September 1, 1998

Where should your systems grow?

Michael Markowitz, Worldwide Editorial Director

For years, the semiconductor industry has been chasing the Holy Grail. Almost from the beginning, the industry objective has been to integrate more and more transistors onto smaller and smaller ICs. Today, silicon vendors have the capability and skills to put millions of transistors onto tiny chips. Their efforts have spawned a new breed of company that aims to relieve you of having to redesign standardized and common functions. EDA vendors have also developed tools meant to help you design million-transistor circuits and integrate your circuits with function blocks acquired elsewhere to fill these multimillion-gate-capacity ICs. All of these companies would like you to believe we're entering a brave new world of systems on silicon.

Now, I'm a bit of a cynic, so when I hear everybody talking about systems on silicon, sirens go off, and red flags go up. I start wondering if we've got an out-of-control bandwagon and if we've abandoned objectivity and reason. So let's look carefully at the situation.

Designing a multimillion-transistor ASIC requires time and money. Rather than designing at the gate level, engineers often design these minisystems using HDLs and function cores that speed the process. That practice is good. Like programming in C instead of using assembly, working at a higher level of abstraction increases efficiency and helps give engineers the productivity they need to design million-transistor circuits in months instead of years.

Then, there are the economics of ASICs. How many designs—of any size—are done correctly without any re-spins? At more than $200,000 per mask set, corrections can be expensive—even before you factor in the time it takes to create the masks and process the silicon. And when you have relatively small volumes across which you amortize the costs, re-spins can be prohibitively expensive—even when you don't need to redo the entire mask set.

In addition, fabs are expensive and need to keep pushing silicon through the lines. To do so, manufacturers generally want assurances of high production volumes. There are two key ways to ensure high production: Promise to buy lots of chips, or be willing to share your design with a bunch of companies that together offer the promise to buy lots of chips. Sharing, because it gives away your proprietary design, isn't such a good idea. There aren't many multimillion-transistor, one-manufacturer designs that ship enough pieces to keep the foundries full. Increasingly, system houses design the chips and license them to the silicon vendors on the condition that the silicon vendors will give them a head start in selling the products. Though they make their proprietary designs available to their competitors, the system companies hope to push technology fast enough to make that design obsolete by the time that head start expires. Playing this game is like walking a dangerous tightrope. If the system folks stumble, they face lower cost competitors that they've bootstrapped with their technology. And, if they succeed in obsoleting their own products before their head start expires, the system companies will find fewer silicon foundries willing to offer another head start.

So, where does that leave systems on silicon? In my mind, it suggests that the million-unit designs need to be "not-so-big." The ICs should be big enough to take advantage of our ability to squeeze more transistors onto the silicon but not so big that the application focuses on an overly narrow niche. Instead, these almost-systems-on-a-chip should be such standard products as microprocessors, memories, and programmable logic that have broad use across thousands of applications and that you can customize quickly and at a low cost with jellybean logic or programmable parts.

You can reach Worldwide Editorial Director Michael Markowitz at 1-617-558-4214, fax 1-617-558-4470, or mmarkowitz@cahners.com.

EDN Access | Feedback | Table of Contents |