TSMC begins 16nm FinFET volume production

[JDG1980]

TSMC recently reported that it has begun volume production of 16nm FinFET products in Q2 2015. This was on an earnings call, so it's not just marketing BS (there would be legal consequences for an outright lie here).

The question, of course, is what kind of products we're talking about. Is it just going to be smartphones and other such crap for the rest of 2015, or might we actually be getting a dGPU on 16nm near year's end? If that does happen, my bet is that it will be a low-end Pascal product (something like GP107) from Nvidia. It's pretty clear that AMD will be going with GloFo's Samsung-derived FinFET process, and Nvidia can't go forward with the big Pascal chips yet because HBM2 isn't ready. But a small Pascal chip which provides better than GM206 levels of performance at GM107 levels of power consumption would be a very strong seller, not only on desktops, but also in the mobile market. And such a chip would still be using GDDR5 because HBM isn't economically viable yet at that low level, and probably won't be for two or three more years.

 

[Azix]

Global foundries same with 14nm. AMD said they taped out 14nm FINFET already. this stuffs late but w/e

 

[JDG1980]

Global foundries same with 14nm. AMD said they taped out 14nm FINFET already. this stuffs late but w/e

But taping out is not the same thing as volume production.

 

[shady28]

I seriously doubt that's going to happen. I think TSMC ramp-up is for the Apple A9. Apple reportedly switched a big part of the A9 production to TSMC back in April because Samsung / GloFlo had poor yields. They probably won't have capacity to do much of anything else.

I don't think we will see 16nm on GPUs until late next year. I even question if all of the next gen GPUs will be 16nm, the financial incentive to go there for multiple SKUs is pretty small, right now it would mean price increases. Add to that TSMC has a new 22nm HP-C process. I wouldn't be real surprised to see that get used for some chips.

 

[Silverforce11]

Mobiles. Others need to catch up to Samsung's 14nm ff already shipped many months ago.

 

[raghu78]

I think TSMC yields on 16FF+ are definitely better than Samsung 14LPE as TSMC has excellent yield learning from 20SOC ramp which shares the same back end as 16FF+. I think we might be in for a surprise as far as Apple A9 and A9X is concerned. Until we see a chipworks teardown I am not willing to write off the possibility of A9 and A9X being built at TSMC 16FF+ . Apple loves margins and better yields means better margins. So it would not surprise me if TSMC edged out Samsung for the A9 and A9X. Anyway Sep-Nov is going to be exciting to find out who has the business of the world's most powerful and largest technology company.

AMD said in their conference call that they taped out a couple of designs on FINFET. They did not mention the foundry and node. But my guess is its TSMC 16FF+ due to better yields.

http://seekingalpha.com/article/333...-results-earnings-call-transcript?part=single

Lisa T. Su - President, Chief Executive Officer & Director
Yes, Matt, so I do think the process technology landscape right now is quite interesting. So on your – the first part of your question, how do we view FinFET technologies? Actually, I think the maturity of FinFET technologies is coming along very nicely, and so we see it as an important part of our roadmap in 2016, across all of our markets. We have actually just taped out our first couple of FinFET designs. Relative to what that means for the competitive landscape going forward, I have been asked that question a couple of times over the last – last year, and my comments have been our focus is on design architecture, and it includes – and ensuring that we use all of our design architecture expertise.

So, Zen is a clean sheet design, and from an architectural standpoint I think it's going to be very competitive. The fact that the gap between foundry technologies and other technologies is shrinking, I think does change the competitive landscape and will be a good opportunity as we go forward competitively. So we are aggressively going after FinFET, and I think that's going to be an incredibly important node for us.

 

[Silverforce11]

This is the full relevant exchange:

Vivek Arya - Bank of America Merrill Lynch
Thanks for taking my question. On the Wafer Supply Agreement with GLOBALFOUNDRIES, I believe it was for $1 billion-plus or so this year. And so far you were done with $367 million. I'm just curious, Devinder, do you think you will be able to fulfill this obligation this year? And if not, are there any financial implications we should be taking into account?

Devinder Kumar - Chief Financial Officer & Senior Vice President
Yes, Vivek, I think on the Wafer Supply Agreement, we have taken about $400 million-plus for the year against the commitment we made earlier this year. But we are working actively to re-profile our wafer purchase commitment with GLOBALFOUNDRIES, in particular, given the business outlook.

Matthew D. Ramsay - Canaccord Genuity, Inc.
Thank you very much for taking my questions. A little bit of a, I guess, longer-term strategic question for you, Lisa. Last night Intel announced the addition of a new chip on their 14-nanometer roadmap and pushed back 10-nanometer. It strikes two things; one, love to get your commentary relative to your foundry partners as to how the Moore's Law progression is going particularly with Zen coming on 14-nanometer next year, and second, it looks like now you will be in a position to potentially overlap your Zen products with a generation of Intel products that is still on 14-nanometer. Just your reaction to that in general and the competitive landscape on the foundry side. Thanks.

Lisa T. Su - President, Chief Executive Officer & Director
Yes, Matt, so I do think the process technology landscape right now is quite interesting. So on your – the first part of your question, how do we view FinFET technologies? Actually, I think the maturity of FinFET technologies is coming along very nicely, and so we see it as an important part of our roadmap in 2016, across all of our markets. We have actually just taped out our first couple of FinFET designs. Relative to what that means for the competitive landscape going forward, I have been asked that question a couple of times over the last – last year, and my comments have been our focus is on design architecture, and it includes – and ensuring that we use all of our design architecture expertise.

So, Zen is a clean sheet design, and from an architectural standpoint I think it's going to be very competitive. The fact that the gap between foundry technologies and other technologies is shrinking, I think does change the competitive landscape and will be a good opportunity as we go forward competitively. So we are aggressively going after FinFET, and I think that's going to be an incredibly important node for us.

Basically AMD's FF is 14nm from Global, because they actually are under using that resource which they sign for. They literally have to use it because of the agreements unless GF just cannot deliver at all.

There was recent talk of NV looking to produce ff chips with Samsung. That suggests to me TSMC is in trouble despite what they publicly say (which from their history, you cannot trust them to deliver on their promise of volume & yields).

Samsung is also the only one outside of Intel to have volume of ff/next-node tech for a long time already, so any kinks, would likely be sorted out earlier than TSMC transitioning to ff.

 

[raghu78]

This is the full relevant exchange:

Basically AMD's FF is 14nm from Global, because they actually are under using that resource which they sign for. They literally have to use it because of the agreements unless GF just cannot deliver at all.

There was recent talk of NV looking to produce ff chips with Samsung. That suggests to me TSMC is in trouble despite what they publicly say (which from their history, you cannot trust them to deliver on their promise of volume & yields).

Samsung is also the only one outside of Intel to have volume of ff/next-node tech for a long time already, so any kinks, would likely be sorted out earlier than TSMC transitioning to ff.

silver if you see its the analyst who makes the statement on Zen being at 14nm. Thats an assumption on his part. Lisa does not confirm or deny anything as far as foundry and FINFET process is concerned. AMD has not made any statements even at their FAD 2015 regarding the foundry and FINFET process. They have just said Zen and next gen graphics products at FINFET in 2016.

http://www.anandtech.com/show/9239/amd-financial-analyst-day-2015-roundup

Nothing more and nothing else. With Fiji now confirmed at TSMC I do not think AMD is going to risk GF for their most important product next gen launches. TSMC has proven track record. We will know when the chipworks teardown for Apple A9 and Snapdragon 820 make it to tech websites in Q4 2015 where Samsung and TSMC stand.

 

[Azix]

But taping out is not the same thing as volume production.

referring to AMD on that second bit

Global is the one doing volume

http://www.kitguru.net/components/g...14nm-chips-for-our-customers-yields-on-track/

 

[Elixer]

With Fiji now confirmed at TSMC I do not think AMD is going to risk GF for their most important product next gen launches. TSMC has proven track record.

As was mentioned, AMD has a contract with GF, and if their 14nm node works, then AMD pretty much must use them.
If it doesn't, then, they will use TSMC, as they did with Fiji.
TSMCs record is far from proven, and while 28nm is proven now, it took them a long time to get there, and delays after delays. The same will most likely be true for anything smaller than 28nm coming from TSMC.

 

[Mondozei]

I seriously doubt that's going to happen. I think TSMC ramp-up is for the Apple A9.

I don't think we will see 16nm on GPUs until late next year. I even question if all of the next gen GPUs will be 16nm, the financial incentive to go there for multiple SKUs is pretty small

From what I've read, the 16 nm is pretty damn expensive.

There have been noises that NV will release Big Pascal in Q1 2016. Yet most people now peg it at Q2 and it will probably be the Tesla and/or Titan cards first, precisely because of the cost of going 16 nm+ HBM2(AMD's tie-up with SK Hynix doesn't help matters) is pretty high.

 

[Silverforce11]

From what I've read, the 16 nm is pretty damn expensive.

There have been noises that NV will release Big Pascal in Q1 2016. Yet most people now peg it at Q2 and it will probably be the Tesla and/or Titan cards first, precisely because of the cost of going 16 nm+ HBM2(AMD's tie-up with SK Hynix doesn't help matters) is pretty high.

Amazing if they can release big Pascal with HBM2 before HBM2 is even made in volume. SK Hynix plans to mass-produce HBM2 in 2H 2016.

 

[ShintaiDK]

Its a LP SoC production with low yields and Apple/Qualcomm gonna sit on it all for the next 12 months+. Forget GPUs.

GloFo couldnt even deliver so TSMC gets increased volume from Apple.

 

[Keysplayr]

I wonder if they'll be stuck at 16/14nm even longer than they were stuck at 28nm.

 

[tviceman]

Amazing if they can release big Pascal with HBM2 before HBM2 is even made in volume. SK Hynix plans to mass-produce HBM2 in 2H 2016.

Who says gen 1 Pascal is going to have HBM? 7ghz vram on a 512-bit bus provides 450 gb/s bandwidth. Given that GM200 is essentially faster than Fury X, despite Fury X having 50% more bandwidth, I think a 33% increase should be sufficient to drive a sizable performance increase if big Pascal truly is ready to deploy. I don't believe the rumors though. I think we'll see small Pascal first.

 

[ShintaiDK]

I wonder if they'll be stuck at 16/14nm even longer than they were stuck at 28nm.

I would be surprised if there ever comes a dGPU below 16/14nm. Its hard to imagine already that anyone but nVidia can make money on 16nm dGPUs. And with a constant volume decline, thats where it will end.

 

[n0x1ous]

I would be surprised if there ever comes a dGPU below 16/14nm. Its hard to imagine already that anyone but nVidia can make money on 16nm dGPUs. And with a constant volume decline, thats where it will end.

With VR around the corner the need for GPU horsepower has never been greater, it will not end the dgpu. i know your rebuttal will be that the need is irrelevant because the chips will be too expensive to make.

to that i say just raise the price even more. nvidia has shown they can do that and not miss a beat and thats what i expect will happen. i and many others will be willing to buy regardless...

 

[Azix]

HBM will likely be the major source of performance increase on FINFET for the first stuff. I'd imagine adding more transistors with the space saved with HBM and increased clocks. Seems most economical. For AMD that would be ideal since it seems GCN is simply superior as an architecture. They can ride it with revisions for a few years. After the first GPUs on the smaller process maybe some real major architecture improvements.

Not sure nvidia could afford staying on GDDR5 AND going FINFET. Maybe rebadge/refresh some 28nm products if there is an issue but they would have been investing in building pascal for HBM I think.

 

[ShintaiDK]

With VR around the corner the need for GPU horsepower has never been greater, it will not end the dgpu. i know your rebuttal will be that the need is irrelevant because the chips will be too expensive to make.

to that i say just raise the price even more. nvidia has shown they can do that and not miss a beat and thats what i expect will happen. i and many others will be willing to buy regardless...

If VR can change the volume decline to a growth, then we can talk about it.

Else its just wishful dreams.

 

[Silverforce11]

Who says gen 1 Pascal is going to have HBM? 7ghz vram on a 512-bit bus provides 450 gb/s bandwidth. Given that GM200 is essentially faster than Fury X, despite Fury X having 50% more bandwidth, I think a 33% increase should be sufficient to drive a sizable performance increase if big Pascal truly is ready to deploy. I don't believe the rumors though. I think we'll see small Pascal first.

Theoretically, big Pascal with a major node jump from 28nm to 16ff, should at least have twice the scaling or more (for uarch improvements). So 33% bandwidth is probably a bottleneck preventing good shader uptime.

NV will need HBM2 because we're tapping out GDDR5 bandwidth. Plus the 512 bit bus is wasteful for TDP that could be better spent on core performance.

They'll need to go with HBM2 to match the big perf leap basically.

I use to think they will go small Pascal first but it makes more sense for them to go big Pascal simultaneous, if the yields will be rubbish it won't matter since $5-10K Teslas will make it worthwhile. Plus it will give them time to sort out yield issues for the consumer variant later.

ps. GM200 is faster at 1440p and below. 4K is the domain of Fury X, since no single GPU can handle new titles at 4K, when we approach multi-GPU territory, XDMA has made SLI look bad. Though I suspect with the new wave of AAA DX12 titles soon, Battlefront, Hitman, Deus Ex, Fable (their devs can't praise GCN enough!), GM200's lead at 1440p will not last very long.

 

[tviceman]

ps. GM200 is faster at 1440p and below. 4K is the domain of Fury X, since no single GPU can handle new titles at 4K, when we approach multi-GPU territory, XDMA has made SLI look bad. Though I suspect with the new wave of AAA DX12 titles soon, Battlefront, Hitman, Deus Ex, Fable (their devs can't praise GCN enough!), GM200's lead at 1440p will not last very long.

PPS Titan X is still GM200 and is faster than Fury X at 4k and has considerably more OC headroom. AMD has better multi GPU scaling but that has nothing to do with this.

 

[railven]

PPS Titan X is still GM200 and is faster than Fury X at 4k and has considerably more OC headroom. AMD has better multi GPU scaling but that has nothing to do with this.

Well, of course it does. Because that's where AMD wins.


Following the CPU section, I though GloFlo was like the devil? Aren't they essentially sucking AMD dry at the moment? Why is it in the GPU section GloFlo is some kind of miracle Fab that's going to save AMD GPUs?

Kind of weird jumping between the two subs. What isn't weird is the hail mary expectations nor the doom and gloom haha.

 

[ShintaiDK]

Well, of course it does. Because that's where AMD wins.


Following the CPU section, I though GloFlo was like the devil? Aren't they essentially sucking AMD dry at the moment? Why is it in the GPU section GloFlo is some kind of miracle Fab that's going to save AMD GPUs?

Kind of weird jumping between the two subs. What isn't weird is the hail mary expectations nor the doom and gloom haha.

This says it all:
http://appleinsider.com/articles/15...tsmc-for-30-of-a9-chip-orders-for-next-iphone

(A9X is already TSMC)

That would mean Global Foundries is out of the game. They couldnt deliver. Nothing new in that tho.

 

[tviceman]

Well, of course it does. Because that's where AMD wins.

Hahahaha exactly. Next gen node, new architecture, bandwidth, MULTI-GPU FURY X MASTER RACE 4k PROJECT CARS IS TEH D3V1L

Following the CPU section, I though GloFlo was like the devil? Aren't they essentially sucking AMD dry at the moment? Why is it in the GPU section GloFlo is some kind of miracle Fab that's going to save AMD GPUs?

Kind of weird jumping between the two subs. What isn't weird is the hail mary expectations nor the doom and gloom haha.

All of this x 1000. This is so true. In CPU's AMD is so far behind that only a few leftover die-hards are preaching from the mount. At least in GPU's AMD is hanging in the game, even though the time it takes to catch up is increasing. I notice how when every new AMD high end GPU is imminent, there is massive Nvidia-killing hype and expectations of 20% better performance than Nvidia's comparable flagships are regurgitate over and over again.

 

[railven]

This says it all:
http://appleinsider.com/articles/15...tsmc-for-30-of-a9-chip-orders-for-next-iphone

(A9X is already TSMC)

That would mean Global Foundries is out of the game. They couldnt deliver. Nothing new in that tho.

What I keep reading, and this isn't the first time GloFlo floundered, either, is it? Yet, that deal AMD made keeps them dropping cash on them with basically no results to show for it.

But I guess, with DX12, and Win10, and GloFlo - AMD is some how going to dominate. Woof. These expectations are so high, I just keep being reminded of the 290X on Mantle beating 780 SLI claim.