HyWire Ltd. has gone back to the drawing board with the search-engine chip announced last year, pushing commercial production into 2004.

HyWire never shipped the original chip for revenues, but it did get samples into customers' hands -- only to find that the chip was too expensive to be viable.

"We were sampling, but then they came back and said the price points weren't good enough to compete with TCAM [ternary content-addressable memory]," says James Goodhart, HyWire director of marketing.

TCAMs are the preferred building blocks for search engines, chips that contain large tables that store routing or classification information. In an IP router, for example, TCAMs would store the "next-hop" choices for different routing destinations.

The fact that HyWire's prices couldn't compete with TCAMs posed a big problem, because HyWire's chips are intended to replace TCAMs. After a bit of hand wringing, the company has come back with a new design, announced today, that promises the same level of performance but with a lower price tag.

TCAMs are relatively expensive and they've got a reputation for using lots of power. As a result, companies such as HyWire are investigating alternatives that would use more conventional memory chips, relying on software algorithms to do the searching (see Search Engines Face Software Challenge).

HyWire's original part -- the one that was too expensive -- used on-chip static random-access memory (SRAM). The new design will rely on off-chip dynamic RAM (DRAM), which is cheaper and takes up less space, although DRAMs do use more power than SRAMs, Goodhart says.

The change was harder than it sounds. HyWire had to add "a new layer of algorithms to use DRAM and avoid bus-bandwidth problems," CEO Moshe Stark says, referring to the fact that it takes time for data to pass to and from an external DRAM chip.

HyWire avoids that problem by accessing the external DRAM only once per query. The HyWire chip, called a search engine manager (SEM), contains most of the information common to all searches. It pinpoints the location of information in the DRAM, then accesses the DRAM once to grab the desired entry. The idea is to avoid a back-and-forth situation, where one table entry prods the router or switch into doing another search for an extra level of detail. A multiple-stage search would take too long using external DRAM.

The trick was to partition the information so that only 1 percent of the stored data resides on the SEM, keeping that chip small and cheap. The rest can be spread among the plentiful -- and even cheaper -- DRAMs.

HyWire has gotten the SEM to work in a field programmable gate array (FPGA), which will be demonstrated at next week's Network Processors Conference. FPGAs are frequently used for this kind of test-run chip; they do the job, but they're larger and and sometimes slower than the real thing. HyWire's actual SEM chip won't be ready until at least March 2004, according to Goodhart.

HyWire claims the SEM will run 400 million searches per second, compared with the 266 million being quoted for most high-end search engines.

Like any startup, HyWire faces the problem of cracking the ranks of incumbent players. That's a particular challenge for search engines, because the market is small and has a thin customer base, making it tough to dislodge market leaders Integrated Device Technology Inc. (IDT) (Nasdaq: IDTI) and Cypress Semiconductor Corp. (NYSE: CY) (see Mosaid Cans CAMs).

Separately, HyWire yesterday (Monday) announced a low-end search engine called HyCognito, targeting 1-Gbit/s speeds. This will be sold not as a physical device, but as a semiconductor "core," a design meant to be incorporated into other chip makers' chips (see HyWire Starts New Search Engine).

— Craig Matsumoto, Senior Editor, Light Reading