Revisiting Moore’s Law

Moore’s Law was predicted to end at 1 micron. It was predicted to die off twice by Gordon Moore himself. And it has vacillated between 18 and 24 months on at least a couple of occasions since it was first introduced in 1965.

From a technology perspective, there is no reason to assume it will ever die. It has gone from microns to nanometers and it can continue well into the picometer range. It can ride a shift from electrons to photons, or spintronics and tunneling electrons, or some combination of all of the above. And there will be exotic ways to cool and develop these chips and to double the number of transistors.

The pertinent questions, though, are for whom, for how much money, and why? Moore’s Law is as much an economic statement as an observation of technology. The reason to double the number of transistors on a piece of silicon, or some other more exotic substrate, is that it reduces the cost per transistor. And for some companies, that’s vital to growth.

It’s not hard to visualize more computing being done on a device with more realistic graphics for games, enough artificial intelligence to make voice commands…well, intelligent…and faster downloads and uploads of just about everything. Transistors are a key part of that, along with massive amounts of memory and throughput.

The problem is that just throwing transistors at the problem isn’t going to solve it anymore. That brute force approach of processing has run out of steam, particularly after 20nm. There are only so many levels of cache you can add before you slow down the entire system and suck the life out of a battery. And while there’s no foreseeable limit to the number of transistors on a chip, there is a very real physical limit to shrinking wires, thinning gate oxides, and a laundry list of physical effects that are the result.

Even lasers refracted through liquid can’t quite bend enough to create these chips, and by the time EUV makes its debut—if it ever really does become commercially viable—it may be too late to make a huge difference. Lithography costs are going through the roof, and the number of companies that can afford to pay those sums is shrinking.

This is hardly doom and gloom for semiconductors. Moore’s Law and semiconductor development may have run in parallel paths, but they are definitely not the same thing. The applications for chips are increasing exponentially, and the spigot is in no danger of being turned off. What will change, however, is the way chips are architected, verified and built. One size does not fit all, and just because it was done that way in the past doesn’t mean it will be done that way in the future.

Rational allocation of resources is a key part of this scheme. A powerful general-purpose processor connected to a large bank of memory may be necessary for intensely computational functions and data creation, but it probably isn’t needed for checking e-mail or even most other data consumption. Checking security on a device or surfing the Web don’t have to be cache-coherent with the data in a spreadsheet. And the audio and video on a portable device don’t have to be completely accurate or even boot in nanoseconds.

This requires some serious engineering, of course. It also will require recognition on the part of OEMs that the features being built into SoCs need to be included in the design and utilized by software. And there needs to be a feedback loop from the OEMs to the design teams about what’s important, with an ability to react more quickly rather to cobble together pre-verified and tested IP and parts. All of this will require some significant shifts in the supply chain around re-use, new packaging approaches and an understanding of how to improve performance and reduce costs without just adding more transistors and more memory.

If Moore’s Law is a road map, then the usual route has been transformed into expensive toll road with too many potholes. At some point, it simply becomes unfeasible to drive on it, and it appears that the road starts getting rougher after 20nm.