Qualcomm arrived late to the ARM server party last week with its 24-core processor aimed at the server market. Lost in the hoopla, though, was a second bit of news that could be far more impactful.
ARM servers have been far more hype than reality. For years there has been talk of ARM-based servers as an alternative to more power-hungry Intel Xeon servers. And the folks over at The Platform have noted, just days before the Qualcomm news, that we've been waiting a long time.
So long, in fact, some early innovators have crashed and burned, most famously Calxeda. One of the first to produce an ARM processor for servers, it collapsed two years ago. HP's Redstone project made a big splash in 2011 but also has gone nowhere. Nvidia was working on its own ARM for the data center project called "Project Denver," but scrapped that in 2014.
AMD made its own go at ultra-dense, low power servers with its purchase of SeaMicro. SeaMicro used Intel Atom chips and AMD planned to add Atom chips, since it has its own ARM projects, but with the company bleeding red ink, it shut down SeaMicro.
So when I see people get all excited about the Qualcomm announcement, I have to wonder what events in the past few years has given you any reason for such optimism?
You can thank the virtualization companies like VMware, Microsoft, Citrix and others for killing the ARM before it could get off the ground. As it was explained to me, with a well-balanced virtualized Xeon server, with two or four sockets and anywhere from four to 18 cores, you can slice that server up into dozens and dozens of virtual machines.
Basically, virtualization on Xeon has become so good that you don't need the ARM, which is still in its early stages and only now is approaching ready for data center use. Plus, the software stack on Intel is way ahead of ARM. So why reinvent the wheel? That's why I'm not getting worked up over the Qualcomm chip as a generic server chip. It's likely going to run into the same obstacles as the others.
What could make a difference, though, is the alliance with Mellanox and Xilinx. This news was treated as an afterthought in most stories I read when it could be rather significant. In the announcement, Xilinx said it would work with Qualcomm to deliver FPGA-accelerated workloads on the new server platform. Mellanox said it will offer Ethernet and InfiniBand interconnect solutions that will be optimized for scalable server and storage infrastructures on the ARM platform.
Suddenly ARM becomes viable for Big Data, because that's where FPGA shines and Mellanox can deliver the throughput needed. FPGAs are reprogrammable processors that are growing in popularity for Big Data and analytics because they are better designed doing repetitive tasks like a search string. It can do in one cycle what a Xeon might need four or five cycles to accomplish.
It gained wider acceptance at least as a concept after Microsoft published a white paper (PDF format) on Project Catapult, an effort where Microsoft used FPGA add-in cards for the Bing search engine and saw a 40-fold increase in search speeds.
The Qualcomm server processor paired with an FPGA could have enormous Big Data potential because the CPU is not the main processor here, it's the FPGA. But the first thing the platform will need is the software, and that's been slow. Oracle announced in 2013 that it would work with ARM to optimize Java for 64-bit ARM-based servers but so far, the only company to show off Hadoop running on ARM has been AMD.
So it looks like Qualcomm is taking its ARM server chip in a different direction. Instead of doing all the simple tasks, it will be the supporting role in a Big Data platform powered by FPGA. Given the dismal failure of the prior ARM server platforms, Qualcomm was wise to at least not go down that road.