Intel will win in foundry wars

Posted by Nick Farrell

Moore's Law is eventually going to crush Intel's opposition as companies are unable to pay the huge costs involved in R&D and new process gear.

In a report, Linley Group analyst  Linley Gwennap has said "Moore's Mountain" will sort out the men from the boys and favour Intel in the long run.

Gwennap said that Moore's Law is not running out of steam, but the increasing cost of producing smaller and smaller transistors faces some companies such as Taiwan Semiconductor Manufacturing with huge bills.

Gwennap uses a case study of the fairly routine shrink from 40nm LP to 28nm LP, TSMC and other foundries rolled out high-k metal-gate (HKMG) technology, which reduced leakage power.

The HKMG process is more expensive than the standard 28nm version, but both have the same transistor density. At 20nm, HKMG becomes mandatory, because otherwise the leakage of the tiny gates would be overwhelming. Double patterning is also needed at this node, adding more cost. FinFETs, another technology to reduce power, compound the cost problem. TSMC is adding FinFETs to its 20nm process and calling it 16nm FF, but the transistor density remains unchanged, Gwennap wrote.

While other foundries have a similar plan for their first-generation FinFETs, wafer cost will more than double between 28nm LP and 16nm FF, but the transistor density will only double. This means that the cost per transistor will increase during this transition.

What we have seen in the past is that wafer costs have risen gradually while transistor counts have doubled every two years.

These trends have slashed the cost per transistor over the past 40 years, leading to chips integrating hundreds of millions of transistors but selling for less than $10.

Gwennap thinks that Nvidia, Broadcom, and Qualcomm can pass along the higher costs of more expensive transistors, while those selling into markets for cheaper devices may not be able to afford the most expensive process technology.

But the real winner will be Chipzilla which is in a better place to use its process technology advantage to deliver on Moore's Law.

Gwennap wrote that Intel seems immune to the troubles that the foundries are facing. It has been shipping PC and server processors built in its 22nm FinFET process since last year and plans to begin production using its 14nm FinFET process by the end of this year.

Thanks to the huge amount of cash Intel has chucked at process development, it can move from node to node, outpacing TSMC and others.

Of course, this is assuming that there is a continued demand for ever shrinking chips.

Gwennap and Chipzilla could become unstuck if the industry said "we don't want smaller chips and are happy with these slightly cheaper bigger ones, thanks very much". So far they have not said that, but it is possible that a pain threshold will be reached where manufacturers can't be bothered writing more cheques for shrinks that do not give them a noticeable benefit.