Moore’s Law gets very warped in the sub 20nm era -

After all the talk of “innovation”, Walden (Wally) Rhines), chairman and CEO of Mentor Graphics, posed many questions about the future of semiconductors in the post 20 nanometre era at IEF2013.  He called it the Big Squeeze. Even Gordon Moore reckons that Moore’s Law is at an end. “No exponential is forever,” he said. When asked what he would like his legacy to be, his answer was “anything but Moore’s Law”.

Wally RhinesMoore’s Law might be going away, said Rhines, but not quite yet. It’s a strange law, he said, but it’s been reliable for nearly 50 years and it’s based on a law of nature – the learning curve. That’s even though Intel’s Gordon Moore changed the “law” from a year to two years and then 18 months.

Basically, it isn’t a law at all, he said. All cost reduction comes from shrinking the die, and a similar rule of thumb operates in many different sectors.

But the cost per function will continue to fall after Moore’s Law is obsolete, he said. Moore’s Law caused people to emphasise feature size. But marketing takes over in the end with people trying to warp peoples’ perceptions by saying, for example that their gate lengths were shorter than their competitors. The claims got out of control on gate length, he said.

The lies started again after everyone in the semiconductor industry decided to agree on a set of rules. The marketing started again. And the ITRS, the regulatory body,  said it was going to abandon the technology roadmap. And so now there are no rules.

The semi industry has been living in a 70 percent linear shrink for a long time – perhaps as long as fifty years, and in that framework everyone wins and that created a healthy environment. 

Marchitecture

The system worked well, he said, but now it’s changed at the 20nm node and there’s no cost/transistor crossover for the first time in history going from 28nm to 20nm.  We’re adding complexity and that’s increasing the price of a wafer to 41 percent for the 20nm generation.

Marchitecture

And if the industry does move to 450mm wafers, it doesn’t solve the problem as it did in the past. A move to 450nm will have an impact but it will be too little and it’s too late.

He suggested that cost per function will continue long after Moore’s Law is obsolete – and that could be because computing changes because of 3D stacking, bio switches, or carbon nanotubes.

So who gets squeezed? He quoted Lisa Su’s ISSCC keymore this year when she said there will be no cost/transistor crossver for the first time at the 28 to 20 nanometre change.

While, in the past, cost per wafer increases by 15-20 percent for each node, at the 16/14 nanometre stage the cost zooms to 41 percent more, with additional cost of double pattern/etch, FinFETs and the now infamous Extreme Ultra Violet delays.

A move to 450mm wafers will reduce costs, but while that will be significant it will be too little, too later. Additionally, it’s unclear which vendors will be moving to 450mm wafers.

The candidates for margins being squeezed include the food chain suppliers – wafer fabs, lithography, assembly, test and EDA software; silicon foundries, semiconductor companies and end users.  The answer is none of them can get squeezed. While Intel, he said, predicts wafer fabrication cost is the answer, a slide it’s doling out doesn’t make sense unless Intel process technology was really bad in the past. No one in the industry he’d shown this slide to said it made any kind of sense.

Marchitecture

One answer to this squeeze is for the entire industry to slow down the learning curve. But that’s probably not feasible. Growth in unit volume makes semiconductors different from other industries. While crude oil had a 1.1 percent ten year CAGR, automotive a 3.6 percent ten year CAGR, and computers an 11.5 percent ten year CAGR, semis have a staggering 72 percent ten year CAGR on the transistor front.

In fact, the insatiable appetite for transistors, doesn’t need Moore’s Law scaling, said Rhines. You can grow integrated circuits vertically by increasing density and functionality; stack dies in 2.5D or 3D; or place your faith in spintronics, silicon photonics, semisynbio or new materials, such as carbon nanotubes. DNA, he suggested, could be the ultimate hard drive.

We just don’t seem to need Moore’s Law any more. But, we suspect, Chipzilla will have a different tale to tell.