[Sweclockers] AMD Zen coming in Q3 2016, will be on 14 nm

[AtenRa]

Right tools is relative. They will still be 1-2 node generations behind with 20FF like they've been for a long time already.

14nm FF is much better than Intels 22nm FF. It will give a huge boost in performance/watt something 28nm cannot directly compete today. The Mobile segment will see the biggest boost in perf/watt ever seen from an AMD APU, even better than Carrizo this year.

 

[III-V]

14nm FF is much better than Intels 22nm FF

In density. It lags elsewhere.

 

[AtenRa]

In density. It lags elsewhere.

Dont think so

 

[III-V]

Dont think so

It's already been proven that TSMC's 16FF lags Intel's 22nm in performance; you think it's any different for Samsung?

These foundries lag Intel in performance by 3-4 years.

 

[AtenRa]

It's already been proven that TSMC's 16FF lags Intel's 22nm in performance; you think it's any different for Samsung?

These foundries lag Intel in performance by 3-4 years.

Where is that proven ???

 

[Arachnotronic]

Where is that proven ???

It's not proven because it's not true. That said, what III-V probably meant to say is that Intel's 22nm process offers very similar drive currents (i.e. transistor performance) to TSMC's 16FF process.

Intel's 14nm drive currents are crazy good, and far beyond anything TSMC has published.

It is wishful thinking to believe that Intel doesn't have a manufacturing technology lead in PC chips and that AMD is suddenly going to catch up because Samsung/TSMC called their nodes "14/16nm".

 

[scannall]

It's not proven because it's not true. That said, what III-V probably meant to say is that Intel's 22nm process offers very similar drive currents (i.e. transistor performance) to TSMC's 16FF process.

Intel's 14nm drive currents are crazy good, and far beyond anything TSMC has published.

It is wishful thinking to believe that Intel doesn't have a manufacturing technology lead in PC chips and that AMD is suddenly going to catch up because Samsung/TSMC called their nodes "14/16nm".

I am not seeing anyone claim that. There are those of us that are interested in seeing the architecture they come up with is all. With no preconceptions, or fan nonsense. Just an interest in technology.

While I do think over the next 5 years or so the foundry business will tighten up, and be a lot closer in nodes that really has nothing to do with what most in the thread are trying to say.

Do I think AMD will pull a magic bunny out of a hat, and beat Intel at IPC all of a sudden? Extremely unlikely. But that doesn't make a new architecture uninteresting.

 

[III-V]

Where is that proven ???

Here:

Here's some hard data to back this up:
http://www-inst.eecs.berkeley.edu/~ee290d/fa13/LectureNotes/Lecture15.pdf

Look at slide 18. TSMC has roughly equivalent performance to Intel's 22nm... on a process that won't show up until next year at best. That lines up exactly with your 1st gen FinFET bit -- a 3 and a half year lag. And their PMOS performance is trailing considerably. And it's unlikely that things have moved much since then, since their targets would have already been dialed in, and they'd just be focusing on yields at this point.

Interestingly enough, this was in a conversation with you.

I don't think he and I were ever talking about the same thing.

I was specifically addressing a technical capability gap in transistors and (separately) interconnect, something which is relevant to an R&D pipeline and the project management decisions made therein.

Whereas it appears to me he is discussing node timelines which is inherently a matter of business strategy and customer's economics, a very differently managed aspect of the foundry business altogether.

In terms of node timelines and technical equivalencies one could attempt to make for comparable products (IC designs) exiting the fabs, I would expect TSMC's 10nm mobile node to be on par with Intel's 14nm mobile. Meaning about 4yrs behind Intel in terms of equivalent yield and production volume timeline but comparable electrical parametrics and design densities. (going back to my comment regarding the 4yr gap in interconnect capabilities)

It's not proven because it's not true. That said, what III-V probably meant to say is that Intel's 22nm process offers very similar drive currents (i.e. transistor performance) to TSMC's 16FF process.

Intel's 14nm drive currents are crazy good, and far beyond anything TSMC has published.

It is wishful thinking to believe that Intel doesn't have a manufacturing technology lead in PC chips and that AMD is suddenly going to catch up because Samsung/TSMC called their nodes "14/16nm".

It is true, though.

 

[Arachnotronic]

Here:

It is true, though.

Perhaps I misunderstood. Are you talking about node-to-node, or are you talking about time-adjusted node comparisons?

If it's the latter then of course you are correct (22nm has been with us for a long time), and I offer my apology for misunderstanding

 

[III-V]

Perhaps I misunderstood. Are you talking about node-to-node, or are you talking about time-adjusted node comparisons?

If it's the latter then of course you are correct (22nm has been with us for a long time), and I offer my apology for misunderstanding

It's true in both cases, in that example. Not only did Intel's 22nm outperform TSMC's 16nm (named 22nm in the presentation -- TSMC had yet to put their marketing spin wizards on the job and label their 20nm process as 16nm), it released nearly 5 years before (1H 2012 vs 1H 2017).

While I do think over the next 5 years or so the foundry business will tighten up, and be a lot closer in nodes that really has nothing to do with what most in the thread are trying to say.

I think the signs have pointed in the opposite direction. Intel not only has the performance lead they've always held, but their density lag has been mitigated entirely at 14nm, and they'll end up having a very clear lead at 10nm.

 

[Fjodor2001]

New article about this topic on Sweclockers today:

"AMD Summit Ridge and Bristol Ridge share the socket FM3"

http://www.sweclockers.com/nyhet/19964-amd-summit-ridge-och-bristol-ridge-samsas-om-sockel-fm3

Google translated to English:

https://translate.google.com/transl...bristol-ridge-samsas-om-sockel-fm3&edit-text=

Some highlights:

The end of 2016 marks not only the beginning of the long awaited Zen, but also the end of the sockets FM2 + and aging AM3 +. These are replaced by socket FM3, which thus could house both performance-oriented Summit Ridge and the more budget-friendly Bristol Ridge with integrated graphics.
[...]

Specifications: Summit Ridge and Bristol Ridge

                       AMD Summit Ridge     AMD Bristol Ridge
Segments               "Performance"        "Mainstream"
Technology             14 nm                28 nm
Architecture           Zen                  Excavator
Cores                  Up to 8              Up to 4
L3 cache               Yes (unknown nr MB)  -
Graphics Architecture  -                    Next gen. GCN
Stream Processors      -                    Up to 512 pcs.
Memory Support         DDR4                 DDR4
PCI Express 3.0        Yes                  Yes (nr channels unknown for both)
TDP                    Up to 95 W           Up to 95 W
Socket                 FM3                  FM3
Southbridge            Promontory           Promontory
Launch Window          Q3 2016              Q3 2016

[...]

The next detail is that both processors share the completely unknown southbridge Promontory, which will be in charge of controlling for everything from storage to USB connections.

Interestingly, according to previous data AMD partnered with ASMedia for the development of future chipsets.

 

[caswow]

28nm q3 2016 cpus must be cheap as hell.

 

[witeken]

Here:

Originally Posted by Homeles
Here's some hard data to back this up:
http://www-inst.eecs.berkeley.edu/~e.../Lecture15.pdf

Look at slide 18. TSMC has roughly equivalent performance to Intel's 22nm... on a process that won't show up until next year at best. That lines up exactly with your 1st gen FinFET bit -- a 3 and a half year lag. And their PMOS performance is trailing considerably. And it's unlikely that things have moved much since then, since their targets would have already been dialed in, and they'd just be focusing on yields at this point.

Back when Homeles was still on this forum, those were good times .

 

[witeken]

it released nearly 5 years before (1H 2012 vs 1H 2017).

What is your math behind 1H'17?

 

[MisterLilBig]

Bristol Ridge, a second Excavator based APU just with DDR4 support? And on 28nm also?

This rumor is unexpected. Doubt this would happen.

And Zen as a non-APU, doubtful.

Why would AMD do this if they are pushing HSA? I would think that they would keep going with APU's and pushing more "Compute Cores" on the high end. And considering all the backers of HSA, APU's are pretty much all the rage in the future.

 

[III-V]

What is your math behind 1H'17?

That's when TSMC will have 16nm products out, as per their recent earnings call.

 

[Arachnotronic]

That's when TSMC will have 16nm products out, as per their recent earnings call.

That's 10nm. 16FF+ will start contributing to revenue in the fourth quarter of 2015.

 

[mikk]

Unless Summit Ridge is going to attack Intels Highend platform I have to wonder why it won't have an integrated GPU. This is a big flaw in my eyes.

 

[Shehriazad]

Unless Summit Ridge is going to attack Intels Highend platform I have to wonder why it won't have an integrated GPU. This is a big flaw in my eyes.

Even if it attacks Intels Highend...it can still fit an iGPU, can't it? Unless they try to make the physical size of the chip so tiny that you might breathe it in on accident...14nm would easily allow for 8 cores, iGPU and L3 cache...or even an HBM "L4" cache.

I mean...if 28nm can fit 4 "cores", igpu and chipset (Carrizo..sure HDL..but still) then I don't see why AMD can't do that for their 14nm Zen chips.

GPU is the part where AMD is ahead of Intel..would be weird of them to not use that..especially seeing how AMD is pushing hard toward HSA...




What I am really hoping for, though, is that Zens' FM3 chips are pin compatible with FM2+(and actually fit on the socket) and have a DDR3+4 controller...I enjoyed having that back in the day (what was the name of those AMD CPUs? I forgot...). I even had a board that could fit DDR2 AND 3...that made a transition hella smooth.

 

[Fjodor2001]

Interestingly, according to previous data AMD partnered with ASMedia for the development of future chipsets.

I think this is a smart move. Frees up more R&D resources for designing the APU, which should be the main focus given AMD's limited budget situation.

 

[StinkyPinky]

Late 2016 is an age away.

 

[ShintaiDK]

Late 2016 is an age away.

And the hype and expectations will only grow in the meantime. Only to be let down dramaticly on release day for close to the last 10 years.

 

[III-V]

That's 10nm. 16FF+ will start contributing to revenue in the fourth quarter of 2015.

Jeez, no wonder why I calculated a 5 year gap. I need to get my prescriptions filled

 

[Idontcare]

Unless Summit Ridge is going to attack Intels Highend platform I have to wonder why it won't have an integrated GPU. This is a big flaw in my eyes.

There will probably be an iGPU version, but the GPU chip would be a 2.5D or 3D chip stacked module thanks to TSV.

TSVs to Split More Chips: Re-Integration is the Focus

“Die stacking is happening and AMD is doing it,” said Bryan Black, senior fellow at AMD as he reviewed today’s 2.5D and 3D packaging solutions, “but why is it happening now?” questioned Black, claiming that AMD had figured out yield issues about ten years ago and is about to use TSVs across all of its product portfolio.

Cost is the first reason, especially when making large dies at advanced nodes becomes cost prohibitive because of decreasing yields. “Silicon integration is running out of gas,” Black says, arguing that the next process node may not necessarily come out cheaper overall.

His analysis is that even though Moore’s law will give us more transistors at each new node, they will not be the right transistors, because process scaling will stop supporting diverse functionalities on a single die such as fast logic, low power logic, analog, and cache.

Hence, logically, engineers will want to break large single dies into specialized components to maximize the value of new and existing process nodes, only to be re-integrated through 2.5D and 3D stacks. In his view, IC integration will never move away from interposers, but on the contrary, silicon interposers will be the SoCs sockets of the future, hosting multi-sourced 3D components whose functionalities can scale at their own pace.

Source: http://www.eetimes.com/document.asp?doc_id=1325440&

Probably make the GPU silicon at TSMC, the CPU silicon at GF, and 2.5D interposer them to make their APU for consumer markets, no GPU chip for their server/high-end markets, etc. Best of all worlds.

 

[Enigmoid]

Even if it attacks Intels Highend...it can still fit an iGPU, can't it? Unless they try to make the physical size of the chip so tiny that you might breathe it in on accident...14nm would easily allow for 8 cores, iGPU and L3 cache...or even an HBM "L4" cache.

I mean...if 28nm can fit 4 "cores", igpu and chipset (Carrizo..sure HDL..but still) then I don't see why AMD can't do that for their 14nm Zen chips.

GPU is the part where AMD is ahead of Intel..would be weird of them to not use that..especially seeing how AMD is pushing hard toward HSA...

What I am really hoping for, though, is that Zens' FM3 chips are pin compatible with FM2+(and actually fit on the socket) and have a DDR3+4 controller...I enjoyed having that back in the day (what was the name of those AMD CPUs? I forgot...). I even had a board that could fit DDR2 AND 3...that made a transition hella smooth.

Its not so much can they fit but can they make it economic to fit the components. 14 nm will be more expensive than depreciated 28 nm.