Network processor vendors claim more than 500 design wins, according to consulting firm The Linley Group (Mountain View, Calif.). But many of the parts haven't lived up to expectations. And for many network processor hopefuls, particularly newcomers, the proving ground of OC-192 (10-Gbit/second) networks is fast approaching.

Four network processor vendors promised OC-192 (10-Gbit/s) packet-forwarding devices by the end of 2001, and none made their goal, said John Metz, president of consulting firm Metz International. "In PC processors and the microprocessor world, people are used to 60 percent more performance per year," said Linley Gwennap, president of The Linley Group. "These 10-Gig parts are four times faster than the OC-48 (2.5-Gbit/s) parts that came out last year. Trying to make a 4x leap in a year — that's huge.

What should result is an interesting first quarter, as companies scramble to be in the first wave of OC-192 network processors. Among the contenders are Applied Micro Circuits Corp. (San Diego) and Silicon Access Networks Ltd. (Ottawa and San Jose, Calif.).

But what if no one's ready to buy? The need for 10-Gbit/s processors comes only at the very high end right now, and even that demand is diluted by continued weakness in the economy, Metz said.

Some companies like Onex Communications Corp. (Bedford, Mass.), which was recently acquired by Transwitch Corp. (Shelton, Conn.), and IBM Corp. are concentrating strictly on the OC-48 and lower speeds, backfilling that market with processors capable of services.

Another question is "scalability" — the ability of existing OC-48 processors to stretch to OC-192. Metz cited Intel's IXP1200 as a scalable chip, as Intel Corp. claims it can take the processor to 10 Gbits/s with some adjustments, like embedded memory added to its data pipelines. Motorola Inc.'s C-Port division, on the other hand, struggled to get to OC-48 (2.5 Gbits/s) and will need some changes to get to OC-192, Metz said.

On a separate front, network processors face a choice of successors to the Peripheral Component Interconnect bus, and the decision is hampering the progress of new designs, Gwennap said. In the absence of standards, companies are continuing to use PCI or have turned to proprietary buses, he said.

Gwennap expects to see a resolution, or at least a de facto leader, by the end of 2002. "Either HyperTransport is going to have enough momentum and be seen as an acceptable solution, or there's going to be enough infighting for something like Arapahoe to get in there," he said.

Memory is an elephant
Another choice unfolding in 2002 is memory.

One option is to avoid off-chip memory by embedding more RAM with the processor, as EZchip Technologies (Migdal Haemek, Israel) and Silicon Access are attempting.

Fujitsu Ltd. has garnered early support among network processors for its fast-cycle RAM (FCRAM), but Infineon Technologies AG (Munich, Germany) and Micron Technology Inc. (Boise, Idaho) are countering with reduced-latency DRAM (RLDRAM). As in the case of bus options, the product shipping first — FCRAM — has taken an early lead. "Particularly for people trying to get products out in the next six months or so, it's pretty risky to try [RLDRAM]," Gwennap said. But RLDRAM should be a competitive alternative if Micron and Infineon can get manufacturing up and keep prices down, he said.

The Rambus Inc. interface is another option, one supported in the network processor realm by Intel and Vitesse Semiconductor Corp. (Camarillo, Calif.). But the key for many packet-forwarding operations is not necessarily bandwidth but cycle time — getting the information back more quickly. For many network processor teams and OEMs, Rambus Inc.'s RDRAM interface fell short on that particular test.

Gwennap believes in Rambus' viability for network processors. But for those who don't, the choice appears to lie between FCRAM and RLDRAM. "Neither of those has the pin bandwidth of Rambus, although they're certainly superior to SDRAM," Gwennap said.

The cheap and simple option of using double-data-rate SDRAM has worked for early network processors but isn't viable at very high speeds, Gwennap said. "For a full-duplex 10-Gbit/s chip, just to buffer all the packets as they come in, you need to sustain 5 Gbytes/s, which means you need a peak [bandwidth] of 10 Gbytes/s. And in SDRAM, that takes hundreds of pins," he said. "As you get to 10 Gbits/s and beyond, SDRAM does not cut it."

Meanwhile, experimental ideas like T-RAM are waiting on the fringes. "There's all kinds of letters in front of 'RAM,' " Metz said.