SAN JOSE, Calif. – Next-generation server products expected from Intel Corp. in March will accelerate the growth of 10 Gbit Ethernet and narrow opportunities at the high end for rival Advanced Micro Devices. Those are just two of many shifts rumbling through the data center, according to analysts at an event here.
“We see rather massive changes in data centers—it’s an amazingly dynamic market—and all these transitions create opportunities to grab market share,” said Bob Wheeler, senior analyst for market watcher The Linley Group (Mountain View, Calif.) and an annual data center conference
In March, Intel is expected to roll out high-end server versions of its 32 nm Sandy Bridge processors. The announcement is expected to include chips supporting four-socket servers, a high-end market Intel has not participated in recently.
“Intel’s 4P will fill a hole it had left open in the past for AMD,” said Jag Bolaria, another senior analyst at Linley Group.
The new Intel processors could outperform Interlagos, AMD’s current 32 nm high-end server chip that uses a new Bulldozer core. AMD is expected to respond with Abu Dhabi, a 32 nm server CPU using an upgraded version of the Bulldozer core called Piledriver.
Intel will roll out a so-called Romley server motherboard using its new CPUs along with its latest 10 Gbit Ethernet chip. At an earlier event, Intel showed the board
using its 40nm Twinville controller supporting the emerging 10GBase-T standard for 10G Ethernet over copper using a physical layer block from startup Aquantia.
Over the next five years, 10GBase-T is expected to become a dominant version of 10G Ethernet, replacing more expensive variants that require optical links. However, Hewlett-Packard, IBM and others are expected to roll Sandy Bridge boards using adapter cards for 10GE. They are opening the door to a wide variety of Ethernet chip vendors and variants because some do not believe 10GBase-T is ready for volume markets.
“We think 28 nm technology will turn the corner on power consumption for 10GBase-T, pushing it down to 2W or less” said Wheeler. “The 40 nm parts now reaching production consume about 3W per port, just getting to where they are viable for switches, so we expect to see ramp begin with Romley,” he added.
The best of today’s Ethernet chips used in switches are hitting power consumption levels of 1W or less, said Wheeler. Most bundle 48 10G links in a chip, although Marvell has debuted a 96-port chip, he added.
As many as 46 million 10G Ethernet switch ports could ship in 2015, most of them from merchant vendors. That’s a dramatic shift from about 10 million ports last year, mainly in OEM ASICs built for their captive use, said Bolaria.
As many as 24 million 10G Ethernet ports shipped last year, counting switches and controllers. “2011 was huge growth and we completely underestimated it,” said Bolaria.
As servers start to build in 10G links, it is driving aggregation switches to shift up to 40G.
“We are beginning to see early OEMs such as Arista and Extreme shipping 40G, and that will accelerate,” said Bolaria. “100G Ethernet is not really in the data center in the next year or so because it’s more of a transport technology for carrier optical networks given its cost is pretty high right now,” he said.
In storage networks, the move to Fibre Channel over Ethernet has been slow, limited mainly to servers links. Less than 15 percent of storage nets used FCoE in 2011, while 4-, 8- and upcoming 16-Gbit Fibre Channel options continue to dominate the field, Bolaria said.
Data center operators are trying to flatten their network topologies, adopting fat-tree techniques to increase link utilization and lower latency and cost. However, the move introduces other problems including increasing complexity of those networks.
One alternative on the horizon is the move to the emerging OpenFlow standard that defines separate network control and data planes. Hewlett-Packard recently rolled out
software supporting OpenFlow on many of its switches, joining NEC and a handful of startups supporting the technology.
Proponents say OpenFlow will ease the process of configuring and virtualizing big computer networks.
“There really are no limits to what can do with OpenFlow, although tables become more complex and not all current switches can support its more advanced features,” said Wheeler.