Multicore Packet Processing Forum

By Charlie Ashton, VP of Marketing – 6WIND

I was pleased to attend this week’s Linley Tech Data Center Conference in San Jose, CA.

During the conference, there were fascinating presentations on data center market and technology trends, on future directions in cloud architectures, on software-defined networks and on new processors announced by Cavium and Freescale. Most of these sessions were non-controversial, with the exception of the Q&A session involving Cavium and Freescale, where most of the questions seemed to be asked by the panelists themselves.

Generally, the experts who were presenting and/or were panelists were in good agreement about overall data center trends, especially about the potential CAPEX and OPEX savings from high-performance Virtual Networking Appliances. More on that subject in another 6WIND post next week……

The session that generated the most debate, though, was one in which several suppliers of non-x86 processors for data centers described the performance-per-Watt advantages of their solutions, spurring much discussion about whether those benefits would be enough to win a significant number of sockets away from standard x86 platforms.

Two ARM licensees (AppliedMicro and Calxeda) along with Tilera (“Manycore” processor architecture) outlined their plans for processors targeted at “Web  2.0 data centers”.

In this context, a “Web  2.0 Data Center” describes a data center designed for high-volume, vertical applications such as Google, Facebook, Yahoo, eBay, Amazon (not EC2), Baidu and Sina. These data centers are forecasted to comprise 25% of all servers by 2013, with a CAGR of 20+%. They are characterized, of course, by the need to support a small number of applications, each of which may run on tens of thousands of servers.

Because they run a small number of proprietary applications, Web 2.0 data centers are well-suited to the use of non-x86 processors, since the amount of software that needs to be recompiled is relatively small and the server volumes are so high. This is completely the opposite situation to enterprise data centers, which must run essentially whatever application software the user chooses to install and therefore must be based on x86 platforms.

Given the assumption that a non-x86 instruction set is not a barrier to adoption of non-x86 processors for Web 2.0 data centers, AppliedMicro, Calxeda and Tilera presented seemingly compelling advantages against Xeon platforms in terms of performance per Watt, extrapolated in at least one case to a detailed analysis of data center OPEX savings over a multi-year period. Not only would these solutions significantly improve energy efficiency, but since data centers are designed with a fixed power budget, they would extend the overall processing capacity of a data center by 2x or 3x.

The wild card in this analysis seems to be Atom, which of course retains full x86 software compatibility. At least one company, SeaMicro, has already announced Atom-based platforms for high-density servers. While these may not achieve quite the same levels of performance-per-Watt as the ARM-based and Tilera platforms, there were interesting discussions about whether they would be close enough, given the inherent advantage of being x86-compatible as well as Intel’s aggressive roadmap for regular “tick-tock” improvements.

What do you think? Do the non-x86 platforms stand a chance of winning significant market share in Web 2.0 data centers, or will Atom (at the low power end), Xeon (for high performance) and, of course, AMD as well, maintain an x86 lock in this fast-growing market segment?