Vol 19, Issue 17

April 25, 2005

Let's Retire Moore's Law at 40

By Kevin Krewell

Intel trotted out Gordon Moore one more time in an attempt to prove that Intel sets the agenda for the industry. But Moore's Law isn't really about Gordon Moore any more; it's about setting the agenda and pace for the semiconductor industry. Intel isn't the first company to build dual-core processors, but when Intel announces that its roadmap is going completely to dual- and multicore processors, the press takes notice.

Moore's law is supposed to be this unstoppable locomotive in semiconductors, yet in embedded processors many companies are getting off the train at 130nm, 180nm, and even 250nm. I was recently briefed by ARM about its DesignStart Program, which gives free early design access to the ARM7TDMI core in 180nm, and was surprised to find out that about half of ARM's new design starts are still in 180nm. Some designs, especially in microcontrollers, will likely stay at older process generations for lower wafer and mask cost, more-robust ESD characteristics, and resistance to higher voltages, as well as because the designs don't need to go any faster or get any smaller.

Tom Halfhill, in his Viewpoint article in 2004 (see MPR 12/13/04-02, "The Mythology of Moore's Law"), talked about the various (mis)interpretations of Moore's law and how it changed over time. It hasn't been a completely stable "law" and has been subject to revisions and interpretation. I recommend subscribers read Halfhill's article to understand the intent behind Gordon Moore's projections.

When Gordon Moore first made the prediction, he was at Fairchild, and the products were MSI logic devices. He last updated his predictions in 1975, when the microprocessor industry hadn't yet taken off. Intel was still primarily in the DRAM and general LSI logic business. Since that time, Intel applied Moore's Law to microprocessors and widened the interpretation to include not just mainstream processors but also the ultralarge-die-size Itanium processors.

What Moore's Law Has Wrought

Moore long ago lost control of his law, and Intel, apparently, believes it owns the intellectual-property rights to it and can do with it as Intel pleases. Which is why we need to retire Moore's law. The semiconductor industry needs to follow the process roadmap that it finds economically advantageous to the industry and not dance to a preprogrammed tune. Each new process generation brings great expenses, greater risks, and, in some cases, diminishing performance returns. On its website, Intel says, "Our R&D investment and silicon expertise support unique Intel breakthroughs that will enable us to drive Moore's law well into the future and deliver more exciting capabilities into our technologies." The question the industry needs to answer is this: Do you want Intel (and the IDMs that follow Intel's path) driving the industry?

AMD, Intel, and server-processor vendors do not focus on making smaller and cheaper chips but rather on putting more features into roughly similar-size die. Some embedded designers are using advanced semiconductor processes to put multiple processor cores and other functions onto one die, which can enable new products that were not possible just a few years ago. That is the progress we associate with Moore's law. But this progress relies on money, and each new process generation requires even more money. Intel has plenty of money to spend, and it uses that money to fund its semiconductor research and fabs. But those fabs and the company's latest-generation process haven't always produced winners. Intel's 90nm process produced a Pentium 4 that had much higher leakage currents than were produced using the preceding 130nm process and didn't reach the 4GHz goal the company set for it. Despite converting most of its products to 90nm, Intel couldn't clearly beat AMD (except in mobile), which continues to ship a considerable number of 130nm products. My point is this: Being first to a new process isn't a guarantee of success, and using an older, but appropriate, process does not signal failure.

Gordon Moore retired; shouldn't his "law" get to retire, too?

Spring Processor Forum Reminder

It's spring, and the first buds of new processors are starting to appear (single-core and multicore versions). They will be in full bloom at the Spring Processor Forum, May 16–19, at the Doubletree Hotel in San Jose. Make your plans now to be there and enjoy the new crop. And we encourage you to consider submitting presentations on your new processor or processor-related technology for our Fall Processor Forum (October 24–27) and Processor Forum Taiwan (November 9–10).