One of the first: Applied Micro's X-Gene
The first server-class SoC compatible with the 64-bit ARMv8 ISA is the X-Gene from Applied Micro, and it's one example of the sort of thing we can expect from ARM partners going into the server space in the coming years. The X-Gene is intended for cloud-style deployments, where lots of small server instances will service workloads that have modest computational requirements or are more I/O-constrained.
The diagram above makes the X-Gene look relatively simple, but don't be fooled—there's lots of parallelism represented. Applied Micro says it has tailored this SoC for specific workloads, and in doing so, the company has created an awful lot of its own IP. The CPU cores, for example, are the product of an ARM ISA license. Applied Micro built its own custom CPU core, compatible with ARMv8. To address the enterprise market, the firm built in ECC support and a number of RAS and reliability features. The first X-Gene chip features eight of these cores clocked at 2.4GHz, a relatively high frequency in the ARM world. The interconnect fabric in the X-Gene is Applied Micro's own design, as well, not anything licensed from ARM. That fabric links the X-Gene's CPU cores and I/O blocks to a total of four memory controllers, twice as many as in Intel's Avoton.
What's interesting is the rationale behind this design choice: Applied Micro says applications increasingly reside completely in memory, so the X-Gene needed to have access to "lots and lots of cheap memory." The primary driver here wasn't bandwidth, but sheer RAM capacity. The result is a fairly low-power SoC that can support a ton of memory—up to 512GB, according to Applied Micro, in an eight-ranks-per-channel configuration. I doubt most X-Gene microserver modules will have half a terabyte of RAM onboard, but this possibility is still worthy of note. Intel has limited Avoton to a maximum 64GB of physical memory, perhaps in part to protect its high-margin Xeon business. The X-Gene permits configurations that might be a better fit for cloud workloads, which is exactly how ARM partners could take business away from Intel.
In a similar vein, Applied Micro has built a form of TCP acceleration into the quad 10-GigE network controllers onboard the X-Gene. This hardware can purportedly reduce the latency for TCP communication from 20-30 milliseconds to roughly five microseconds. Applied Micro says cloud providers like Facebook provision their servers on the basis of request latency, and it believes the X-Gene's TCP acceleration could allow the chip to deliver a substantially higher number of requests per second.
These things sound good in theory, but we don't yet know how they'll work in practice. Applied Micro didn't have any performance numbers of consequence to share with us yet, just a vague claim of being able to support twice as many instances per unit of power as an Intel CPU. (We don't know which one.)
Also, in a reminder that the X-Gene comes at things from a very different angle, this first chip is built on an antiquated 40-nm fabrication process. Production of the first X-Gene chips began in March, and Applied Micro is currently shipping pre-production silicon to system builders. Happily, a 28-nm X-Gene follow-up is in the works, and the first samples are scheduled for this quarter. The second X-Gene shouldn't be dramatically different from the first one, but tweaks to the CPU core are expected to bring a 15% gain in the number of instructions retired per clock cycle.
If none of this sounds good enough to persuade customers to leave the existing x86 hardware and software infrastructure behind, well, just know that Applied Micro has been working with some very influential partners. HP hasn't officially announced any products, but it has repeatedly demonstrated an X-Gene based cartridge for its modular Moonshot servers.
Also, last week, some folks from Canonical showed up at ARM's event to demo a 64-bit ARMv8 version of the Ubuntu Linux distro running on X-Gene. Canonical intends for Ubuntu 14.4 for ARMv8 to be a first-class server operating system, complete with a five-year support lifetime.
The demonstration platform consisted of a stack of 14 X-Gene servers with very little active cooling running in a centrally controlled OpenStack Icehouse environment. Christian Reis from Canonical kicked off instances of several different server applications, including MediaWiki and Hadoop, with all of the necessary components natively compiled for ARMv8. Although the process of deploying a cloud application environment didn't make for the most breathtaking real-time theater, the apps did seem to work as advertised once they were up and running.
Reis reported good progress in getting Ubuntu ported to ARMv8. He said that "99% of the main universe" is already up and going. Some of the important remaining "gaps" he identified have to do with proprietary components, like Oracle's Java virtual machine. There's also apparently work yet to be done in order to ensure ARM-based systems support low-level firmware standards for broad interoperability. UEFI support is now ready to go, but ACPI is still a work in progress, for instance.
A couple of other major software vendors, Citrix and Red Hat, were also on hand last week to signal their support for 64-bit ARMv8 servers. Both Xen and Red Hat Enterprise Linux are in development now for the 64-bit ARM ISA, and both firms appear to be committed to producing ARM versions of these core products for the long term.
We are at the beginning of something, obviously, and there's much to be done before ARM-based SoCs can truly challenge Intel for the highest-profile roles in the data center. But the foundation is being laid, brick by brick, by software and hardware engineers from a range of companies whose names are familiar and not so familiar. This week's revelation that AMD is joining the fray opens up new possibilities for ARM-based servers to challenge Xeons toe to toe, assuming the K12 core turns out reasonably well. It's hard to say exactly what happens next, but it's possible the data center will look very different five years from now, thanks to a swarm of invaders, big and small, that share almost nothing in common but an ARM license.