And then there were three: GloFo exits 7nm chip process
Image credit: Global Foundries
The number of companies planning to continue on pushing Moore’s Law dwindles further.
Having pulled together not just the foundry interests of Chartered Semiconductor, AMD and IBM, GlobalFoundries seemed set to stay at the bleeding edge of Moore’s Law. Two of those companies remain customers for their multicore processors. But the expense of trying to get the 7nm process to production has proved too much. Now Glofo (an abbreviation that the company’s management dislikes, but which rolls off the tongue very easily) has thrown in the towel.
Technically, the work on the 7nm finFET is only suspended, but Glofo has said the move will involve layoffs and the movement of its top technologists to other projects that are meant to shore up the existing finFET and silicon-on-insulator (SOI) processes the company already operates. As a result, the number of companies planning to continue on pushing Moore’s Law dwindles further. It’s pretty much down to three. There is Intel, which has been struggling to bring its equivalent to market, though its future depends on doing so successfully to a much greater degree. Samsung has been much more aggressive in pushing its 7nm project, to support both inhouse demands and a foundry business it would like to establish. Then there is Taiwanese foundry TSMC.
When it comes to money, Intel and Samsung are better placed to continue to invest in silicon scaling for some time. However, with the lion's share of the foundry business, TSMC has a fairly clear interest in maintaining momentum. It’s been making GPUs for AMD for some time and has now said TSMC will take over production of future x86 processors aimed at the 7nm node.
Although its fab in Malta, New York, provided Glofo with the ability to sell to what one chipmaker exec calls “paranoid people in the US government”, even the military-industrial complex has the option to use Intel. For its mainframe Power processors, IBM perhaps has the knottier problem. It has not just to decide whether to buy local from another high-end processor maker or go to the Far East, but also how to deal with a world where it no longer has the same strong influence on process technology that it had with its former operations inside Glofo.
Anyone designing for the 7nm generation and looking to go beyond has to wonder: who next? It’s possible that Glofo is the last to back out of leading-edge manufacture before CMOS scaling comes to its final, physics-limited grinding halt. It’s also likely to become a lot more expensive to get processes ready and designs done before that point. If someone cracks monolithic 3D integration in the way that has happened for flash memory, that puts a lot more years on Moore’s Law and provides the last few standing with a lot of revenue.
If not, we are looking at a different industry that could coalesce to just one or two out in front. This would be a straightforward process of assimilation by standard, software-programmable devices that can be made in huge volumes. Customers will simply have to adapt those devices to their own applications through software. This would seem more viable were it not for the small problem that the indicators are pointing in the opposite direction. Software-programmable devices are generally much more power-hungry than custom chips that perform the same functions. Software only delivers better energy efficiency if it allows a dramatic change in the way functions are implemented.
One possibility is that, rather than monolithic 3D integration, the investment goes into multichip packaging. This is something Intel is pushing following its purchase of Altera. With its EMIB technology, you can barely see the join between chips that are nudged up against each other on their shared substrate. The economics are not as good as one-chip CMOS and the performance suffers because you need to restrict how many lines of interconnect go from one chip to another, but it makes it possible to customise around a common core.
The other option is for computer architectures to change much more dramatically. With the rise of machine learning, this is entirely possible. In many of these systems, absolute transistor density is not as big an attraction as it was for von Neumann processors because the interconnect becomes much more important. It’s possible that things made on ‘14nm’ will become denser in terms of actual useful functions than their Moore’s Law equivalents. That’s a big bet, though, and one the industry has rarely taken in the past.
Generally, the path of least resistance has been the lucrative choice and that points to the last supplier standing, with everyone else trying to figure out how to work with them.