GlobalFoundries Announces 12LP+ FinFET Solution for Cloud and Edge AI Applications

They can't sit on their ass with 14nm forever, was a major disappointment that they stopped developing further process nodes.
Nothing great will come with TSMC being one of the few (Samsung isn't sleeping either) major players.

 

If they find ways of innovative increases in performance that are more than lithography shrinking, then I wouldn't count them out just yet. 7nm rumoured to be having some yield issues, plus what you said, there aren't major players in the field so demand will drive up cost. 12nm in comparison remains cheap and available with high yields. If performance is good enough for purpose, why pay more?

 

They kind of can - there are tons of customers that use older nodes. 50% of TSMC's revenue comes from nodes 28nm and higher. They can trail behind, adopt smaller nodes as the equipment/R&D becomes cheaper (others pave the way) and still function perfectly well.

 

https://www.techpowerup.com/259536/tsmc-to-begin-mass-production-of-5nm-chips-in-2020#comments 5nm is better though

 

How is exactly the size of the process related to the AI application?

 

It is super related almost as related are "gaming ssds" with rgb leds on em!

 

Apparently, their statement about new process says that improved performance or power efficiency. Which is what matters for most of clients.
And since I have seen TSMC's revenue split up between different manufacturing nodes, I can say that there is good business even for lithography on 10 times bigger node.

Same way as is it relevant to CPU vs GPU. Different lithography achieves better results for GPU that is switching states in large blocks with massively parallel processing.
And different tuning is better for CPU which is not massively parallel device at all.

Many AI applications would be closer to GPU. So, they could have analyzed results on different use cases and determine that this does provide better improvement for given use case than it does for others.