June 17, 2009

The End of Moore’s Law in Five Years?

Michael Feldman

The era of the ever-shrinking transistors may be coming to a end. According to market research and consulting firm iSuppli, Moore’s Law is going to run out of money before it runs out of technology. If true, this would be bad news indeed for the IT-industrial complex, since semiconductor components (CPUs, GPUs, memory devices, etc.) depend on Moore’s Law for their roadmaps, and many businesses directly or indirectly count on the ensuing technological advancements to drive revenue growth and worker productivity.

Moore’s Law, of course, is the observation that the density of transistors on computer chips doubles approximately every two years. Intel co-founder Gordon Moore originally described the trend in a 1965 paper, at a time when transistor densities were actually doubling ever year. More importantly though, Moore observed that the cost per transistor decreased in concert with the shrinking geometries. And it is really this aspect of the model that is breaking.

In fact, it has been apparent for some time that the Moore’s Law curve is running counter to the escalating costs of semiconductor manufacturing, which are rising exponentially as process technology shrinks. This is the result of the increased cost of R&D, testing, and the construction of semiconductor fabrication facilities.

The price tag on a new 45nm fab is over a billion dollars today. AMD’s new foundry partner, Globalfoundries, is constructing a 32nm fab in New York with a budget of $4.2 billion, and Intel has already committed $7 billion to upgrade its fabs to produce 32nm chips. You have to sell a lot of chips to recoup those kinds of costs. And those are just capital expenditures.

In the iSuppli announcement, Len Jelinek, the firm’s director and chief analyst for semiconductor manufacturing, explained it thusly:

“The usable limit for semiconductor process technology will be reached when chip process geometries shrink to be smaller than 20 nanometers (nm), to 18nm nodes. At those nodes, the industry will start getting to the point where semiconductor manufacturing tools are too expensive to depreciate with volume production, i.e., their costs will be so high, that the value of their lifetime productivity can never justify it.”

The operative word is “never.” The iSuppli study predicted that in 2014, when the 18nm and 20nm process nodes are introduced, there will be no economic incentive to build volume semiconductor components below those geometries.

If true, this will tend to level the playing field for semiconductor vendors and especially fabless chip companies. For example, Intel would lose its current chip manufacturing advantage if everyone was stuck on the same process node. More importantly, if transistor size becomes a constant, much more of the burden of computer advancement will be shifted onto other elements of the ecosystem, mainly the folks that do design — chip/device, board, system, and even software.

There would also be increased pressure to abandon legacy architectures in favor of more efficient designs that need proportionally less silicon to do comparable work. Products based on x86 processors and Ethernet networks have been able to advance partly thanks to the ever-shrinking semiconductor components upon which they are based. Without that crutch, more advanced processor designs and interconnects may come to the fore.

To a certain extent, this is already occurring in the high performance computing sector. Moore’s Law is already too slow to keep up with the performance demand of HPC users, and the difference is being made up by aggregating more chips together and attaching accelerators like GPUs, Cell processors and FPGAs. That’s why interconnect technologies have become so important in HPC, which has largely abandoned Ethernet in favor of InfiniBand, and why x86 chips are playing a supporting role on some supercomputers, like the Roadrunner machine at Los Alamos National Lab and the TSUBAME super at Tokyo Tech. I imagine if Moore’s Law comes to a halt or even slows down, non-legacy architectures will become more commonplace in HPC and even generally throughout the ecosystem.

Of course, none of this may come to pass. Moore’s Law is periodically declared dead and has thus far defied its doomsayers. Additional transistor density may be achieved in other ways, such as 3D semiconductor structures. And there’s no shortage of more exotic approaches like carbon nanotubes, silicon nanowires, molecular crossbars, and spintronics. In any case, whatever happens in 2014, we’re bound to be living in interesting times.