After Moore's Law

Transistor density has been doubling roughly every eighteen months like clockwork for 40 years now. In a nutshell, that's Moore's Law.

Not only has chip density increased, but because of increasing density and an expanding market, prices for the same amount of memory or CPU power continually decrease. (Okay, there are glitches in memory pricing, but over time the trend is more for less.)

Bad news: it can't go on forever.

In fact, as Roy Brander observes in The end of Moore's Law, within ten years we may run up against the laws of physics. At a certain point electrons will freely wander between circuits, so circuits must remain above that minimum size or they will not work reliably.

Exactly where that point and our ability to manufacture circuits end up is probably somewhere in the 0.10 to 0.03 micron range. (Today's newest designs use 0.18 micron elements, so we can still expect at least four times and possibly twenty to thirty times higher density.)

Even if we make breakthroughs in optical computing or quantum computing, we will eventually hit a limit.

Corollaries to Moore's Law

Moore's Law specifically predicts transistor density. Corollaries derived from Moore's Law include:

• cost per transistor drops correspond to density increases
• transistor speed increases as density increases
• CPUs roughly double in power every 18 months

Power is intangible - it isn't just MHz. Sometimes a chip with a slower clock speed can vastly outperform one with a higher speed. Mac users saw this when the 25 MHz 68040 consistently outperformed the 40 MHz 68030 in the "wicked fast" Mac IIfx. We saw it again when the Power Mac G3/233 held its own against the vastly more expensive Power Mac 9600 with a 350 MHz 604e processor.

And we've claimed the same advantage for the G3 (and now the G4) compared with Intel's Pentium line.

But let's first look at CPU speed. If today's processes can eek out a 733 MHz processor (Intel's just-announced Pentium III), and we can expect to reach at least 0.09 micron chips, we could see 3 GHz processors in about three years. Over the following 4-5 years, technology permitting, we would see dies using 0.03 micron traces and achieving speeds of 20-25 GHz.

Those are simply mind-boggling numbers, but if engineers can maintain Moore's Law for another 7-8 years, that's what we'll see. And then we'll finally have all the processor designs on relatively even footing - they'll all run at the same clock speed, so the design efficiencies will clearly show themselves.

The Real World

But it won't necessarily happen. As Brander points out in his article, at a certain point the cost of building the factory (or fab) becomes prohibitive, more than any company or consortium can afford. That could happen within five years.

For the sake of argument, and because it's a very nice round number, let's say the industry achieves 10 GHz in about five years - and can go no further.

What's an industry based on constantly faster, bigger, more power, and lower costs going to do? How can you make a better computer when Moore's Law finally runs into the brick wall of the laws of physics?

You innovate. One way to increase computing power is to put more on the chip: more instruction pipelines, more registers, more specialized circuitry (such as AltiVec). By taking a processor and giving it more execution units, it can do more operations in the same amount of time.

You also improve the instruction set. Some commands on the G4 run twice as fast as on the G3. Chip designers will keep pushing the envelope to optimize each instruction a CPU processes.

And you use a bigger on-chip cache, so the CPU spends less time waiting for data on the relatively slow memory bus. Back in the Mac IIci era, a 32 KB level 2 (L2) cache was huge; today we're seeing Intel and Motorola designs that support a 2 MB L2 cache.

And finally, you abandon the CPU.

The SETI@home Supercomputer

The world's most powerful supercomputer doesn't exist as a single computer. It's the collection of well over a million computers across the globe working on the SETI@home project. At over 6 TeraFLOPs/sec., it puts even the G4 to shame.

Some of today's computers, including some older Macs and clones, support multiple processors. Not a central processing unit (CPU), but multiple processing units (MPUs). Rumors are that Apple will introduce MPU G4 machines next year, possibly at the Macworld Expo in January 2000. Based on efficiencies of the G4 design (see Macintosh CPUs, Part 2), a dual processor G4 system could be more powerful than two single Power Mac G4 computers.

Whether that actually happens or not, adding a second processor will roughly double the computer's performance. With Apple, Motorola, and IBM losing the MHz race to Intel and AMD, a quad-G4/500 system could claim 2 GHz overall speed - and the four AltiVec processors would put it in the 16 GFLOPs range!

Down the road, when we reach the GHz wall, when the engineers have tweaked every instruction, and when each processor has reached a limit on how many pipelines it can reasonably handle, we'll still be able to make our computers faster by strapping on more processors.

Already with OS 7.x Apple had patches (created by Daystar) to support two to four processors. That's vastly improved in OS 9, despite the fact that Apple hasn't produced a dual-processor system since 1997. And the support improves even further when you combine OS X with the expected multi-CPU G4 systems.

Conclusion

Transistor density will stop growing, probably within the next five or ten years. CPU power will probably peak at about the same time, as engineers use every trick in the book to give their processor the edge over the competition.

After that, the only ways to grow a faster computer will be by clustering multiple computers (as SETI@home does), using multiple processors, or doing both.

Further Reading

• Computer scientists poised for revolution on tiny scale, NY Times, 11/1/1999 (registration required). Molecular circuits could be 1/100 size of those used in today's chips.