Intel 10GbE chip chucks TOE, extends reach

Intel Corp. has developed a 10GbE controller that sidesteps some of the advanced Ethernet features. So far the chip giant is opting not to support TCP offload engines (TOEs) or remote direct memory access (RDMA), which are becoming popular in the industry. It did take into its latest chip a feature being broadly debated in Ethernet circles, a pre-standard version of congestion management.

In terms of media support, Intel is now evaluating PHY chips for long-reach copper connections from two startups, believed to be Solarflare Communications Inc. and Teranetics Inc. Intel plans to ship early next year networking cards using its latest controller for 10GBase-T networks that carry signals up to 100m. The cost and power consumption of the 10GBase-T PHY chips will likely make that technology a relatively small part of the market for the next couple years, Intel said.

1st offering since 2003
The 31mm x 31mm controller consumes 4.8W average and 6.5W peak. It is Intel's first new 10Gbit part in four years. Intel was early to market with 10Gbit Ethernet, but its first controller supports the aging PCI-X bus.

"They have been losing market share to startups such as Neterion Inc. because Intel has had no 10Gbit new product since 2003," said Bob Wheeler, an analyst with market watcher The Linley Group. Neterion held 60 percent of the market for 10GbE controllers in 2006 compared to just 20 percent for Intel.

Intel's new 10Gbit part, called the 82598, is compliant with the 2.0 version of PCIe but has a maximum Express data rate of 2.5Gbps, the theoretical rate of the 1.1 version of Express.

The new Intel part is expected to appear in Caneland, a new four-socket platform based on Intel's latest quad-core Xeon processors and expected to be released later this year. Previously, Intel only shipped two-socket systems using its quad-core processors.

The new 10Gbit part will be available in September at a price Intel will not disclose.

TCP termination tradeoffs
The industry is engaged in an intense debate about the cost and performance tradeoffs of handling TCP termination on the host processor or the network controller. A handful of companies have introduced controllers with TOEs in the past several weeks. However, performance and price information about all the parts is still held close to the chest by most vendors.

Intel is taking the approach called I/O Acceleration Technology (I/OAT) of terminating TCP traffic on its processors using a variety of functions in its CPUs, chipsets and controllers to ease that job. Intel claims its controller will achieve 10Gbps throughput while only using 10 percent of the resources of a Xeon server CPU. The devices have an 8.9ms latency between applications, Intel added.

Hewlett-Packard Co., one of Intel's biggest customers, is taking a different approach, using controllers from NetXen Inc. that handle the TCP work.

"I/OAT is an example of an approach that has been used before and really just moves the processing from one part of the CPU complex to another. It doesn't completely move the processing off the CPU and on to the [networking] card," said John Gromala, a product marketing manager for HP's x86 servers. "We don't want [the job of processing] the TCP stack using up all the CPU's horsepower," he added.

HP also has been a big backer of RDMA. The feature was first developed for Infiniband and brought to Ethernet to help reduce latency in networking. Neither NetXen nor Intel support RDMA yet. Netxen is working on the feature for its next-generation part, and Intel suggested it will at least consider adding RDMA to its controllers in the future.

Intel is aware even server makers using I/OAT also use TOE controllers in some of their systems.

"There will be some environments where you get better response from TOEs," said Steve Schultz, a product marketing manager in Intel's Ethernet group. "We think there is a niche in the market for state offload and we have not ruled out supporting it in hardware," he added.

In addition, Intel has opened up for licensing a key element of its I/OAT approach, the direct memory access engine that Intel calls Quickdata. The company would not comment on whether anyone has licensed the technology.

Congestion management
Separately, the new Intel controller does incorporate a form of link-based congestion management called per-priority pause. An IEEE group is debating a standard approach to the feature, but it is not expected to have a draft of its work until next year. The feature is one of a handful seen as key in the drive to support Fibre Channel over an enhanced version of Ethernet sometimes called Convergence Enhanced Ethernet, or CEE for short.

The Intel controller does support the KX4 standard for 10GbE backplanes. That standard is expected to drive use of the chip in server blade systems, especially in 2009 and beyond. Intel projects as many as 20 percent of the 10GbE deployments by 2010 will be on the motherboard, mainly due to use of KX4 for backplanes.

Intel has implemented a new feature on the controller it calls Virtual Machine Decision Queues. It is meant to offload some of the work for creating virtual I/O on servers by creating as many as 16 virtual queues. The chip also supports 32 transmit and 64 receive logical queues per port.

The company also rolled out a new GbE product that includes some of the new features such as the virtual queuing. "Gigabit Ethernet is still where all the volume is," said Wheeler.

- Rick Merritt
EE Times