Celebrating 20 Years of Innovation Xilinx Staff editor@xilinx.com (1/15/04) As Xilinx marks its 20th anniversary, a Xilinx Fellow recalls the birth of programmable technology.

Twenty years in any industry is a long time; in the lightning-paced semiconductor business, it can seem like a lifetime. But for Bill Carter, the first chip designer hired by Xilinx shortly after the company was founded, programmable technology really is just entering its adolescence.

“I’m surprised at how far we’ve come in a relatively short period of time,” admits the understated Xilinx Fellow, who doubles as the unofficial company historian (which essentially means historian for an entire industry). “But we’ve got a long way to go to reach maturity, simply because the application potential for programmable technology is so vast and still largely untapped.”

Indeed, compared to its more staid silicon cousins such as microprocessors and memory – the embodiments of fixed architectures – programmable technology is still a wild-haired teenager, in some ways battling for respect and searching to find itself. But no one can deny the impact FPGAs or programmable logic devices (PLDs) – today a multi-billion-dollar market – have had on the semiconductor industry and on products that touch our lives every day.

In the context of an ever-changing electronics industry and relentless improvements in semiconductor technology, the unique benefits of programmability seem destined to be a cornerstone of innovation and progress for years to come.

Challenging a Mindset
Of course, no one could have predicted that 20 years ago. When the founding fathers of Xilinx – Bernie Vonderschmitt, Ross Freeman, and Jim Barnett – launched their oddly named start-up venture (perhaps appropriately in the Orwellian-prophesized year 1984), the semiconductor world was a vastly different place than it is today. The PC, destined to be the heavyweight champion of silicon consumption, was just emerging from Silicon Valley labs into commercial viability. The Internet was an arcane communication link for scientists and the government, wireless telephones were about the dimension of a cinder block, and Bill Gates still had to work for a living.

More importantly to the Xilinx founders, many of the engrained ways of thinking and doing business in the semiconductor industry, while seemingly permanently rooted, were in their minds becoming a bit misguided and shortsighted.

“It was Ross Freeman, really, who had the radical notion that transistors are free,” remembers Carter. “In those days, gates were precious and everyone thought, ‘fewer transistors is better.’ Ross challenged all that and saw the potential for leveraging the available real estate on chips to allow customers to customize their devices. It was contrary to everything most chip designers had learned, including me.”

Of course, Moore’s Law saw to it that eventually semiconductor designers would have more transistors than they knew what to do with.

The other tenet of the semiconductor industry that Xilinx immediately challenged was the concept of a company owning its own manufacturing capability. Fabs were then, as they are today, an expensive proposition, but also considered a closely guarded competitive advantage for chip companies.

Vonderschmitt, through past relationships and a straightforward and fair business style, managed to convince Japan’s Seiko Corp. that allowing Xilinx-designed chips to be built in their fabs was a good idea for both companies. Little did he know that this would launch a whole new approach that today is commonplace – the fabless semiconductor company.

“That part of our strategy was borne out of practicality more than anything else. We knew we didn’t have enough money to build a fab, and we certainly didn’t have enough customers to fill one,” says Carter. “Bernie was able to put together a deal that was truly win-win for both sides.”

Perhaps the only thing that would be familiar to today’s semiconductor participants was that in 1984, the industry was in a slump. Undeterred, the Xilinx founders shopped around an ambitious business plan, ultimately securing just over $4 million in funding to launch their venture. Their initial plan called for first silicon by mid-1985, and $200 million in sales by 1990 (a figure they would ultimately reach in 1993). They quickly assembled a team of software and chip design experts that shared their vision (some, such as Carter, took no small amount of convincing) and set out to change the world, transistor by transistor.

It’s the Interconnect!
Programmable devices were not a new concept in 1984, but they were anything from mainstream. Programmable logic arrays (PLAs) had been around since the 1970s, but were considered quirky, slow, and hard to use. In the early 1980s, configurable programmable array logics (PALs) had begun to emerge, offering a limited ability to implement flip-flops and look-up tables enabled by crude software tools.

Manufacturing processes were in the 2-3 micron range, and transistors still were the key to performance. PALs were seen as a replacement to small-scale integration/medium-scale integration (SSI/MSI) glue-logic parts, and slowly gained favor with the more aggressive engineering set.

But programmability remained a foreign and risky proposition for most, further compounded by attempts in the mid-1980s at “mega PALs,” which suffered from critical drawbacks in terms of power consumption and process scalability that would ultimately limit broader adoption.

Xilinx’s technical strategy was based around Freeman’s belief that for many applications, flexibility and customization would be an attractive feature if implemented correctly – perhaps only for prototyping at first, but potentially also as a replacement to more rigidly-defined custom chips. ASIC design was starting to take root, pushed along by better design tools such as simulation and other computeraided engineering (CAE) capabilities, as well as the increasing spectrum of applications for which silicon technology could be used. But Xilinx saw an opportunity to offer an even more customizable approach.

The linchpin of their innovation was the idea of programmable interconnect. In fact, the Xilinx name is drawn from this concept: The Xs at each end of Xilinx represent programmable logic blocks (or configurable logic blocks [CLBs]). The -linx represents programmable interconnect connecting the logic blocks together.

The founders took a page from printed circuit board (PCB) design (and a precursor to today’s system-on-chip [SoC] design) and envisioned arrays of custom logic blocks surrounded by a perimeter of I/Os, all of which could be assembled arbitrarily, thus overcoming the scalability issue PALs had run into (which were constrained by fixed I/Os).

The concept borrowed from the increasingly popular gate array technique, but supported the notion of post-manufacturing customization. Programming would be enabled by a set of graphical and intuitive PC-based design tools, and customers could quickly change the functionality of the chips.

Best of all, it was scalable to new manufacturing processes, a benefit perhaps even the founders may not have fully appreciated when the first chip rolled off the production lines in 1985, containing 85,000 transistors in a 2-micron process. The XC2064 was conservatively designed even for that era, containing 64 logic blocks and probably no more than 1,000 gates. However, Xilinx pioneers were determined to “only take risks with the concept, not the technology.” And the rest, as they say, is history.

Fast and Furious Progress
By the third generation of Xilinx FPGAs, the 4000 series, people were beginning to take programmable technology seriously. The XC4003 contained 440,000 transistors and was implemented in a much more leading-edge 0.7-micron process.

The performance and capacity of FPGAs were closing in on fixed architecture alternatives. In fact, FPGAs were beginning to be looked at as good vehicles for process development by manufacturers (at this point, dedicated foundries had emerged as a viable component of the semiconductor supply chain).

“By the mid 90s, we were secure enough in the concept that we could become more aggressive with how it was implemented,” Carter explains. “Plus, it turns out that FPGAs provided excellent observability into new processes, so they were being used in a way that memories had been to qualify each new generation. That put us on the leading edge, to the point now that I believe we have surpassed Moore’s Law.”

Xilinx had shipped its one-millionth device by 1989; a public offering in 1990 established the company’s sustainability. More success came quickly after that as it rode the wave of silicon proliferation. By 1995, the company was ranked as the 10th largest ASIC supplier, revenues were close to a half billion dollars, and the company had grown to more than 1,000 employees, with offices around the world.

A steady stream of innovation and increasingly competitive capabilities won Xilinx new customers across a variety of application spaces – and many new converts to the “programmable way.” The Xilinx product line expanded into highend, high-volume, and low-power variations, giving customers even more freedom of choice. Its Spartan™ family set a new price/performance standard in 1998 when it was fist introduced, while the high-end Virtex™ device became the first milliongate FPGA in that same year, enabled by aggressive use of 0.25-micron processes from Xilinx’s manufacturing partners.

By 2000, sales had topped $1 billion and Xilinx was among the first semiconductor companies to have a reachable roadmap to 90 nm processes and 300 mm wafer manufacturing.

Heading into its 20th year, Xilinx continues to set new standards, this year releasing the world’s first one-billion transistor device – a platform FPGA that not only breaks down another barrier in terms of complexity and integration, but establishes an innovative approach to bring the power and flexibility of FPGAs to a variety of applications.

The Value is in the Words
Today Xilinx ranks as the fourth largest ASIC supplier, demonstrating that programmability is not only here to stay, but is quickly becoming the customizable alternative of choice among product developers. “ASICs are already dead as far as I’m concerned,” declares Carter. “The technical barriers are gone. FPGAs are fast enough, big enough, and economically viable in high-volume applications. It’s really just a perception issue standing in the way of more widespread adoption.”

Carter feels the mindset challenges are slowly fading away as older engineers are replaced by a new generation of designers. Plus, FPGAs offer undeniable benefits in a world where products go out of style in months, an alphabet soup of standards changes at a dizzying pace, and companies require multiple variations of the same core design.

“Nothing can exploit the improvements in process technology; nothing can leverage the available transistors, like FPGAs can. The truth is that the value-add in laying out transistors, a traditional advantage for semiconductor companies, is diminishing. The way of the future will be to think about delivering intellectual property to more people, not in optimizing layouts,” Carter says.

Using a publishing analogy, Carter explains, “The value is in the words, not the ink and paper. We can supply the foundation on which to build an infinite combination of words that can be customized for any number of applications.”

When asked if there are any applications that can’t benefit from using programmable technology, Carter replies, “I really can’t think of any. Put it this way – the threshold for where FPGAs make sense is getting higher and higher.”

In fact, Carter and his colleagues at Xilinx envision FPGAs as the foundation to a whole new approach to product development, analogous to current software development approaches. Designers would use FPGAs to do weekly design builds, running tests on both hardware and software, essentially having access to a dynamic prototype. In some ways this returns FPGAs to their original roots, but the difference is that now you can “ship the prototype,” says Carter.

Beyond Silicon
The full potential of programmable technology may be beyond any single person’s wildest imagination. Exciting and unthought-of breakthroughs will come when research disciplines are crossed, such that ideas from biology or genetics, for example, are merged with silicon physics to create bold new applications.

Programmable biological systems? Not as crazy as you may think for an uninhibited adolescent who is just coming into its own.

Printable PDF version of this article with graphics.  (1/15/04) 170 KB