14nm gpu's

[Commenter]

Do you think it's likely given Nvidia's recent comments about 20nm not being a big enough incentive to go to a lower process that AMD, Nvidia are waiting before 14nm process becomes available before we see the next architecture shrink?

 

[Squilliam]

I think AMD are likely to release on 20nm next year. I don't know about nVidia though...

 

[Arun]

Do you think it's likely given Nvidia's recent comments about 20nm not being a big enough incentive to go to a lower process

Recent? Have they said anything about it in the last 2.5 years(!) since that leaked presentation? I'd be very interested if they did and you had any links
EDIT: But yes, some TSMC customers will probably go straight from 28HP(M) to 16FF+. Whether NVIDIA and/or AMD are amongst them remains to be seen...

 

[entity279]

The use of apostrophe when constructing plural should be banned at least for topic names.. oh how I hate it..

 

[Esrever]

I have my doubts of 14/16nm being as on time as they say. 20nm is probably needed for at least 1 refresh unless they want to keep on 28nm for at least 1.5 years.

 

[Squilliam]

I saw somewhere that TSMC expects to have 900,000 16nm and 1,000,000 20nm wafers completed in 2015. Based on this they could release a 16nm GPU (20nm with finfet really) in 2015.

 

[Grall]

Btw, why do TMSC call their 20nm process 16nm, and why do people let them get away with it...?

 

[Alexko]

Btw, why do TMSC call their 20nm process 16nm, and why do people let them get away with it...?

Well, it's shorter than "20nm with FinFETs". But it will get awkward when they introduce a real 16nm process with the corresponding BEOL.

 

[Squilliam]

They'll just call their real 16nm process 14nm or something...

 

[3dilettante]

Btw, why do TMSC call their 20nm process 16nm, and why do people let them get away with it...?

Process node numbers have been made up for years now by everybody.
There's no single feature that physically matches the numbers used for node names, just some sort of "effective" improvement.

 

[Alexko]

Process node numbers have been made up for years now by everybody.
There's no single feature that physically matches the numbers used for node names, just some sort of "effective" improvement.

True (from what I've read) but you could argue that the 20nm-to-16nm transition fails to match that effective improvement.

 

[3dilettante]

Because the metal layer pitch didn't improve for the foundries? It wouldn't be the first time in the last several nodes that a company took its foot off the gas in terms of wire pitch.