"I got started in engineering in Argentina, but it was not as easy as it is in the United States," said Nemirovsky, who left that South American nation in 1982 for the University of Kansas at Lawrence, where he earned his master's degree and, later, a PhD. "That was a time of great political turmoil in my country."

It's a long way from Buenos Aires to Los Gatos, Calif., where 40 of Xstream's employees are designing a programmable CPU that must walk a very fine line. On the one hand the device is being optimized for such applications as streaming video and audio, quality of service, remote monitoring and security-all the complex technologies that next-generation communications will need. On the other hand, it must remain flexible enough to accommodate the unforeseen demands of the rapidly evolving Internet.

The business plan for XStream Logic was simply to apply learning from computer CPUs to the special needs of the Internet.

"We started out thinking about what we need to do to speed up packet processing," Nemirovsky said. "Starting from a clean sheet, what do we need to do? How will branches behave? How do we do the cache locality? All these things-we know how they behave in a generic computer processor, but the needs are different for a packet processor. We took everything that is known today from superscalar processors and we looked at it and made a decision: How does this apply to packet processing? We aimed the design for that."

Nemirovsky compared the burgeoning packet processor field with the digital signal processors of 20 years ago. Just as DSPs were optimized for a certain class of problems, but have evolved into programmable engines, the Internet requires a similar evolution in packet processors.

XStream, which describes itself as a "pre-IPO startup," will detail its basic architecture at the Microprocessor Forum in the second week of October. A product rollout is expected next year.

Nemirovsky founded XStream Logic in his living room, but it was not his first startup. After studying engineering, math and physics in Argentina, he launched his own company there before moving to the United States in the early 1980s.

Upon completing his master's, Nemirovsky shuttled back and forth between Texas and Indiana for six months. He was developing a real-time engine control IC at Delco in Kokomo, while at the same time working with Motorola Inc.'s semiconductor operation in Austin.

Then came PhD studies, after which Nemirovsky went to work at Weitek on a MIPS-architecture design and then at Toshiba Corp., where he said he developed the 64-bit MIPS 8000 design.

Moving on to Apple Computer Inc.'s Advanced Technology Center in Cupertino, Calif., Nemirovsky worked on the PowerPC architecture, and then spent 16 months at National Semiconductor Corp. on an X86 design team. He holds a string of patents, many of them in the multithreading-processor arena.

In July, he hired another computer processor veteran as XStream's chief executive and president. David Epstein worked in processor design at Data General Corp. and was among the supporting cast in the Pulitzer Prize-winning 1981 book Soul of a New Machine, by Tracy Kidder, about the development team for the Eagle super-minicomputer. More recently, Epstein worked at NexGen, where he helped develop the processor that would morph into the AMD K6.

"Rather than have multiple processors chained together to get the performance needed, our CPU will have a single processing element and a single programming model," said Epstein. "And we do this architecturally rather than through circuit design."

In this way, the XStream team is working counter to the conventional wisdom about network processors. "Virtually everyone else in the network processor space has taken the perspective that-because they think Moore's Law has run out of steam-they need to apply multiple processors to the problems presented by the Internet," said Ron Barr, the vice president of marketing. "We take the perspective that they haven't applied advanced enough technology to handle packets. We are taking concepts that are only just now emerging at the very high end of the server space and are applying them down to the embedded space, making modifications that apply to networking."

But what about the half-dozen network processors that have hit the market over the last year? Aren't they developed from the same "clean sheet" that XStream Logic's founders started with?

"They do architect with a clean sheet," Nemirovsky said. "The question is, from what viewpoint? We would argue that they architect from the system level. They are working with pieces that are not the best ones. They say, 'I have a microprocessor. I have a CAM [content-addressable memory]. I have these different pieces that I will use to architect a system solution.' And since the building block they are using is really not the best processor, they say, 'OK, let's put two together, four together, eight together.'

"What we do is start from the viewpoint of the processor. What we architect is more the fundamental piece, the process engine.

"We go much deeper than that. We cover not only contexts, not only latency. We cover everything on the processor side to make it a packet processor."

Nemirovsky said the other network processor vendors are "more like system integrators, like system-on-a-chip people, who try to figure out 'What do I need to put together to satisfy the need?' That is why there is a difference in the people at this company compared with the people at the other companies. Our expertise is in the processor architecture."

Said Epstein, "If you talk to SiTera, or Intel, or IBM, or SiByte, you'll see that they started with a building block, some core processor that was never optimized for packet processing. It was optimized for arithmetic operations. And yet, specialized functions have to happen in non-specialized hardware, he said. "We can't tell you today exactly what the algorithms might be two years from now."

Available bandwidth is catching up, Nemirovsky said, and "processor intelligence" is about to move to the upper end of the teeter-totter. As wave-division multiplexing allows existing fiber to handle many more channels, available bandwidth will explode and pressure will shift to optimizing the silicon on the pipe.

What's more, new services are emerging that require deeper looks inside the packet. While higher bandwidth allows the packet to zip across the country-or the world-faster than ever before, CPU speed will limit how quickly systems can peek inside the packet and decide what to do with it, he said.

XStream Logic is working with existing vendors of software development tools and operating systems to develop the programming environment. In addition, Epstein said the company has paid attention to debug support within the initial design.