NVIDIA (NVDA +0.10%) has been building high-end graphics processors on the relatively old 28-nanometer manufacturing technology for some time. However, the advances NVIDIA has made in moving from its first 28-nanometer architecture (known as Kepler) to its second-generation architecture (known as Maxwell) have been quite impressive.
Nevertheless, I am sure NVIDIA is ready to move to more advanced technologies in order to substantially boost performance and power consumption. The company has talked about two upcoming graphics architectures, known as Pascal and Volta, but it has not mentioned which manufacturing technologies it will use to build them.
I would like to offer an educated guess based on what NVIDIA and the semiconductor foundries have said thus far.
Pascal, Volta, and Taiwan Semiconductor timelines
At the NVIDIA GPU Technology Conference this year, the company showed the following slide:
Source: NVIDIA
Note that the successor to Maxwell, known as Pascal, is expected to arrive in 2016. The next successor, Volta, is then anticipated in 2018.
Given that Taiwan Semiconductor(TSM +0.80%), NVIDIA's major foundry partner, should be in mass production on its 16-nanometer FinFET technology later this year, it makes sense that Pascal would be a 16-nanometer part. I do not believe NVIDIA will bother with a 20-nanometer GPU if 16-nanometer manufacturing technology is readily available.
Perhaps the more interesting question involves the manufacturing technology choice for Volta.
16-nanometer or 10-nanometer?
The road map above suggests Volta will launch in 2018. Furthermore, given comments from Taiwan Semiconductor that it plans to go into "risk production" on its 10-nanometer technology by the end of 2015 for a production ramp in 2017, Volta should be a 10-nanometer part.
Note also that NVIDIA is projecting a pretty significant performance-per-watt increase in moving from Pascal to Volta. Although NVIDIA has shown it can wring significant performance out of a manufacturing technology with good architectural decisions, I am sure the company is eager to get as much help as possible by adopting next-generation manufacturing technologies.
There is also another reason why a 10-nanometer Volta around mid-to-late 2017 might be viable.
10-nanometer could be a better fit for large, expensive GPUs initially
A recent note from Randy Abrams at Credit Suisse (via Barron's) said Taiwan Semiconductor plans to ramp 10-nanometer during 2017 but added that the "cost is high." For most mobile processors, which need to be low-cost, this is obviously an issue.
However, for very high-performance Tesla accelerators that NVIDIA sells for literally thousands of dollars apiece, low initial yields (which might be driving the "high cost") are probably not too much of an issue -- the margins on these chips are just insanely high.
Source: NVIDIA
NVIDIA will just need to make sure it packs in sufficient additional features and performance to command higher selling prices for chips that are more expensive to produce.
That said, I expect Apple will require significant 10-nanometer volumes for its A11 chip in 2017. This means Taiwan Semiconductor will probably be working frantically to bring yields on the technology as high as possible in order to compete for Apple's orders. This could have a positive impact on NVIDIA's Volta, should it ultimately turn out to be a 10-nanometer part.
