The chip vendors are clearing technical hurdles to deliver high-performance fabrics. For example, a competitive fabric must support moderate throughput with only a few chips while scaling to terabit speeds just by adding chips.

High-speed serializer/deserializer (serdes) design has become critical as fabric vendors cut system chip count by integrating those interfaces. A state-of-the-art serial link can move 3.125 Gbits/s-three to four times faster than Rambus and other high-speed interfaces.

Achieving that speed across a backplane is not easy. And with next-generation serdes headed to 5 Gbits/s, fabric designers must possess high-speed analog expertise to keep pace.

Handling failures is another challenge. Commercial fabrics must offer N+1 and 1:1 redundancy, allowing performance to degrade gracefully in case of one or more hardware failures.

Designing gigabit fabrics is simple enough that most equipment vendors have done it in-house. But few have the skill set to design terabit fabrics. Fewer still want to support the cost of in-house design when attractive third-party options exist.

Fabrics are available from both large and small suppliers. AMCC, Agere, IBM, Mindspeed, PMC-Sierra and Vitesse offer high-end fabrics to complement their other networking chips.

Those vendors have observed that many OEMs start with the backplane when designing a platform. The backplane is likely to have the longest life of any component in that platform. A fabric design win opens the door for selling network processors and other components.

The fabric market continues to attract startups, like Erlang, PetaSwitch, Power X, Tau Networks and ZettaCom. The startups often partner with, and may be acquired by, fabric-less networking-silicon vendors such as Broadcom or Intel.

Exploiting the tight coupling of fabric and network processors, several startups are developing both products in parallel. That group includes Internet Machines, Paion and TranSwitch (Onex).

The fabric sector has too many vendors. I expect to see the top startups acquired while others fail. Even so, this group offers many reasons for OEMs not to design another in-house switch fabric.

Linley Gwennap is founder and principal analyst of the Linley Group (www.linleygroup.com; Mountain View, Calif.).