Why Maxwell will probably launch on 28nm process.

[d_brennen]

It's hugely unlikely that Maxwell will appear on both 28 and 20nm IMO. It's not just a case of turning a dial down to bake on a smaller process node, it's a lot of work and GPU architectures are tuned to their target gate size

 

[Marine-RX179]

Now I'm not saying that no Maxwell chips will be made on 28nm but I'm certain that none of the top end chips will be on 28nm.

Not if Nvidia just release entire line of product based on the GK110 (think GTX760. 770 etc with 384-bit bus and may be some core counts bump), and rebadge the GTX780 as GTX870 as next gen and delay Maxwell to one more gen later

 

[bru]

Not if Nvidia just release entire line of product based on the GK110 (think GTX760. 770 etc with 384-bit bus and may be some core counts bump), and rebadge the GTX780 as GTX870 as next gen and delay Maxwell to one more gen later

Ok I'm just being picky, in what way would your scenario be high end Maxwell chip being made on 28nm, 760/770 remade with 384 bit busses would still be kepler based chips.

 

[Rroff]

It's hugely unlikely that Maxwell will appear on both 28 and 20nm IMO. It's not just a case of turning a dial down to bake on a smaller process node, it's a lot of work and GPU architectures are tuned to their target gate size

With the changes in lithography there is no direct optical shrink path from 28nm to 20nm which also means the reverse is true to some extent taking a 20nm design for Maxwell and building it on 28nm would require quite a bit of reworking (on other node changes there has been the easy optical shrink possibilities) - as per what JediFragger said I'm pretty sure we'd have heard something about it before now if it was the case due to the work involved.

As per my previous post tho that doesn't rule out nVidia doing something on 28nm and calling it Maxwell.

 

[Kaapstad]

There is no point in launching Maxwell on 28nm

NVidia can continue using GK110 chips which will remain competitive for quite some time.

 

[Marine-RX179]

Ok I'm just being picky, in what way would your scenario be high end Maxwell chip being made on 28nm, 760/770 remade with 384 bit busses would still be kepler based chips.

I was saying Nvidia "don't have to" launch Maxwell as next gen if they are still stuck on 28nm...they can just call the GK110 based next gen cards product line Kepler II or whatever. People will bitch and whine about no Maxwell for next gen, but the rage/complain would be gone within the first month, and the it would be business as usual with the Nvidia cards flying off the shelf, Maxwell or no Maxwell.

 

[Orangey]

http://forums.anandtech.com/showthread.php?t=2335739

Depends on TSMC, they are serial bunglers.

 

[3t3P]

Where is duffman in our hour of need

 

[Dicehunter]

Wondering if I should sell my 780 now while prices are up'ish and wait for Maxwell as we don't have long to go :-/

 

[eddyr]

They don't necessarily need a highly tuned and clocked chip to launch. A low to mid tier part, a smaller design on 28nm with good yields, introducing the new arch as alternative to gcn 2.0 and getting their new features out in the wild. Replaced later on at 20nm with a revised chip(or not) under the same model name (perhaps with some go faster suffix). The latter wouldn't be the first time, been done on mobile a few times iirc (by that I mean same chip moved to smaller process later on).

 

[Diagro]

They don't necessarily need a highly tuned and clocked chip to launch. A low to mid tier part, a smaller design on 28nm with good yields, introducing the new arch as alternative to gcn 2.0 and getting their new features out in the wild. Replaced later on at 20nm with a revised chip(or not) under the same model name (perhaps with some go faster suffix). The latter wouldn't be the first time, been done on mobile a few times iirc (by that I mean same chip moved to smaller process later on).

Wont work. People seem to not realise that 20nm design wont work on 28nm.

It wont happen

 

[eddyr]

Aside from needing new masks and perhaps some layout tweaking because of transistor design/lithography there is not reason why not. I would be surprised they hadn't been prototyping the architecture on 28nm (or cheaper) during R&D. There is no doubt starting a production run has a high start up cost and requires work but that doesn't exclude it from possibility.
So while transitioning a chip design to another lithography scale is not a case of turning a dial it doesn't make transitioning the architecture impossible and if economically viable or worthwhile (short term or long term) then there is reason enough.
I am not saying there is or isn't Maxwell on 28nm just that it is not flat out impossible and they they don't have to be large die parts. The economics and strategy will determine whatever they do in the end.

 

[Orangey]

Aside from needing new masks and perhaps some layout tweaking because of transistor design/lithography there is not reason why not. I would be surprised they hadn't been prototyping the architecture on 28nm (or cheaper) during R&D. There is no doubt starting a production run has a high start up cost and requires work but that doesn't exclude it from possibility.
So while transitioning a chip design to another lithography scale is not a case of turning a dial it doesn't make transitioning the architecture impossible and if economically viable or worthwhile (short term or long term) then there is reason enough.
I am not saying there is or isn't Maxwell on 28nm just that it is not flat out impossible and they they don't have to be large die parts. The economics and strategy will determine whatever they do in the end.

Don't bother mate, he's a fanATIc. His argument will simply be "it's bad for nvidia if it's impossible so I want to believe that".

 

[Rroff]

Aside from needing new masks and perhaps some layout tweaking because of transistor design/lithography there is not reason why not. I would be surprised they hadn't been prototyping the architecture on 28nm (or cheaper) during R&D. There is no doubt starting a production run has a high start up cost and requires work but that doesn't exclude it from possibility.
So while transitioning a chip design to another lithography scale is not a case of turning a dial it doesn't make transitioning the architecture impossible and if economically viable or worthwhile (short term or long term) then there is reason enough.
I am not saying there is or isn't Maxwell on 28nm just that it is not flat out impossible and they they don't have to be large die parts. The economics and strategy will determine whatever they do in the end.

Its not flat out impossible no - but its harder than a typical node change as 28nm to 20nm lithography can't be simply scaled (in either direction) its a bit more complicated this time around than simply new masks it would require major layout tweaking - it has been done already with stuff like SRAM tho in the 28->20nm direction.

A certain amount of prototyping for Maxwell was done on Kayla but beyond that don't know too many details as to how much has been tested on 28nm and or if any of that is even relevant to porting it to 28nm. As TSMC released the design kits early I suspect they are in a better position to build it on 20nm than 28nm however.

 

[3t3P]

I love how people like to bash TSMC when they manufacture some of the most elaborately engineered precision structures known to mankind.