Advertisement

Only Thing Missing in New Super-Chips Is Imagination

Share

As Donald J. Trump will tell you, being a billionaire can be such a headache. The press hounds you, lawyers do locust impressions--and then there’s all that darn money to keep track of.

Except for the press, the lawyers and the money--this is exactly the same problem the semiconductor industry is facing. Today, semiconductor moguls can squeeze a couple of million transistors on a chip. By the end of the decade, however, they’ll be able to pack more than a billion transistors on a sliver of silicon the size of your thumbnail. So what do you do with a billion transistors?

A panel of experts at the recent International Solid State Circuits Conference--the Mecca of semiconductor engineering--ran aground trying to answer that. No bold insights, no creative sparks, just the same old chips but faster, smaller, cheaper and more memory. “There’s so much capacity with that number of transistors that it’s hard to imagine what to do with it all,” says Richard Shaffer of Technologic Partners, a technology analysis group.

Advertisement

In other words, faced with this embarrassment of transistorized riches, the semiconductor industry is pleading conceptual poverty. What this field clearly needs are visionaries; circuitpreneurs comfortable with their parvenu billion--pioneers who know how to boldly develop silicon real estate. Unfettered by prenuptial agreements or quantum electrodynamic zoning boards, these Trumps of Transistors could take their billions and build sweeping submicron vistas of computational skyscrapers and data superhighways. These wouldn’t be a mere Tokyo or Los Angeles of silicon, but entire continents of circuitry stretching farther than the average electron can see. There are fortunes to be made here.

Ready access to a billion of anything ought to change one’s view of the world, and transistors will be no exception. This submicron implosion of circuitry will transform silicon design as surely as prestressed steel, black glass and reinforced concrete resculpted urban landscapes. Beyond a certain point, quantitative differences make qualitative differences. The leap from millions to billion is past that point.

“What’s going to happen is when you have a billion transistors on a chip is that all kinds of chips--even memory chips--will have central processing units on them,” says John Moussouris, a co-founder of MIPS, the pioneering microprocessor design company who now runs Microunity Systems, a chip design firm in Palo Alto. “These CPUs will serve as traffic cops on the chip, doing real computation and coordinating communications.”

This blending of processor and memory, says Moussouris, basically means that silicon becomes a platform for huge computational systems--not just denser and denser componentry. Integrated circuits will become even more integrated. The line between processor and memory blurs. The chip becomes a vast network of processing power and data fuel. “What we’re going to see are more and more CPU cycles (capacity) whose role it will be to make all this complexity transparent,” says George Heilmeier, a senior vice president at Texas Instruments.

In the 1960s, Heilmeier points out, “the percentage of CPU cycles for direct number crunching was 100%. Today, that number is no more than 30% and dropping.” More and more CPU cycles are being dedicated to the interfaces between systems and their users. So instead of just doing calculations, the CPU is also managing the chip.

“My view of the world is that the big consumer of transistors at the chip level isn’t only memory,” Heilmeier says, “but also signal understanding, speech understanding and image understanding. The goal is for these chips to support interfaces that adapt systems to the user instead of vice versa.” That means a computer that could respond to your voice and to your handwriting.

Advertisement

The economic beauty of these billion transistors is that you get to plunge down the cost-performance curve. The per-transistor cost is dropping into a submicron abyss. Today, the same penny that buys you two sheets of toilet paper gets you a couple of hundred transistors on a chip. By 2001, it may buy you a couple of thousand.

When things become orders of magnitude cheaper, explorers may take chances on radically new architectures, ranging from neuron-like neural networks to Connection Machine-like supercomputers-on-a-chip, with tens of thousands of communicating processors. According to one Bell Labs design, a billion-transistor neural network chip could hold the equivalent of 600,000 neurons.

Technologic’s Shaffer points out that cheap transistors let you “put a lot of stupid machines together in ways that make the wiring more important than the switches.” That allows an extraordinary capability for high-speed parallel processing. Instead of CPUs, you’ll have DCPUs, or decentralized processing units.

That means, for example, you can easily manipulate three-dimensional images in real time. “It will be possible to do three-dimensional ultrasound diagnostic imaging in real time, so that the physician would be wearing a head-mounted display and looking at the inside of the patient,” says Justin Rattner, an Intel Fellow and director of technology for Intel Scientific Computers. “He can see the fetus just floating in space in three dimensions.” Today’s CPUs would take hours to assemble such imagery.

It seems as if linking these billionaire chips together gives you such overwhelming capacity that your major constraint is imagination.

Which is the problem. When you can put a billion transistors on a thumbnail, the real challenge becomes allocating your computational riches. “The U.S. semiconductor industry, by and large, does not distinguish the value of its design from the value of its manufacturing,” asserts Carver Mead, the Caltech professor who is a pioneer in both very large scale integrated circuit design and neural network technology.

Advertisement

“Last year, 80% of Intel’s profit came from its 386 (microprocessor),” Mead notes. “That ended up being less than 5% of the chips Intel manufactured. So tell me that this is a manufacturing-intensive business. . . . Tell me where the value is; is it really in the printing press or what is being printed?”

It’s not that the ability to manufacture chips is unimportant, Mead stresses, it’s that value will increasingly come from the systems one designs from these billions of transistors--not the transistors themselves.

“The semiconductor industry isn’t the semiconductor industry (as we know it), any more than the book industry is the printing industry,” Mead says. Printers focus on paper and type; publishers focus on content and distribution. The business is accelerating towards the management of intellectual property versus the management of silicon.

“It’s clear that the semiconductor guys don’t know what to do with a billion transistors,” says Mead. “But that doesn’t mean that nobody doesn’t know.”

The semiconductor status quo is clearly asking the wrong question. “What do you do with a billion transistors?” is like asking what do you do with a billion sheets of blank paper or a million feet of celluloid. The real question is: What do you do when your major constraint to using a resource is your imagination?

At a time when budget crises and environmental challenges present society with a relentless series of Hobson’s choices, we have to face up to the embarrassing reality that we don’t do so hot when we’re faced with the prospect of nearly unlimited computational riches. Whether our chips are made in Japan, Korea, Taiwan or Silicon Valley, the reality is that we have the real opportunity to become multibillionaires. It’s not yet clear who is going to do the best job of grasping that opportunity.

Advertisement
Advertisement