EZ-Chip and Raza Micro at Gilder Telecosm 2006

NPU companies consistently make the case the market is moving into their domain and that technology is their edge, right up to the point they go out of business.

During their session at Telecosm 2006, EZ-Chip (LNOP) and Raza Microelectronics read from the same script NPU companies have been using since 1998.

1. The network is filled with changing requirements, and therefore you need a flexible hardware solution.
2. There is this new market coming that will be the turning point (this changes about every 1-2 years, current flavor of the month is Metro Ethernet, last year it was access networks)
3. NPUs put companies on the same footing as a Cisco, allowing networking hardware to be commoditized. When this takes place the NPU will be to networking what the CPU is to computing.

Reasons 1 and 2 are linked. Whenever there is a new market (and quantities are low) there is a high degree of uncertainty in what features are required. If you speak with carriers today about Metro Ethernet requirements it is clear they have no idea what functions are needed. Ask them if they prefer A or B and they will answer “yes”. Ergo, an NPU is a great fit because it provides flexibility.

The problem arises once the market takes shape. The equipment provider now understands what features are needed, which are not, and which are niche. Once these requirements are well understood they can and will be codified in silicon.

When VLANs were first implemented by Xylan in the 90’s it was done in custom ASICs. Eventually, those features were cooked right into Broadcom (BRCM), Marvel (MRVL), Vitesse (VTSS.PK), and Agere (AGR) L2 switches that sell for $1 a Gigabit port today.

If something is required to be flexible and heterogeneous it is hard to see how it is adopted as a widely accepted networking standard. Unlike computing, networking hardware converges to a common set of features due to strict interoperability requirements. In Enterprise Networking, the Venn diagram of flexibility and high volume is a null set.

I see no reason why Metro Ethernet is any different than Enterprise Ethernet. Companies like Transwitch (TXCC), Zarlink (ZL) and Cortina Systems are already building dedicated silicon for Metro Ethernet.

Another barrier to the widespread adoption of NPU’s is that no common instruction set exists among all vendors. Given the frequency that NPU companies have failed there is great hesitation on the part of equipment vendors to put an single-sourced NPU in the critical path of product success. I’ve discussed this effect in depth. (see Wintegra - Following Up on a Comment)

NPU’s play a role, but only in situations where the flexibility they bring to the table matches the cost and risk premium . If dedicated hardware can process 90% of the traffic transiting a switch or router, then the remaining 10% is shunted to the side for processing by the NPU. This radically alters the market size calculations for NPU’s because instead of having one or two per linecard you have two or four per rack. Design win numbers offer no details on the magnitude of future