Nvidia Pascal GPU to be built on TSMC's 16nm FinFET process

[HEXUS]

Sources say that Samsung lost in its bid for part of the manufacturing action.

Read more.

[ValkyrieTsukiko]

i wonder will using 16nm cause any problems instead of using 14nm?

[shaithis]

i wonder will using 16nm cause any problems instead of using 14nm?

TSMC must be confident of their 16nm process to sign up nVidia and Apple. If they cannot meet demand economically, it would get very bad for them due to the volume they will need to produce. Apple made 200m iPhones in the last year.

[DanceswithUnix]

TSMC must be confident of their 16nm process to sign up nVidia and Apple. If they cannot meet demand economically, it would get very bad for them due to the volume they will need to produce. Apple made 200m iPhones in the last year.

Nvidia will only have enough HBM memory to send a few sample cards to journalists, so TSMC only have Apple to worry about

Seriously though, TSMC is the main player in this game, is it really surprising when they get big contracts?

[Corky34]

Do we know who AMD are using for their next gen GPUs yet?
Would I be guessing right that they maybe going with GF 14nm, if so it should make for some interesting comparisons.

[HalloweenJack]

It`ll be late then and low yields - cause TSMC have been so reliable on new process haven't they

[watercooled]

I don't remember any credible claims that Samsung would be making Nvidia's GPUs anyway, only nebulous 'Samsung making Nvidia stuff' which was far more likely to be related to their Tegra processors. But considering Qualcomm (whose mobile capacity completely dwarfs Nvidia's) may be moving a large amount of capacity to Samsung for their FinFET APs and Nvidia have decided to try suing Samsung, I'm not sure they'd be first in line for that either.

With regard to 14nm vs 16nm, really don't pay too much attention to it - they're just names of the node really and are having less and less relationship to any actual measurements. For example, Samsung/GloFo's 20nm are far more dense than Intel's 22nm, but Intel's 14nm is considerably more dense than Samsung/GloFo's 14nm. But there's more to a node than just transistor density, and indeed both TSMC and Samsung/GloFo have at least two versions each of their FinFET nodes.

[MarcelTimmer]

I hope these new chips are full dx12 compatibele.

[kalniel]

I hope these new chips are full dx12 compatibele.

Well given existing graphics cards are, there's not reason to believe the new ones would break that.

[Corky34]

They are? I thought some guy from AMD said no current GPU is fully DX12 compatible.

[kalniel]

They are? I thought some guy from AMD said no current GPU is fully DX12 compatible.

They're all compatible, you'll have no problems playing a DX12 game on the current AMD or nVidia cards.

[Corky34]

They're all compatible, you'll have no problems playing a DX12 game on the current AMD or nVidia cards.

I'm guessing you've misread the word "fully" as there's no doubting most GPUs have some support for DX12 but AFAIK no card on the market fully supports DX12, fully meaning all available feature sets at their highest tier.

[kalniel]

I'm guessing you've misread the word "fully"

No more than you misread the word 'compatible'

[Corky34]

Compatibility is very different than "fully compatible" though, the former just means able to be used with a specified piece of equipment or software without special adaptation or modification, the latter means completely or entirely; to the fullest extent.

[DanceswithUnix]

Compatibility is very different than "fully compatible" though, the former just means able to be used with a specified piece of equipment or software without special adaptation or modification, the latter means completely or entirely; to the fullest extent.

Wouldn't that be "fully accelerated"?

My shoes are fully compatible with a road in that there is no part that I am incapable of traversing, but I can't do 30mph unless I swap my shoes for a car

[watercooled]

I still don't understand why DX feature levels are getting so much attention for DX12, I don't remember people worrying or complaining so much about 11_2 (which Nvidia never supported)/11_1 vs 11_0, 10_1 vs 10_0, etc. They're just an arbitrarily selected set of features which might be useful for developers in some cases, stuck into a table by MS. Consumers really needn't worry about it, at least not to the extent it seems to be getting publicised lately.

Feature level != DX version support. Cards going back a few generations from both Nvidia and AMD support DX12 (with FL <12_0); they won't all implement the more recently specified feature levels of course, but they can still use and benefit from the newer API. 'Fully support' is a bit of an ambiguous term, the meaning of which can be mangled one way or another as it has indeed been for marketing reasons on both sides of the argument. From my perspective it appears both brands 'fully support' DX12 in the same way that the 5870 and 480 'fully support' DX11, not claiming they don't because they're 'only' FL 11_0. It's really that silly.