support DDR3 or DDR4 memory with ECC at up to 1,866 MT/second; and coprocessors for cryptography and data compression.
The development kit packages the processors into a Micro-ATX form factor, along with the necessary connectors for developers to throw memory, power and communications at it, and a basic software stack of GNU/Linux, device drivers, Apache, MySQL, PHP, and Java 7 and 8.
The question is: will the world do anything but yawn?
The problem for AMD is that there's only one thing likely to drive any appetite for the work and investment needed to turn a new processor into a new server – and that's a vast improvement in power consumption.
Analyst Kevin McIsaac of IBRS told The Register the move to an ARM architecture is interesting, but he's sceptical about whether that's the same thing as “compelling”.
“For most people these days, the choice of processor is like arguing about whether you put Pirelli or Bridgestone tyres on a car,” McIsaac told Vulture South.
“If you care about it, it's either because you're a propeller-head, or you're a monster like Google or Facebook, deploying so many servers that it actually makes sense to build your own.” Those monsters, McIsaac said, are likely to be the target customers of greatest importance to AMD.
Even the most interesting aspect of the announcement, AMD's adoption of the ARM architecture, is something that fails to excite McIsaac, since while it means that the chip-shipper might have a lower-power offering than Intel, there are other ARM vendors in the world who've been doing it for longer.
McIsaac is not alone in his scepticism: VMware CEO Pat Gelsinger last year opined that x86 would still the be the data centre's CPU of choice even if ARM silicon consumed no electricity whatsoever. Gelsinger feels that the ecosystem that has grown up around x86 gives the architecture such a head start that it would be impossibly expensive to replicate, even if anyone had the incentive to do so.
But the current server ecosystem may not be quite so important at cloud scale. It's important for operating systems that run on x86 to have access to all the drivers and tools that talk to the host bus adaptors and myriad other exotica needed to assemble a conventional data centre. A Facebook-style rig requires fewer such niceties, because it's more homogenous and involves fewer classes of device.
Even if AMD ends up selling a few million ARM CPUs a year to operators of the planet's very largest data centres it will have made a tidy business for itself while at the same time taking out some of Intel's largest and highest-margin customers. ®