Network processors enter new generation

A new generation of programmable network processors coming out this year may have a better shot than previous such chips at getting into switches and routers, allowing vendors -- as well as service providers and enterprises -- to tackle new requirements with software instead of hardware upgrades.

The first of this type of processors, commonly called NPUs (network processing units), were introduced several years ago to solve a persistent problem: Developers and users of network equipment to some extent have had to choose between speed and flexibility. Introducing more programmable chips proved harder than some in the industry had expected, but advancements in NPUs and better development tools are easing the transition, vendors and analysts say.

The new generation of NPUs is being led by major names. Intel Corp. in February charted a course for networking dominance based on two new NPUs that use its XScale architecture and are coming this year. IBM Corp. in March extended its NPU line with a chip for access routers and other network-edge devices. Motorola Inc. isn't standing still either. Its Semiconductor Products Sector, which incorporates a pioneering NPU company it acquired about two years ago, in late April introduced an edge-oriented network processor as well as a coprocessor for traffic management.

The increased integration that has powered faster PC processors is driving the new NPUs, too. Vitesse Semiconductor Corp. next month will begin shipping sample quantities of its IQ2200 NPU, which it says will be the first NPU manufactured using a .13-micron process. That process, the state of the art for commercially available processors, will let Vitesse cram more processing power into a smaller chip that consumes less power than earlier products. Intel's first NPU built with a .13-micron process will ship in sample quantities at about the same time -- in the third quarter, according to the company.

While the computer industry for decades has been able to marry ever-faster processors with a wide variety of new software, makers and users of data network gear essentially have had to choose between speed and freedom of choice. Fast forwarding and some other functions have been built in to fast ASICs (application-specific integrated circuits), while data packets that need special treatment have gotten it from general-purpose CPUs (central processing units). Those chips can be given new jobs through software upgrades, but they generally don't process packets as fast as ASICs do.

An ASIC can work with packets at "wire speed" -- sending as many bits per second as a network connection can carry -- but in most cases its operating instructions are either coded permanently into the chip or configurable only in certain limited ways. An ASIC can be made to do many different functions, but those functions are fairly well set when the chip comes out of the foundry. Therefore, making network devices do new things to packets at high speed usually requires a new part that is expensive and time-consuming to develop.

Enter NPUs, which are designed to handle packets about as fast as ASICs, but get their instructions from firmware on the chip. That firmware can be changed significantly in the field, and with relative ease, vendors say. It can also be written by a switch or router maker for its own product and loaded on to a processor developed by a third party. This saves the system maker the hard work of chip development.

In addition, NPUs for both low-speed and high-speed network interfaces, as well as successive generations of chips, can be built with the same software architecture. That means once a system vendor settles on an architecture, it can leverage that architecture and the software already written on it to build new products for years to come, vendors say.

NPUs got off to a slow start in the late 1990s. Chip makers fumbled ship dates, and system makers missed out on the hoped-for time-to-market advantage because adopting the new chips turned out to be a lot of work for both sides, said John Metz, a longtime network chip analyst at Metz International Ltd., in Harvard, Massachusetts.

"A lot of (the vendors) have gotten extremely disappointed," he said.

However, equipment vendors finally are buying into the concept, Metz said. Although chip development as expected has led to smaller and faster NPUs, he cited external factors as the biggest drivers.

-- Software simulations of the chips, which system makers can use to start designing a device before even samples of the NPU have been delivered, are becoming more accurate and reliable guides.

-- NPU vendors now provide better reference boards that equipment makers can use to make sure a chip performs and works with their software. The reference boards typically include memory, network ports, an interface to a switching fabric and other elements that are needed alongside the NPU. After the chip is qualified, the system maker makes some of those elements itself and turns to third parties for others.

In addition, NPUs increasingly are aimed at the place where they are likely to do the most good: the edge where service-provider and enterprise networks meet. That is where new and varied services such as VPNs (virtual private networks), firewall capability and QoS (quality of service) mechanisms often are implemented.

"(These functions) require a high degree of programmability and flexibility. Many were done in software in the past, but a pure software approach on a CPU had some extreme performance issues," said Bob Wheeler, an analyst at The Linley Group, in Mountain View, California. One other place to look for NPUs is in routers for the core of an enterprise, he added.

For enterprises building LANs and for customers of service providers, the new chips mean new networking products can ship earlier and start delivering new services sooner, analysts and vendors say. NPUs should help routers and switches keep up as network standards evolve over the next few years. The flexibility of NPUs also means new boxes may let creative system makers do things that were never imagined before.

"The network processor is flexible enough that people can do all kinds of things with it, and it goes beyond even what the (chip) designer intended for it," said Linley Gwennap, principal analyst at The Linley Group.

For example, an NPU could make it feasible to build a niche product, such as a network testing device, that could monitor a wide range of statistics at wire speed. Building those functions into an ASIC would require a level of investment that the product's relatively low sales volume might not justify, The Linley Group's Wheeler said.

Taking chip development off equipment makers' plates also should clear the way for more startups and greater competition, Gwennap added.

"A couple of years ago, I'd have to hire 50 or so ASIC designers and spend a couple million in foundry fees just to produce my first system," Gwennap said. "It increases the amount of innovation going on, and ultimately it will reduce prices."

SVA Networks Inc., a Plano, Texas, subsidiary of SVA Information Industry Co. Ltd., in Shanghai, turned to an NPU and a switch fabric from IBM to build inexpensive LAN switches. It saved ASIC development costs and provided flexibility, according to Jerry Pate, senior vice president of engineering at SVA.

"We're associated with a pretty large company, but our part of it is pretty much like a startup. ... From a financial standpoint, it makes it possible to even get into the market," Pate said.

SVA hopes to give its switches more features in the future, such as more security capabilities, without having to change chips. An ASIC typically is designed to carry out a certain number of functions and can even be programmed in some ways, but there are limitations.

"It does those 43 functions as fast as it can, but if you say, 'Now I want 44,' you're stuck," Pate said.

When startup Laurel Networks Inc. began developing a router for the edge of service-provider networks that can bring traffic from services such as Frame Relay and ATM (Asynchronous Transfer Mode) into a core IP infrastructure, it got to market quickly using an NPU, said Stephen Vogelsang, Laurel co-founder and vice president of marketing.

The NPU, which Laurel has not identified, delivered other benefits, too, he said. Carriers will be able to roll out new types of services to their customer base every time Laurel ships updates to its router software, probably once per quarter. In addition to downloading newly developed capabilities for their devices, carriers will be able to adjust services for particular customers by changing the software on the NPU that serves the customer's port. That means the carrier won't have to move the customer to another port or start up a new service for it.

Not all network vendors embrace the vision of an NPU future. Juniper Networks Inc., which has based its business on routers for the service-provider core, is committed to ASICs in its core products so it can keep up with speed requirements, said Scott Kriens, its chairman, president and chief executive officer. Kriens spoke before Juniper's agreement in May to acquire Unisphere Networks Inc., a deal that would significantly extend Juniper's business into the edge. That acquisition is expected to close in the third quarter of this year.

A big question mark sits next to Cisco Systems Inc., the dominant vendor of routers and data switches and a big player in both the core and the edge. The company produced its own NPU, the Parallel Express Forwarding Network Processor (PXF), in 1999 and has used it in its 7200, 10000 Series ESR (Edge Services Router) and 7600 OSR (Optical Services Router). However, Cisco spokesman Larry Yu said the company was not able to comment on its NPU strategy at this time.

That suggests the company may be close to a major decision on NPUs, perhaps whether to use third-party chips or continue going it alone, analysts said. In time, it will make sense for Cisco to use merchant NPUs, they said.

"There's nobody that could do a better job than Cisco, but we've got six or seven others out there who've already done all the work," Metz said.

Although ASICs still predominate, NPUs are on the agenda for big vendors that dominate the equipment market, Metz added.

"There isn't one of those Tier 1 companies that doesn't have a team looking at network processors. It's protection for the future," he said.

The message about NPUs appears yet to reach many enterprises. Several network engineers who attended the Networld+Interop trade show in Las Vegas were not familiar with the technology and expressed mixed opinions about its potential value.

Network devices made with NPUs might be a godsend if they reduce the need for expensive forklift upgrades of equipment, said Matt Doidge, team leader of network operations at the Aerostructures Group of Goodrich Corp., in Chula Vista, California. Mark McNeil, director of network services at Fordham University, in New York, was less enthusiastic. Useful updates already can be made to Fordham's existing ASIC-based LAN equipment through firmware changes, McNeil said.

Moreover, the NPU industry, along with the technology itself, is still young, vendors and analysts cautioned during a conference on NPUs held at Networld+Interop. Debates still swirl around issues such as the best hardware approaches and memory technologies. Chip makers' dreams of equipment vendors using their NPU architecture to build generation after generation of easily designed products could easily be dashed in the coming years, analysts and some chip makers said.

One fracture in that idea comes from the fact that the speed of network interfaces has grown much faster than the speed of conventional processors: by about 10 times in the past few years, according to Chris Hoogenboom, president, CEO and founder of NPU maker Internet Machines Corp., in Agoura Hills, California.

"The only way we come up with chips that are 10 times faster is by coming up with clever architectures, which makes it hard to preserve old architectures," Hoogenboom said.

Despite the work required to adopt a new chip, equipment vendors may change NPU suppliers in midstream if they see one leap ahead in technology or their current vendor stumble in executing its strategy, Gwennap said in an interview at the conference.

"It's a Wild West kind of thing," he said.

Nevertheless, with network requirements and standards bound to change frequently in the future, programmability in any single generation of a device could make a big difference out in the field, according to Metz. As the benefits of the technology become better known, both service providers and enterprises will flock to it.

"The enterprise guys at this point don't care that much, but they will start to care," Metz said.