Many observers think blades will become a vital form factor for a welter of converging computer, storage and networking systems. But just what interconnect dominates in these systems is still a matter of intense industry debate.

Ethernet had the edge in the first two generations as designers set out to pack as many cards as possible into a compact, low-power system. And it may be the long-term winner because of its broad industry support and, thus, low costs. But with the downturn, many now see a significant market window for Infiniband in high-performance systems, one of the few sectors not averse to extra costs and new technologies.

Broadcom says it will put 1000 Base X and Xaui interfaces for Gigabit and 10-Gbit Ethernet in its media-access controllers and switches as an optimal interface for backplane designs. The Irvine, Calif., company is helping define an autonegotiation protocol for the two interfaces that's expected to become part of the PICMG 3.1 standard for next-generation passive-backplane systems that use Ethernet at 1 and 10 Gbits/second.

Early next year, Broadcom will add TCP offload engines to its 10-Gbit MACs. Such engines are required to get full 10-Gbit Ethernet performance without swamping a host processor. "At the gigabit level, today's CPUs are doing just fine, but at the 10-G level TCP offload will be essential. Stay tuned for this spring," said Allen Light, a product-line manager for Broadcom.

The company claims its Fast Ethernet silicon played a prominent role in first-generation sever blades. Its Gigabit Ethernet chips using the 1000 Base X interface can be found in the current crop of blades from companies such as IBM and Dell.

Looking to the next generation, Broadcom is part of the RDMA Consortium defining remote direct-memory access capabilities for Ethernet. RDMA, first developed as part of the Infiniband specification, lets one system transfer data directly to the appropriate spot in another system's main memory without interrupting a host processor or accessing an operating system kernel. Such operations would slow down the process.

However, the RDMA feature is not expected to show up in Ethernet chips until late next year. That gap, and the extra overhead and expense of TCP processing, is nudging performance-sensitive systems being developed in the next 12 months toward Infiniband.

"We believe there's a great opportunity for Infiniband as the node interconnect," said Scott Farrand, vice president of systems engineering at RLX Technologies (Woodlands, Texas). That's in large part because Infiniband parts are now becoming available at 10 Gbits/s with native RDMA capabilities and without the overhead created by TCP.

RLX, a startup formed by former Compaq server engineers, was the first company to roll out a server blade design. A form factor essentially borrowed from telecom, blades put next-generation servers on adapter cards in a rack-mounted chassis. The first RLX system was based on 10/100-Mbit Ethernet and Transmeta CPUs. More recent designs have been built around Ethernet and low-voltage, 800-MHz and 1.2-GHz Pentium III processors.

'There's a window'

The company is expected to upgrade these designs and sees Advanced Micro Devices' Hammer processors and the Infiniband interconnect as strategic for grabbing business in the market's sweet spot: high-end server and database clusters.

Separately, at the recent launch of IBM Corp.'s BladeCenter servers based on Gigabit Ethernet, an IBM server executive said his company, too, has Infiniband on its road map.

"There's a window for Infiniband, but its long-term prospects have been damaged" by the lack of silicon support from Intel Corp., said Richard Somes, editor of the PICMG 3.0 chassis specification and director of standardization for Force Computers Inc. (Westborough, Mass.).

The PCI Industrial Manufacturers Group now has separate standards efforts for versions of the 3.x chassis using Ethernet, Infiniband, Star Fabric and PCI Express. Force is putting most of its development dollars into Ethernet as a backplane interconnect, Somes said.

As for who is adopting which interconnect, "I can't see any clear trend," said Somes. "There might be a little more thrust in Ethernet than the others, but there's still interest in Infiniband, and there's a community that wants to use Star Fabric. PCI Express is a little behind the others, but it has Intel's investment." Other options for 3.x backplane interconnects, including RapidIO, are "nothing more than rumbles," he said.

Somes said a final vote to ratify the PICMG 3.0 base specification for a standard passive-backplane system is set for Dec. 30. A draft version, 0.9, was posted for members last week. The 3.1 standard for Ethernet should be ratified during the Bus and Board conference in the third week of January, Somes added, with other interconnect standards following shortly.

A September target deadline for 3.0 ratification was delayed by a decision to change the mechanical spec for the front panel and the card insertion mechanism.