The onslaught against Intel’s hegemony in the market for data center servers has begun in earnest, with Microsoft leading the charge. Microsoft is going to start populating its data centers with servers based on Qualcomm’s new Centriq 2400 processor, Qualcomm and Microsoft announced today.
The collaboration is squarely in the context of an overarching Microsoft project to encourage the development of a variety of open source data center technologies. Microsoft’s Software for Open Networking in the Cloud (SONiC) project aims to have SONiC software run across a variety of Ethernet switches that can be used in data center networking from multiple vendors.
Part of the announcement with Qualcomm is that Qualcomm has joined the Open Compute Project (OCP). OCP is the home for the SONiC project. All of these announcements were made at the annual OCP conference being held this week.
Qualcomm and Microsoft said their collaboration in data center technology is going to be a long-term relationship. Qualcomm and Microsoft seem to be aligning their orbits with each other. The two said they have been working together on using the ARM architecture for years.
In December, the two said they will partner on helping Qualcomm’s ARM-based Snapdragon processors better support Microsoft’s Windows 10 operating system. That collaboration is basically about enabling mobile devices (from smartphones to tablets) to behave more like PCs.
Aligned Against Intel
The new data center server collaboration is all about grabbing market share from Intel. Intel owns well over 90 percent of the data server market, a dominance based in part on familiarity – programmers are thoroughly acquainted with Intel processors based on Intel’s x86 architecture. Intel solidified its lead in server ICs when it introduced a 48-core version of its x86-based Xeon chips, designed using newer finFET technology and processed at the 14nm node.
Power consumption in data centers is enormous, and the management of power is literally critical. The x86 architecture is practically ancient, with roots in the late ‘70s, a time when power management was a secondary consideration. If Intel has a weakness in the server IC market, it might be that the adaptation of x86 circuitry to applications where power management is paramount has been fitful. That’s the knock on the x86 architecture, anyway.
Qualcomm and ARM
Qualcomm, meanwhile, is celebrated for its expertise building chips for communications products, including smartphones, a low-power application notable for having strict power management requirements. Qualcomm has been basing its communications processors on the ARM architecture (as have many other suppliers to the smartphone market) in part because ARM was designed with low-power applications in mind.
There are several proposed advantages to using ARM, but Intel’s would-be competitors have long argued that an ARM-based alternative might thrive in the market for data center server processors in large part because of its inherent facility with power management.
For the last year or more, Qualcomm has been explicit about its intention to compete with Intel in the server market, and in December it started sampling its ARM-based Centriq 2400, also a 48-core processor, also using finFET technology. The unexpected twist was that the Centriq is being fabricated at the 10nm node, the current state of the art. It appears to be the first server processor fabricated at 10nm (other types of ICs have previously been built at 10nm).
Traditionally, each progression from one node to the next automatically buys a set of performance improvements. An IC fabricated at 10nm might be expected to outperform an IC built at the 14nm node. But as IC manufacturers push the very edge of device physics at these smallest possible design rules – 14nm, 10nm, and soon enough 7 nm – it’s not entirely clear that that rule still strictly applies.
Microsoft and Qualcomm trumpeted that their server partnership will “accelerate cloud services” – the reference explicitly being Microsoft’s Azure cloud service. But neither company provided details, including performance details. In fact, Qualcomm isn’t going to disclose any performance measures of either the Centriq 2400 or servers based on it until “a later point in the year,” a Qualcomm spokesman told SDxCentral.
Until then, there will be no way to tell if an ARM-based processor has any significant performance advantage over an x86 product. But ARM might have another advantage inasmuch as it’s an open technology.
This is an argument that is hard to parse for a lot of reasons, some of which border on the religious. One thing that open systems irrefutably do is defeat vendor lock. Anyone can use the basic technology, and anybody can modify it as needs be. By joining the OCP, Qualcomm is making it easier for any OEM to build servers based on the Centriq. That could also serve to widen the market for Centriq-based data center servers. Again, this is in the context of SONiC.
More SONiC Support
Cavium today said it will support SONiC on its XPliant Ethernet switches; at the OCP conference. Cavium is demonstrating its switch running on a 100-Gb/s Ethernet network platform built by Edgecore Networks.
Centec said it will support SONiC with its 10- Gb/s white box switches. And Mellanox announced its Spectrum switches will support SONiC; the Spectrum line supports connections at 10-Gb/s, 25-Gb/s, 40-Gb/s, 50-Gb/s and 100-Gb/s.
The potential downside of open systems is that since anyone can modify an open system, an open system environment can end up bordering on (relative) chaos. That’s the knock on the vast ARM environment anyway.
Making the situation even murkier, as it pertains to the server IC market, late last year Intel became an ARM licensee. That seems to have been a maneuver mainly to support its contract manufacturing business, but that’s part of the murk.