Comms chips

Kichips' Krazy Switch

What do you do if you're out of money in St. Louis? How about start a company and move to California?

That was the career path for switch fabric startup Kichips Inc., although it happened over the course of a few years. But now the company's finished a relocation to the southern California city of Orange, and it's preparing for third-quarter sampling of KiSS, the Ki Super Switch.

Sound Krazy? Not Kompletely. Kichips' founders are Korean engineers who were doing research at Washington University in St. Louis, courtesy of grants from Korea's Electronics and Telecommunications Research Institute (ETRI).

Under ETRI, the team was developing Asynchronous Transfer Mode (ATM) switch fabrics. They weren't building the actual chip, but creating a design that could be implemented in programmable logic devices, generic logic chips sold by the likes of Altera Corp. (Nasdaq: ALTR) and Xilinx Inc. (Nasdaq: XLNX). The designs were successful and are running in live Korean telco networks, Kichips president Jay Hoge says.

But ETRI cut the funding in early 1999, which meant decision time for the engineers. Part of the team split off to work on intrusion-detection systems, remaining in St. Louis, while the core switch-fabric guys stayed together to form Kichips. They recruited a few outsiders such as Hoge, garnered $3 million in angel financing, and got to work.

The 12 employees began Kichips' relocation to Orange last year, and the company also maintains a design office in Korea. "The quality of Korean engineers is extremely high, and frankly, they work really cheap -- so we can get talented guys at half the price we could get in Silicon Valley," Hoge says.

As an extra (and admittedly justified) dig against Silicon Valley, Hoge notes that Orange has more reasonable housing costs, which made it easier to convince the founders to move from the affordable environs of St. Louis.

Kichips' switch fabric uses a shared-memory architecture, a contrast to the crossbar-based designs that most startups are pursuing (see Erlang Preps Switch Fabric Push). The heart of the chip is a block of static random-access memory (SRAM), a cheap type of memory that's used all over the computing world. Kichips adds the necessary logic to communicate with line cards, and the result is a device consisting of just 1 million gates, compared with several million for more elaborate switch fabrics.

Other switch fabrics require a second chip that resides on every line card and adds relevant info to the packet header. Kichips uses one too, but because its header information is just 10 bits long, the company gets away with a cheap PLD in place of an elaborate, costly line-interface chip, Hoge says.

Kichips adds the extra kicker of allowing variable-length packets, meaning the data doesn't have to be segmented into the fixed-size cells used by most switch fabrics.

It all sounds nice, but Kichips has to contend with a crowded market (see Packet Switch Chips) and will be behind the curve a bit. "There's 20 other guys out there that have silicon," says John Metz, principal analyst of consultancy Metz International Ltd.

But Hoge contends that Kichips has something those other architectures don't: economy. He says the company's chip design is simple enough that another $8 million is enough to carry Kichips to profitability.

By contrast, he contends other switch-fabric startups will need $50 million to $75 million just to get their chips to market -- and he emphasizes Metz's point that the market's competitive.

"The stuff is being commoditized. Guys who got $120 million in VC money, there's no way in hell they can pay that back," Hoge says.

The KiSS fabric is also likely to be inexpensive -- critical, because Kichips is setting sights on companies such as D-Link Systems Inc. that are expert in providing low-cost networking. Many of these box makers want to expand more deeply into enterprise-infrastructure territory by providing features such as 10-Gbit/s Ethernet uplinks, Hoge says.

"Thirty percent of the market in Ethernet is made up of 'Other.' We want to help the Others get on the pie charts."

Kichips' next step will be to do a classification engine, again striving for low cost. Most classifiers are built to ride alongside network processors, but Kichips is again targeting a lower-cost market -- specifically the PowerPC, a general-purpose microprocessor that's usable as a network processor. In fact, IBM Corp. (NYSE: IBM) recently confirmed it will stop marketing its specialized network processor, concentrating instead on offering PowerPCs to that market.

— Craig Matsumoto, Senior Editor, Light Reading

Be the first to post a comment regarding this story.
Sign In