Get email delivery of the Cadence blog featured here
One of the big themes of the Linley Data Center Conference last week was the possibility that ARM could finally start to get traction in the data center. In the opening keynote, Linley Analysts Jag Bolaria and Bob Wheeler said that microservices and hypercovergence are creating opportunities for ARM but that they would be less than 5% of the market this year. Actually, considering that they are at pretty much zero today, that would be something that looks like the beginning of success.
In fact, with perfect timing, just before the conference opened, Google and Qualcomm announced that they would be working together. Or at least there were off-the-record reports that they would. Since Google installs over 300,000 CPUs per year, even a small percentage being ARM would start to be a large number. Other providers, in particular Amazon, install CPUs at a even higher rate.
The keynote on the second day was by Jon Masters of Red Hat, where he is the chief ARM architect. His talk was titled, How ARM Servers Can Take Over the World. He subtitled it, "or how an industry is coming together to do something disruptive." Red Hat have been involved with ARM servers since the beginning, including co-intitiating many standardization activities associated with ARMv8. He gave a brief history of their involvement:
What is driving potential growth of ARM servers? Jon pointed out four trends:
I don't think I need to tell any reader here about SoC integration.
Changing workloads refers to the fact that traditional, often proprietary, workloads are being replaced with open-source software that doesn't have the same porting challenges. In fact, a lot of software runs at even higher levels. For example, all big data runs on top of the JVM (Java Virtual Machine), meaning that it is insulated from the underlying architecture. There are already two good JVM implementations on ARM (OpenJDK and Oracle).
Migration to the cloud, with the hundreds of thousands of servers typically involved, means that it starts to be economical to invest in customized designs. In fact, Intel does this themselves already, with semi-custom Xeon designs that public clouds such as Amazon run today. Jon reckons that 20% better price/performance is enough to justify the move, although in the panel session later in the day people thought 2X was more compelling.
The fastest growth market for data center solutions is the Chinese domestic market. Just like in the US where we have Amazon, Microsoft, Google, Facebook and more, they have Baidu (search), Alibaba (shopping) and Tencent (social media). But demand exists for locally developed technologies (and government pressure may change to mandate). Since ARM is an IP company, it is easier to build a local-content ARM server than a local-content x86 server.
Jon said that there are many serious proof-of-concept projects in flight with lots of rumors around the big public clouds (such as the Google/Qualcomm one). Server silicon is now available from AppliedMicro, Broadcom, Cavium, Qualcomm, HiSilicon and AMD. A large number of operating systems, not just Red Hat, are available. Hypervisors such as Xen are available. JVMs are available, as I mentioned above. Various higher levels of the stack, such as openstack, are available.
ARM servers require standards to take off. They need to "just work" the way that Intel-based servers do today, out of the box. Unbox, plug into rack, provision via the network. In short, give the users a phenomenally "boring" experience.
To wrap up Jon talked about what not to do, and then what you should do:
Jon pointed out one great example of what you should do: the Qualcomm Software Development Platform. The engineering is phenomenal and Red Hat (and presumably Jon himself) have found them a pleasure to work with. The trick is to make using an ARMv8 server as boring and uneventful as using an Intel-based server, and ensure that whatever software you need "just runs."
We can only wait and see now.