September 1, 1997

Putting pc boards on chips

Think of real functions - the PCI bridge, DMA controller, interrupt sequencer, UART, and other blocks - that will allow you to design 1 million-gate ASICs or 100,000-gate FPGAs as though you were going to the stockroom for parts.

Here's a revelation for you: Electronics is changing. Big deal. Faster, smaller, cheaper. So what? Actually, it is a big deal, and it has far greater effects on the way you work than the usual "faster, smaller, cheaper."

For decades, we've used the venerable pc board as the foundation upon which we build most of our products. The pc board has been a convenient, reliable, and inexpensive vehicle to mount, connect, and ultimately deliver our circuits to customers. Over time, as IC-lithographic and -fabrication processes have improved, we've integrated multiple ICs and functions into few-er packages. The reduction in the number of separate packages has al-lowed us to use a smaller board or add stuff (the technical term for more functions and capability) or both.

So, today you can build a computer, for example, by buying a CPU, memory, a graphics accelerator, VRAM, and a few other pieces of silicon. Then, you can place these chips on a pc board, connect the appropriate signal lines, and have a working computer. And it is a whole lot easier than it used to be.

But check this out: When I started designing ICs (back in the Dark Ages of NMOS logic and 6-µm feature sizes), I built my circuits using individual transistors that I hand-sized, based on Spice simulations, to achieve the necessary speed and drive requirements. Within a few years, I started using CMOS NAND and NOR gates--chosen from among presized collections of four transistors. That approach accelerated my design flow by removing me from the tedium of simulating and tweaking each transistor. The faster design process came at the expense of a small amount of design control and silicon real estate. In the years since I started picking NAND and NOR gates from my design libraries, we've further accelerated the process by automating logic design using such tools as synthesis. But we've largely stuck to using four-transistor NAND and NOR gates as the basic logic element.

Now, as the discussion among the ASIC and programmable-logic communities turns to intellectual property (IP), we are rapidly reaching the point where you can forget about NAND and NOR gates and think of real functions--the PCI bridge, DMA controller, interrupt sequencer, UART, CPU, and other blocks--that will allow you to de-sign 1 million-gate ASICs or 100,000-gate FPGAs as though you were going to the stockroom for parts.

That prospect could--and should--scare people. Dropping a microprocessor core onto an ASIC requires faith that the core's designers didn't make any mistakes and that they properly documented the core's operation and connections. That's why shifting to this approach won't happen overnight. But, like the earlier shift to NAND and NOR gates, the efficiencies of the new approach will over time create astounding improvements in design productivity that will allow you to design multimillion-gate ASICs in the time it now takes you to design multithousand-gate ASICs.

EDN Access | Feedback | Table of Contents |