AMD Greenland GPU 2016 to feature 14nm node

[CoolaxGaming]

 

Seems like AMD wants to catch up on the performance per watt market. According to kitguru, the core on the Greenland gpu will be a successor to Fiji XT with memory and node improvements.

A jump from 28nm to 14nm will be really nice indeed. The node being used will be Samsung node.

But why not go to 20nm this year? Well seems like AMD decided to skip it because the yields were too low.

 

 

 

Greenland is the successor of Fiji, and we don’t think it will be radically new core. The main goal for Greenland is to bring more performance per watt to AMD GPUs in 2016.

AMD's highly anticipated Fiji GPU is a 28nm design, and so are the GM200 based GPUs that ended up in Nvidia's latest Titan X cards. The Geforce GTX 980 based on GM204 GPU is also manufactured in 28nm manufacturing node at TSMC. Back in November, we said there was simply no place for 20nm GPUs in 2015. The yields are horrible and this is one of the reasons why neither Nvidia nor AMD went for TSMC's 20nm node.

 

 

Our early information is that Greenland will be made in GlobalFoundries / Samsung's 14nm manufacturing process and it is not clear if AMD will stay loyal to TSMC for its GPUs and use its 16nm manufacturing process. Korea Times thinks that Nvidia might go after Samsung's 14nm instead of 16nm TSMC manufacturing process for its next generation GPUs, but we are not convinced.

 

 

Source -

http://www.fudzilla.com/news/graphics/37584-amd-greenland-2016-is-a-14nm-gpu

[Interfectorem]

So they're just shrinking FIji XT down to 14nm?

[CoolaxGaming]

So they're just shrinking FIji XT down to 14nm?

Well and some memory improvements.

[Interfectorem]

Well and some memory improvements.

That seems kind of dumb.

[Comessy]

28nm to 14nm is a massive jump

 

This may save AMD

[CoolaxGaming]

That seems kind of dumb.

Well you cant really blame em. If the Fiji XT holds out against Pascal, the 490x might actually be good

[BobbyG2]

I would imagine there will be at least be a card with more SPs than Fiji, just maybe sharing the same architecture.

[Opcode]

So they're just shrinking FIji XT down to 14nm?

Jumping from 28nm down to 14nm FF is a huge node improvement. It could be why AMD decided to call their next line of GPUs "Arctic Islands". As Fiji itself is suppose to see substantial power improvements as it is.

[Sakkura]

28nm to 14nm is a massive jump

 

This may save AMD

 

It won't make any difference, because Nvidia will almost certainly have access to the same process node at the same time.

[CoolaxGaming]

It won't make any difference, because Nvidia will almost certainly have access to the same process node at the same time.

Are you sure?

Nvidia and Samsung aint the best of budies right now...

[Kamina]

Now it all comes to architecture with pescal. 

[Sakkura]

Are you sure?

Nvidia and Samsung aint the best of budies right now...

 

Well they could pick TSMC too, if their process is competitive. But the article does mention a rumor about Nvidia having a contract for 14nm GPUs.

[SurvivorNVL]

*Throws money*  Take it, gimme.  Please.

[aom]

Well they could pick TSMC too, if their process is competitive. But the article does mention a rumor about Nvidia having a contract for 14nm GPUs.

doesn't seem likely, yields from TSMC havent been great with their 16nm process. Afaik, they dont have 14nm for mass production ready yet.

[Delphi1662]

Well they could pick TSMC too, if their process is competitive. But the article does mention a rumor about Nvidia having a contract for 14nm GPUs.

Every one seems to be jumping ship from TSMC as of late

Samsung's making a killing 

[Sakkura]

doesn't seem likely, yields from TSMC havent been great with their 16nm process. Afaik, they dont have 14nm for mass production ready yet.

 

TSMC has multiple 16nm processes, and they aren't in mass production yet. Their next process node after 16nm would be 10nm. Those numbers are somewhat arbitrary and inaccurate though. A 16nm process can be comparable to a 14nm process depending on the specifics of each process.

[Opcode]

It won't make any difference, because Nvidia will almost certainly have access to the same process node at the same time.

This is why they just need to keep stacking on the shaders.

 

"AI-100"

~500-550mm² @ 14nm GloFo

~6000-8000 Shader Cores

4096 Bit DDR HBM2

"AI-104"

~350mm² @ 14nm GloFo

~4000-5000 Shader Cores

2048 Bit DDR HBM2

"AI-106"

~200mm² @ 14nm GloFo

~2000-2500 Shader Cores

256 Bit DDR GDDR5

(or 1024 Bit DDR HBM2)

"AI-107"

~120mm² @ 14nm GloFo

~1000-1200 Shader Cores

128 Bit DDR GDDR5

doesn't seem likely, yields from TSMC havent been great with their 16nm process. Afaik, they dont have 14nm for mass production ready yet.

TSMC Capex is up there with Samsung's this year both spending more than Intel. So I would expect a mature 14nm process from them sometime this year.

[patrickjp93]

It would be smart of AMD to adopt the same tick-tock strategy Intel did. Especially when it comes to lack of research funds, it would be a boon for such a company.

[Opcode]

It would be smart of AMD to adopt the same tick-tock strategy Intel did. Especially when it comes to lack of research funds, it would be a boon for such a company.

They should no matter what even if Zen happens to beat out the competition. AMD has done this time and time again thinking they could build a better architecture. They did it with K8 and they shouldn't of done it with K10 because there's no doubt Phenom III, IV or even V would of been a beast right now if they kept refining it.

[patrickjp93]

They should no matter what even if Zen happens to beat out the competition. AMD has done this time and time again thinking they could build a better architecture. They did it with K8 and they shouldn't of done it with K10 because there's no doubt Phenom III, IV or even V would of been a beast right now if they kept refining it.

Particularly if they hadn't cut corners on Bulldozer the landscape would be very different right now, though I'm sure Intel would still be in the lead since software has been so slow to evolve.

[Mew]

This is no surprise- Intel's already on 14nm. Samsung *is* also on 14nm. Apple will also be at the end of the year. nVidia's lineup will most likely be on 14nm as well.

[Opcode]

Particularly if they hadn't cut corners on Bulldozer the landscape would be very different right now, though I'm sure Intel would still be in the lead since software has been so slow to evolve.

They hit a frequency wall with K10 and couldn't squeeze anything else out of it. Although they should of went back to the drawing board with it and worked around that problem. A Phenom III on 28nm that can hit 5.0 GHz like the current Bulldozer revisions wouldn't of been so bad. Although it's hard to say how a chip will perform (or even work at all) until it hits the fab. One of AMD's biggest downfalls is they think they can always make a better architecture. They've been kind of sticking to Intel's tick/tock cycle with Bulldozer mainly because they have no choice without the funds to throw around for R&D except for K12/Zen. If Kellers team delivers I think they will stick to that architecture for a while (as they should).

[patrickjp93]

They hit a frequency wall with K10 and couldn't squeeze anything else out of it. Although they should of went back to the drawing board with it and worked around that problem. A Phenom III on 28nm that can hit 5.0 GHz like the current Bulldozer revisions wouldn't of been so bad. Although it's hard to say how a chip will perform (or even work at all) until it hits the fab. One of AMD's biggest downfalls is they think they can always make a better architecture. They've been kind of sticking to Intel's tick/tock cycle with Bulldozer mainly because they have no choice without the funds to throw around for R&D except for K12/Zen. If Kellers team delivers I think they will stick to that architecture for a while (as they should).

Nah, must develop new Nirvana architecture for 10nm and plan Paradise for 7nm. All will be ground-up brand new designs, and to save costs AMD will let Koduri and Keller go to the highest bidder to add more money to the R&D budget.

Furthermore, Rory Reed will be put in charge of finances and oversight for the R&D departments. Every penny will be counted twice before a decision is considered.

[Lawliet93]

Well, we can clearly say, that nvidia will not have the access to the 14nm manufacturing process, due to their relationship with SAMSUNG (they will use the 16nm process from TSMC). In my personal opinion, amd will ad some GCN-cores to the FIJI-XT GPU and improve the architecture a little bit, therefore, I estimate that, there will be a something like a 20% performance improvement over the original FIJI GPU.

[kacper]

Well, we can clearly say, that nvidia will not have the access to the 14nm manufacturing process, due to their relationship with SAMSUNG (they will use the 16nm process from TSMC). In my personal opinion, amd will ad some GCN-cores to the FIJI-XT GPU and improve the architecture a little bit, therefore, I estimate that, there will be a something like a 20% performance improvement over the original FIJI GPU.

Why would Samsung want to lose such a big client like Nvidia? Yeah, I know about lawsuits , but they don't really matter when there is that much money on the table.

 

Anyway ,I stopped believing that gpu will see new manufacturing process. We'll be on 28 nm forever.