Nvidia experimenting with HBM?

[patrickjp93]

Werent they taping out good wafers of it last quarter? Otherwise, im sure intel wouldnt mind lending some 22nm fabs to them its basically the same process anyway

They are taping out 16nm FF. 16nmFF+ is still predicted with a Q1 2016 tapeout. Xilinx is the only one optimistic about it being sooner.

[patrickjp93]

Whatever happened to the Denver arm chip on a gpu?

I just want sub 20nm HBM cards naaooooo.

Volta, 2017.

[LukaP]

They are taping out 16nm FF. 16nmFF+ is still predicted with a Q1 2016 tapeout. Xilinx is the only one optimistic about it being sooner.

Oh, do you have any technical details on the changes? or is it just another name for 16FF LP and 16FF HP?

[patrickjp93]

Oh, do you have any technical details on the changes? or is it just another name for 16FF LP and 16FF HP?

It's been forever since I've seen the papers, but 16nmFF is closer to 20nm in terms of density due to that first jump to FinFETs. 16+ actually gains ground in density, and 16+ is meant for higher power at first, though given Xilinx is jumping on it, there will be medium and lower power variants of the node most likely.

[LukaP]

It's been forever since I've seen the papers, but 16nmFF is closer to 20nm in terms of density due to that first jump to FinFETs. 16+ actually gains ground in density, and 16+ is meant for higher power at first, though given Xilinx is jumping on it, there will be medium and lower power variants of the node most likely.

ah ok. so basically they are marketing 16FF as 1x to "compete" with intel, while FF+ will actually be the real 16 node... then that node isnt coming until 2017 (thats when intel tapes out cannonlake on 10 iirc)

[patrickjp93]

ah ok. so basically they are marketing 16FF as 1x to "compete" with intel, while FF+ will actually be the real 16 node... then that node isnt coming until 2017 (thats when intel tapes out cannonlake on 10 iirc)

16nmFF+ last I knew was due for tapeout in Q1 2016, about the time Knight's Landing is being ramped for production on 14nmFF.

[LukaP]

16nmFF+ last I knew was due for tapeout in Q1 2016, about the time Knight's Landing is being ramped for production on 14nmFF.

i dont believe them. i mean intel had problems with the 1x node, and they poured millions into it. with TSMC not being really the best fab company, i dont see them doing a real 1x node in 2016

[Prysin]

,Isn't it known that pascal will have HBM ?

AMD doesn't have copyright over it ( that I know of )

 

Pascal gpu's will hopefully blow our socks off

AMD has a "1 year headstart right". since they were the ones to help fund the research alongside SK Hynix i believe.

[patrickjp93]

i dont believe them. i mean intel had problems with the 1x node, and they poured millions into it. with TSMC not being really the best fab company, i dont see them doing a real 1x node in 2016

If TSMC isn't "the best", then who is? Samsung is a 1-trick mobile SOC pony. Global Foundries has had worse transistor characteristics (leakage, capacitance, etc.) than TSMC at the same densities for more than a decade. Intel...doesn't exactly operate as an open foundry. It only opens up the old ones to outside product makers when the Intel product lifetimes run out. 32nm only recently opened up, and Intel still hasn't announced any plans for 14nm Altera silicon. Of the four big foundries, TSMC serves the most markets and tends to have very innovative designs, so much so I'd say they overcomplicate each node jump to the point they are forced into delays. They made an enormous mistake chasing 20nm planar HP for as long as they did, but honestly what else has TSMC really screwed up recently?

 

And Global Foundries seems to be squandering the foundry they just got paid to take from IBM. 22nm FDFFSOI (Fully Depleted FinFET Silicon On Insulator) has better properties than any other ~20nm process, but yet GloFo seems to not be using it as a tool to drive fresh production sales. If I were Nvidia I'd be making preparations to use it as a Maxwell refresh if both TSMC and GloFo fail to deliver on a 14/16nm HP solution. 12 cores at 4.3 GHz at only a 145W TDP, even as "RISC" architecture on the Power 8 flagship, is damn impressive.

[BogdanM]

So here we are over two years later and we have 0 mainstream HBM cards from the little green men and promises for Volta and 2018...

 

Looks like AMD will have a fantastic 2017;)