TSMC Shows Path to 16nm, Beyond

[jpiniero]

That's probably the least informed statement I've seen on this board...today. You do realize ~2% of Xeon Phi's logic is x86, right? Most of the work is done by a fat vector unit.

It's still the old x86 way though. They built it that way because they wanted it to be compatible with the current AVX units in the Xeons. It's just not as efficient as the GPU strategy that nVidia is taking, and will be tough to overcome with just better fabs.

What really looks silly is that Broadwell GT4 might be able to deliver SP Flops comparable to the current 220W Phi in a laptop.

 

[ShintaiDK]

It's still the old x86 way though. They built it that way because they wanted it to be compatible with the current AVX units in the Xeons. It's just not as efficient as the GPU strategy that nVidia is taking, and will be tough to overcome with just better fabs.

What really looks silly is that Broadwell GT4 might be able to deliver SP Flops comparable to the current 220W Phi in a laptop.

Really. You need to do your homework.

AVX is NOT cross compatible before the 14nm Xeon Phi and Skylake. And its about DP flops, not SP.

One could also say a 250W GTX780 only does 165Gflops DP....

 

[krumme]

Risk production is something you call your production run when the process flow itself is still not functioning within the spec'ed electrical parametrics but you intend to tweak and fiddle with the process flow in parallel to your production run.

Everybody does it. It is all about managing time-to-market risk.

So let's say for example that your have a 20nm process flow which is spec'ed on paper as delivering 1.2mA/um drive current at 100pA/um off current. But in reality the process flow is delivering wafers with xtors that are both too cold (low drive current, say 0.9mA/um) and too leaky (say 200 pA/um), oh and the yield is a paltry 8%.

Obviously your R&D team is still hustling like mad to get the process flow tweaked so the electrical parametrics come in on target, and in parallel they are hustling like mad to get the yield to 80% instead of 8%.

But in the meantime there will always be a customer or two waiting in the wings who just cannot get their hands on the 20nm early production stuff soon enough. They'll pay a price-premium for it, on top of knowing full well that the yields are very bad and the electrical properties are not yet up to spec.

That is risk production. And if you are an IDM like Intel then your customer is an internal customer but a customer all the same. Someone will always want your buggy unpolished unfinished process node if for no other reason than to start getting through there own stages of design validation or early sales of limited run products.

And the principles here is the excact same for other products.

There is solid benefit beeing an idm as Intel. Knowledge sharing and the ability to prioritze long term strategic important technology, processes or platforms. A more holistc aproach.

The drawbacks is the valuation; price of the products and services. It can get cloudy and unhealthy cultural and historic "as we use to do", can hinder more rational approach.

(Edit: btw notice why the emphasis on customer colaboration from tscm old man is a very important message here. To lessen the drawbacks of not beeing an idm. I think its brilliant seen, where tscm is now)

 

[NTMBK]

Considering that Nvidia is now a legit competitor to them in the server market place, I don't see how you can say this.

Don't make me laugh. NVidia is a rival in very niche HPC markets, but the bulk of the server market is off limits to them. Call me when a Tesla runs a SQL server.

 

[Ajay]

A report from DigiTimes indicates that TSMC has been busy lately with acquiring equipment and preparing to cross over to 20 nm lithography. According to the report, the company will be making the switch in Q1 2014. Shipments for the 16 nm HKMG process have also started, and the company expects mass production of FinFETs to commence about a year from now.

Busy acquiring equipment?? If TSMC is on target for Q1 2014, I'd think they'd be pretty nearly done acquiring equipment. Right now I think they'd be doing test batches, QA/Verification, calibration and systems tests and running around with their hair on fire trying to hit their targets on yeilds without changing their electrostatics (as IDC eloquently informed us).

Intel's lead is in no danger unless their margins collapse for some unexpected reason. Then they'd have to rethink their leading-edge fab strategy; but right now they have the cash and talent to continue forward. 7nm could prove challenging if EUV isn't ready by then, but Bohr has mentioned that Intel has other options.

 

[ShintaiDK]

If Digitimes is right it looks real bad for TSMC and 20nm.

 

[NTMBK]

If Digitimes is right it looks real bad for TSMC and 20nm.

Yup, just taking a look at GPU roadmaps would have made that pretty clear. Intel's process lead is safe for a long time to come.

 

[Ajay]

If Digitimes is right it looks real bad for TSMC and 20nm.

If so, yes - not that it'll hurt them much (unless by some miracle, GF is doing better). This could explain rumors that Apple is going back, at least partially, to Samsung for its A9 SoC.

Yup, just taking a look at GPU roadmaps would have made that pretty clear. Intel's process lead is safe for a long time to come.

Yeah, I thought the GPU delays were just because Apple and Qualcomm paid more for a larger share of the wafer starts, but now we have some other data points that indicate it could be because of yield problems.

 

[Arachnotronic]

If so, yes - not that it'll hurt them much (unless by some miracle, GF is doing better). This could explain rumors that Apple is going back, at least partially, to Samsung for its A9 SoC.



Yeah, I thought the GPU delays were just because Apple and Qualcomm paid more for a larger share of the wafer starts, but now we have some other data points that indicate it could be because of yield problems.

Samsung's 20nm is doing worse.

 

[ShintaiDK]

Yeah, I thought the GPU delays were just because Apple and Qualcomm paid more for a larger share of the wafer starts, but now we have some other data points that indicate it could be because of yield problems.

Its due to both. I am slowly starting to ahve doubts about 20nm GPU releases in 2014. I was rather certain pervious if would be in about a year. But if that Digitimes story is right, then it could just as well be 2015.

 

[jdubs03]

Its due to both. I am slowly starting to ahve doubts about 20nm GPU releases in 2014. I was rather certain pervious if would be in about a year. But if that Digitimes story is right, then it could just as well be 2015.

It has to be very frustrating for companies like Nvidia trying to release Maxwell, where the bump to 20nm would certainly help in power reduction as well as higher performance.

If anything they'd likely release Tesla on Maxwell then consumer cards later on like they did with Kepler (GTX 700 series, GK110). Or maybe they'll release a 28nm Maxwell GTX 800 series in 2014, then wait and go 20nm Tesla then GTX 900 series in 2015.

When those 20nm Tesla's come I'm curious to see the performance difference against the future Knights Landing Xeon Phi's on 14nm. Both are targeting similar GFLOPS/W.

In the overall landscape I suspect a mass release of 20nm products will be in early 2015, where preferred customers have late 2014 releases, but oh well. That just pushes out that 16nm FinFet out, but maybe the transition to the will be easier because it'll be a hybrid 20nm/16nm.

 

[krumme]

I am optimistic about first 20nm tscm products phones q2 2014 and gpu q4. And 16nm max 1.5 year later.
Even if that happens it just shows Intel is in another league for the forseable future.

 

[krumme]

Perhaps when i think about amd historic advantage to nv for gpu processes we might see 20nm amd gpu q2 or q3. They need a 200-300usd gpu here. And the smaller die would be perfect for that.

As shown in the comments for the eetimes article 20 nm will probably bring a solid 29% cost advantage.

Hell. When i think about it. I think there is a good probability we will get the usual amd gpu process pipecleaner 4-6monts earlier than nv. But for midrange.

Whats your take on it?

 

[Fjodor2001]

Busy acquiring equipment?? If TSMC is on target for Q1 2014, I'd think they'd be pretty nearly done acquiring equipment. Right now I think they'd be doing test batches, QA/Verification, calibration and systems tests and running around with their hair on fire trying to hit their targets on yeilds without changing their electrostatics (as IDC eloquently informed us).

Well, isn't it hard to draw too definitive conclusions from that single DigiTimes statement?

I mean, perhaps TSMC already have some 20 nm production lines completed or close to being completed, and they "lately" (for the last couple of months, or what does it mean?) have been acquiring additional equipment to ramp up production capacity.

How do we know for sure this is not the case?

 

[Nothingness]

Digitimes is so unreliable, I wouldn't base any prediction based on their "info".

 

[NTMBK]

If 20nm were anywhere near ready, AMD wouldn't be launching a massive new 28nm ASIC.

 

[mrmt]

If 20nm were anywhere near ready, AMD wouldn't be launching a massive new 28nm ASIC.

And Globalfoundries wouldn't be talking about 28nm FDSOI

 

[krumme]

If 20nm were anywhere near ready, AMD wouldn't be launching a massive new 28nm ASIC.

Half a year accounts for something. Besides launching eg a 425mm2 die on a new tscm process will kill yield.
I am not saying amd will use early 20nm for eg. 130mm2 size midrange gpu. Cost for 20 vs 28 will decide that. But there is a posibility. But if its all for a57 arm then yeaa, they probably have to wait for prices to come down even for a relatively small die.

 

[krumme]

And Globalfoundries wouldn't be talking about 28nm FDSOI

Wrong interpretation imho.
Go read the comment on the eetimes link from op. I think its right to see 28fdsoi as a cheap mans future low power node. 20nm is like 28nm hpm.

 

[krumme]

Anyway what does it matter from consumer perspective we get a midrange 20nm gpu q2 at the same cost. The only upside should be for mobile? Hmm. For sure we are still at q4 for the highend desktop replacements when tscm is ready for the big stuff.

 

[Borealis7]

there is an Intel slide from about the time IVB launched showing 14nm in 2013. that was...optimistic.

 

[CHADBOGA]

there is an Intel slide from about the time IVB launched showing 14nm in 2013. that was...optimistic.

That was no doubt referring to shipping to customers in Q4 and as far as I know, that has recently been confirmed.

 

[Ajay]

Its due to both. I am slowly starting to ahve doubts about 20nm GPU releases in 2014. I was rather certain pervious if would be in about a year. But if that Digitimes story is right, then it could just as well be 2015.

Hmm, Jen-Hsun must be throwing a fit!

 

[Ajay]

Samsung's 20nm is doing worse.

Wow, it really is Intel and then everybody else (well IBM might be the best second, but they are in it really just for themselves) - everyone else is dropping the ball.

I think Samsung is supposed to be a second source for redundancies and bargaining power (though running behind schedule will nullify their benefit as a negotiation counterweight).

TSMC really needs to step up their game, otherwise Intel might just decided to go all in on being both and IDM and major Fab provider.

 

[ShintaiDK]

TSMC really needs to step up their game, otherwise Intel might just decided to go all in on being both and IDM and major Fab provider.

Intel roughly spends 3 times more on foundry R&D than TSMC.