Startup aims at DSP algorithm development

Startup Catalytic Inc. is targeting the apparently narrow niche of floating- to fixed-point model conversions for DSP software developers. But the company says it is pioneering an approach to design that combines EDA with embedded-software development.

"We view ourselves as the next stage of EDA," said Randy Allen, founder and CEO of Catalytic. "The way we see things going, many people will program and few people will design. It just doesn't make sense for as many people to make ASICs as they do right now."

At Texas Instruments Inc.'s Developer Conference, Catalytic demonstrated its technology, which sits on top of the Matlab development system and converts floating-point models to fixed-point, automating a tedious manual transition. Users then can run the floating-point models within Matlab, which doesn't natively support them.

Catalytic does not plan a formal rollout of its software until later this year, after the company makes sure the product is "rock solid," Allen said. The startup is soliciting designer feedback.

Catalytic employs 12. The startup received $6 million in first-round venture funding--from New Enterprise Associates, ITU Ventures and angel investors--which is being used primarily to expand marketing and distribution channels.

Although Catalytic speaks of taking DSP algorithms to "implementation," that does not mean RTL; it means implementation on an existing application-specific processor. Catalytic thus targets software developers, although Allen said he believes many hardware designers will move in the direction of software.

"We're coming from a higher-level language," he said. "In the case of DSP, it's Matlab. We're mapping down to the execution level, where we take advantage of the parallelism in the processor."

Floating- to fixed-point conversion is a "huge problem" when done manually, Allen said. Problems are often found at the end of the design cycle, such as discovering one has chosen the wrong processor.

"Everyone doing DSP starts with Matlab and then has to rewrite a series of models," Allen said. "Generally they write a C floating-point model, do quantization and then write a C fixed-point model. Then they take the output of that and go into assembly language and tune it. We're looking at automating that rewriting process."

Catalytic's software lets a DSP developer stay within the Matlab environment, using the Matlab language, Allen said. The user can access Matlab's numerical-analysis and visualization capabilities, although a C-language model must eventually be written.

Catalytic's FXT Toolbox lets users convert a Matlab floating-point algorithm to fixed-point, based on user-defined parameters. A GUI assists in that process. A third element accelerates Matlab simulation to handle floating- and fixed-point models more quickly.

Allen said the tool could potentially complement DSP EDA environments like CoWare Inc.'s SPW or Synopsys Inc.'s Cossap. While those tools' emphasis is hardware design, Catalytic focuses exclusively on software development.

Is Catalytic an EDA or embedded-software tool vendor? "We're between the two," Allen said. "I see those coming together."

- Richard Goering

EE Times