Comms chips

Net Processors Aim for Access

SAN JOSE, Calif. -- As demonstrated by announcements at this week's Network Processors Conference show, high-end network processor manufacturers are moving toward lower-priced chips targeted at the access and edge markets.

Announcements from EZchip Technologies and Applied Micro Circuits Corp. (AMCC) (Nasdaq: AMCC) today showed how both companies are taking advantage of semiconductor technology to improve network processors, but the increased power appears most likely to target systems designed for the edge of telecom networks, rather than the core. Separately, startup Azanda Network Devices is daring to push even closer to the edge, with a traffic management chip aimed at OC12 line speeds.

Special attention is being placed on the edge and access markets, and even enterprise markets, because that's where money is being spent. What's interesting is that even those that appeared to be developing chips geared for the core of telecom networks are able to morph themselves into access players.

Take the case of EZchip, which burst on the scene in 1999 bragging of a super processor to handle all 10-Gbit/s requirements in a single device. The company this week announced a faster, more powerful version of that chip, but officials also stressed that their products are finding uses in lower-end systems.

The new NP-1c is an enhanced version of the NP-1 network processor, which began shipping last spring (see EZchip Sallies Fourth). The NP-1c takes advantage of improved semiconductor technology -- transistors with line widths of 0.11 micron, rather than 0.18 micron for the NP-1. This helps increase the clock speed and add more processing elements on chip. The NP-1 funnels packets through a pipeline of these processors -- like a car going through a car wash -- so that more processors are required in order to keep up with faster line speeds.

In all, the NP-1c uses 30 percent less space than the NP-1 and boasts twice as much processing power, chief executive Eli Fruchter says.

The suitability for lower-end boxes comes in when you consider the price. The NP-1c sells for $795 and the original NP-1 for $1,175. This is lower than EZchip had first hoped, but the company had to make that tradeoff as the core-router market degenerated, Fruchter said. Moreover, EZchip adds features such as classification and table lookups without the use of additional chips such as external CAMs or SRAMs, further lowering the cost of the resulting line card, Fruchter says:

"We see ourselves today going into Layer 2 switches. We are getting into boxes that would not be using network processors if it were not for the price."

AMCC, meanwhile, is preparing its next wave of network processors with a 5-Gbit/s full-duplex chip (that is, it can process 5 Gbit/s of ingress traffic and another 5 Gbit/s of egress traffic simultaneously). The chip targets edge-minded systems such as DSLAMs as well as metro switches.

The new chip includes the traffic-management hardware that AMCC sells in its nP5700, a chip that sits between the network processor and the switch fabric. Line cards sometimes use a hardware-based traffic manager to juggle the multitude of traffic streams heading for the switch fabric (see our recent research report, Traffic Manager Chips), and many companies -- EZchip included -- have discovered the need to offer a traffic manager alongside their network processor.

Previously, AMCC didn't merge the nP5700 with a network processor because the resulting chip would be too large and would probably be too expensive for access and edge markets. The availability of smaller semiconductor geometries -- line widths of 0.13 micron -- is making the integration feasible, says Keith Morris, director of product strategy.

AMCC is referring to the technology behind its new chips as nP5, although the actual product numbers will be of the form nP3xxx. In addition to integrating the traffic manager, AMCC is adding a buffer unit where packets await to be reassembled upon arrival. For line cards dealing with hundreds of channels per port, this kind of reassembly is often handled by an ASIC or FPGA that sits in front of the network processor, Morris says.

The most dramatic thrust towards the edge has to be Azanda's. Founded as a traffic-management specialist with aspirations of OC768 (see Azanda Network Devices), Azanda scaled back its ambitions to the more realistic OC48 market earlier this year (see Azanda Flips Its Chips).

Subsequently, customers told Azanda about applications even further out from the core, in legacy OC12 networks. So Azanda is taking another step back to offer a chip, called Saber, targeting legacy OC12 networks.

"A lot of what you're doing in the access space is taking all these subscriber connections and aggregating them onto these Layer 2 VCs [virtual circuits]," says Greg Wolfson, Azanda's vice president of marketing, "and that takes a lot of shaping and scheduling."

The first Saber chip, the AZ61120-100, handles IP, ATM, and Frame Relay traffic simultaneously, and can sort traffic into four priority levels and eight service classes.

The NP-1c is slated to sample in the first quarter of 2003. AMCC expects to deliver the first of its nP5 chips in the second half of 2003. And Azanda's Saber is sampling now, with production shipments expected in the first quarter of 2003. — Craig Matsumoto, Senior Editor, Light Reading
buddha_bone 12/4/2012 | 9:30:22 PM
re: Net Processors Aim for Access This is lowwer ...?

Peter Heywood 12/4/2012 | 9:30:21 PM
re: Net Processors Aim for Access Fixed
edgecore 12/4/2012 | 9:30:19 PM
re: Net Processors Aim for Access Most of these NPU's complement their micro engines with a seperate 32 bit general purpose CPU built in for control plane duties.

Access, edge or core, what are the typical control plane tasks that are targeted to run on the NPU vs a dedicated traditional control card!


sircarrana 12/4/2012 | 9:30:17 PM
re: Net Processors Aim for Access What would be interesting to know is the trend in the co-processors with these NPU's to do deep packet look-ups
LiteRave 12/4/2012 | 9:30:05 PM
re: Net Processors Aim for Access I'm starting to wonder if the Network Processor will be relegated to a niche market. Their cost / performance model cannot seem to be justified as they continue to search for new markets to gain traction. The vendor claim was that the network processor would do to networking gear what the microprocessor did to the minicomputer - in the euphoria of the dot-com environment analogies like this were readily accepted and investments were not heavily scrutinized. The assumptions were that with the large number of MIPs relative to the fixed increments in Bandwidth and 'fat pipes' (Fast Enet, GigE, OC3 - OC192), was increasing leaving excess 'free MIPs'. While this analogy made a nice sound bite there seems to be a few differences. First, the microprocessor was smaller, and less expensive, and ultimately the physics allowed it to even be faster then a mini / mainframe. Secondly, the microprocessor did not require an architectural change to the computer system - i.e. it provided the same general purpose programmability - in the case of the PDP8 and PDP11 (popular DEC mini's) complete software compatible chips became available. Thirdly, once designed in, microprocessors were embraced to the point where next generation product was snapped up as fast as possible (as witnessed by the x86 family and 680x0 families from Intel and Motorola). The microprocessor spawned a whole new generation of companies (680x0 - Sun, Apollo, Stratus, 6502, 6809, 680x0, PPC - Apple, along with the myriad x86 PC companies including Compaq, Gateway among the largest). Some of these companies even designed their own microprocessor (e.g. Sun - Sparc). In contrast, the NP trades off performance, including speed or amount of packet processing for flexibility; requiring what was done in hardware to be implemented in software. This paradigm is counter to the past in that silicon is usually optimized and not further generalized as product matures and becomes a commodity. Examples of this include DMA, floating point accelerators, and Graphics and rendering engines to replace commonly used and intense software routines, allowing software to be applied where programmability is really needed. At computer companies for example, it is common to profile target customer applications in an attempt to optimize hardware and software specifically for those applications - uses of large multiway caches, spin-locks etc. were found to enhance system performance. Finally, while the NP alone was not responsible for the recent demise of many network startups, planning to become the next Sun or Apple relative to Cisco, companies such as Gotham Networks and Tiburon are examples of failed companies who's business plans courted and counted on the cost model of the NP. Companies that are more successful with NPs use them at the periphery of their product space (i.e. on specific high margin line cards) looking to get the feature to market and then design them out as soon as possible; not a recipe for NP success.
light-headed 12/4/2012 | 9:29:59 PM
re: Net Processors Aim for Access NPs are a great concept but the reality (despite what all the NP vendors tell you) is that they are different/difficult to program and lack development tools (learning curve), they lack some features/functionality, they are not wirespeed with normal features @ 40 or 64 byte packets or with extra features turned on. All NPs are not the same either with intel and ibm way ahead of everyone else in features, quality and performance.

additionally the only way you can compete against ASICs is if your cost is the same which it cannot be when you are paying $700-1000 for an NP or FPGA. All your competitors have to be using NPs or FPGAs or you will be much more expensive and thus non-competitive.

The winning strategy and the only one you will see succeed in the immediate future is where companies build there own NPs and use them in their products (Procket?, maybe some others). Time to market would be better in using off-the-shelf silicon but the cost and lack of control over features and performance will kill you in the long run.
woodstock 12/4/2012 | 9:29:55 PM
re: Net Processors Aim for Access >>>
Time to market would be better in using off-the- shelf silicon but the cost and lack of control over features and performance will kill you in the long run

For NP companies to succeed, they need multiple skilled employees from current reputable routing and switching companies who can redesign components consistently and repeatedly with high rate of success. Most NP company skilled employees aren't from a reputable routing and switching companies. They have to streamline and focus on fewer component NPs.. niche players.

Is it going to be a niche? You bet. Convincing a reptuable routing and switching company to use your silicon never worked in the past. They don't want NP's to take them out of business!

The winning strategy and the only one you will see succeed in the immediate future is where companies build there own NPs and use them in
their products

you must work for one of them? hope they heard you. :-)
Sign In