Options, profits, and jellybeans
Certainly, there are other routes that chip and system designers may take – 3D stacking, improved packaging, battery breakthroughs, faster memory, higher-bandwidth I/O, better architectures, snappier system interfaces, new materials such as carbon nanotubes, probabilistic or approximate computing, quantum computing – you name it. But none will show the 3500X improvement over time as was accomplished by CMOS and Dennard scaling.
"In the end, if you want to know what's going to happen," Colwell advised, "follow the money. That's not meant cynically, it's meant to remind you of how expensive this business is."
If you work for, say, Intel, and you're asking your company for billions of dollars to make a significant step forward in chip design, your company needs to decide whether to make a bet on you and your design team in the hopes that some multiple of those billions comes back in the form of profits. If your company doesn't see that profit potential, it won't make that investment – and the profit potential in small increments of performance or power improvements just isn't there, Colwell says.
And it's critical to remember, he said, that chip companies don't make the bulk of their profits from the top-of-the-line chips, but instead from the huge numbers of run-of-the-mill follow-on chips that they peddle.
"At least the way Intel did things," he said, "is that we would design a big, fat, hot flagship – we didn't call it that at the time, but that's what it was. Then we would sell a couple of million of those – and those were cool, but they were sort of tolerated. They were tolerated within Intel because they were necessary to get to the parts [Intel management] really wanted to get to," meaning the mass-market chips.
As an example, Colwell said that Intel sold three million Pentium Pro microprocessors for "about a thousand bucks" over about a year and a half. "That's not a terrible business," he said, "but in terms of Intel's business, that ain't so good." What Intel management cared about, he said, was taking the basic design of those high-end chips and translating that microarchitecture down to a smaller, higher-yield process technology, "then do it again, then do it again."
Those iterations would tweak the original design somewhat, but not much. After all, he said, the design teams that worked on those shrinking chips were smaller, had fewer resources, a shorter design cycle, and didn't have the original – expensive – engineers working with them. As a result, the shrunken chips were therefore cheaper, which helped provide higher margins for the iterative designs.
"What they were doing," he said of those secondary design teams, "was cranking out the parts that really made the profits for the place."
That iterative, shrinking process will not be possible after Moore's Law is repealed. "If you start designing a flagship at seven nanometers, you may or may not care about how many of those chips you sell at seven nanometers if you're at a place like Intel and you're designing the flagship-style machine," he said. "But what you do care about is, 'Can I proliferate this to five nanometers and make a hell of a lot of money there?' And if the answer turns out to be 'No,' suddenly my interest in seven nanometers has gone down."
The chip industry, after all, like any other industry, is about profit. Colwell recounted a conversation he once had with former Intel CEO Andy Grove, discussing profitability in the chip biz. "He said, 'Hey, look – if I could make jellybeans and make more of a profit than CPU chips, I'd do it. We're going to make a profit; we answer to the stockholders'."
And as Moore's Law arrives at the inevitable end of its exponential growth, those profits are going to be harder and harder to make. Follow the money. Or the jellybeans. ®