TSMC Shows Path to 16nm, Beyond

[Arachnotronic]

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

No, Broadwell will be "well into production" by Q4 2013. Intel is right on target.

 

[Ajay]

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

True, but the trajectory isn't looking so good for TSMC, IMHO. They have been on a capacity drive lately, but they are basically N-2 (counting FinFET as a node jump) at this point and it isn't getting any easier to stay there apparently.

Just saying, really - there is a risk that Intel decides it wants all the higher margin marbles. With this fumble (if it's true) some companies with the cash (i.e. Apple, Qualcomm) could be lured into a more secure arrangement if Intel can accept margins of ~50% for fab only customers.

Actually, at this point, with 450 mm wafers and EUV/Quad patterning on the horizon, I'm surprised Intel isn't pushing harder. They must really believe that they are going to own the phone and tablet markets with x86 by the time they get to that point. Interesting times...

 

[ShintaiDK]

True, but the trajectory isn't looking so good for TSMC, IMHO. They have been on a capacity drive lately, but they are basically N-2 (counting FinFET as a node jump) at this point and it isn't getting any easier to stay there apparently.

Just saying, really - there is a risk that Intel decides it wants all the higher margin marbles. With this fumble (if it's true) some companies with the cash (i.e. Apple, Qualcomm) could be lured into a more secure arrangement if Intel can accept margins of ~50% for fab only customers.

Actually, at this point, with 450 mm wafers and EUV/Quad patterning on the horizon, I'm surprised Intel isn't pushing harder. They must really believe that they are going to own the phone and tablet markets with x86 by the time they get to that point. Interesting times...

I dont think the margins is the issue. TSMC already gets around 45%.

But you cant give everyone 14nm for example. Just look at the Haswell Refresh. Intel is not selling desktop Haswell 2 years for fun. But because its not possible in the world to do it practically faster. 14nm goes to Atom as new product and left a hole for desktops. While Intel could maybe run 1-2 fabs extra outside own needs with 22nm is another matter tho. But again, it gives other challenges like intercompetition. Intel only allows companies access that doesnt compete with Intel.

 

[Khato]

Intel only allows companies access that doesnt compete with Intel.

It'll be quite interesting to see if this stance changes going forward - probably really depends upon how well received Intel's push into tablets and smartphones is next year. If it goes well then there's no reason to allow the competition into its fabs. If it continues to stagnate however then there's something to be said for selling wafers at a premium. Margins for Intel would be comparable or possibly even better than what they could sell their own designs for (all indications thus far are that Baytrail-T is priced well below the competition.) Whereas the increased silicon cost would take a hit on the competition's margins that they'd have to take lest they lose a competitive edge to the other ARM players... and best of all (for Intel) is the fact that they'd be taking all the lucrative sales from the other foundries.

 

[Blandge]

Nevermind

 

[blackened23]

It'll be quite interesting to see if this stance changes going forward - probably really depends upon how well received Intel's push into tablets and smartphones is next year. If it goes well then there's no reason to allow the competition into its fabs. If it continues to stagnate however then there's something to be said for selling wafers at a premium. Margins for Intel would be comparable or possibly even better than what they could sell their own designs for (all indications thus far are that Baytrail-T is priced well below the competition.) Whereas the increased silicon cost would take a hit on the competition's margins that they'd have to take lest they lose a competitive edge to the other ARM players... and best of all (for Intel) is the fact that they'd be taking all the lucrative sales from the other foundries.

I think it's safe to state that firms will line up around the block, so to speak, to get a chance at using Intel's fabs. It's a fine line for intel, though - how do they do this without enabling a competitor? I think Bay Trail will do fine but we will see.

 

[Exophase]

I think it's safe to state that firms will line up around the block, so to speak, to get a chance at using Intel's fabs. It's a fine line for intel, though - how do they do this without enabling a competitor? I think Bay Trail will do fine but we will see.

You make it sound like it'd be easy for anyone to just go from whatever manufacturing they're using to Intel. First and foremost, Intel isn't going to offer every last thing that their competitors will, particularly because they've developed their manufacturing capabilities to suit only their needs. That will include at least some equivalents for IP libraries. And they won't use the same tools. And you don't know what their pricing would be like. Companies are going to consider more than just which process has the best tech specs.

 

[sm625]

Their path is simple: Intel + 2 years.

 

[jdubs03]

2 years sounds about right.

I know this is a little bit off topic, but I am kind of surprised that Bay Trail at 22nm FinFet is not trouncing the ARM vendors at 28nm planar. I know its Intel's first try at it, but with their process lead they should be smacking their competition around. Fast forward a couple years with TSMC at 16/20nm FinFet, ARM will gain ground on Intel if they don't step up.

Though this is just hypothetical, but say Intel and ARM are on the same performance level next year (likely, Intel may have a slight advantage). If TSMC had FinFet at the 20nm node, they'd experience a decent bump in performance while lowering power consumption just by switching to FinFet. Potentially being the top performer.

That leads me back to 16/20nm FinFet at TSMC in 2015/2016. Its gonna get even more interesting (keep in mind that Intel stays a node ahead at 10nm).

Honestly I just hope Intel can get 1.7x performance out of Airmont - Cherry Trail.

 

[Exophase]

I know this is a little bit off topic, but I am kind of surprised that Bay Trail at 22nm FinFet is not trouncing the ARM vendors at 28nm planar. I know its Intel's first try at it, but with their process lead they should be smacking their competition around. Fast forward a couple years with TSMC at 16/20nm FinFet, ARM will gain ground on Intel if they don't step up.

What were your expectations exactly? We don't have a lot of exact measurements but based on what we do know I'd say that BayTrail performs comparably with the best Android ARM systems while using maybe 40-70% as much power. That's a huge deal and easily exceeds what I'd expect from the process advantage alone. I think Intel has stepped up tremendously not just with its manufacturing but experience in CPU design at all levels.

It seems people are comparing it with A7's CPU, but I'm not aware we have any really useful power numbers for it. I would expect the gap there to be narrower, it's only natural given what Apple is putting behind this vs what their competitors are.

 

[jdubs03]

What were your expectations exactly? We don't have a lot of exact measurements but based on what we do know I'd say that BayTrail performs comparably with the best Android ARM systems while using maybe 40-70% as much power. That's a huge deal and easily exceeds what I'd expect from the process advantage alone. I think Intel has stepped up tremendously not just with its manufacturing but experience in CPU design at all levels.

It seems people are comparing it with A7's CPU, but I'm not aware we have any really useful power numbers for it. I would expect the gap there to be narrower, it's only natural given what Apple is putting behind this vs what their competitors are.

Really its just an educated guess based on prior and current data. I haven't really seen much of a power consumption test comparing Bay Trail vs. ARM vendors, except that Bay Trail was around 2-2.5W for intensive CPU benchmarks. If that is 40%+ less than ARM on the same benchmark then yes I completely agree Intel has a concrete edge. Do you have any links with Snapdragon or Tegra power consumption data?

 

[Ajay]

Really its just an educated guess based on prior and current data. I haven't really seen much of a power consumption test comparing Bay Trail vs. ARM vendors, except that Bay Trail was around 2-2.5W for intensive CPU benchmarks. If that is 40%+ less than ARM on the same benchmark then yes I completely agree Intel has a concrete edge. Do you have any links with Snapdragon or Tegra power consumption data?

Most phone SoCs use allot less, but we don't know what the top ARMv8 SoC will (or do) use. The A7 is the only ARMv8 SoC out there right now and I think AT estimated it at <= 2 Watts peak. I imagine there are some numbers out there for top of the line ARM Quads (ARMv7) which might give a better picture.

 

[Khato]

Most phone SoCs use allot less, but we don't know what the top ARMv8 SoC will (or do) use. The A7 is the only ARMv8 SoC out there right now and I think AT estimated it at <= 2 Watts peak. I imagine there are some numbers out there for top of the line ARM Quads (ARMv7) which might give a better picture.

Out of curiosity, where did AT estimate A7 power consumption?

The only data I know of to go by are the rough whole-phone power measurements done by notebookcheck. There definitely seem to be the occasional outlying data points, but for the most part it does give a good ballpark estimate for SoC load power consumption.

For example, on the Snapdragon 600 you have results going from a delta power consumption between maximum idle and maximum load of 2.3W for the LG Optimus G Pro, 2.9W for the HTC One, and 3.9W for the Samsung Galaxy S4. Sure it's a decent spread, but that power increase matches quite well with the differences in benchmarked performance between the devices when you consider that getting that last bit of frequency is done at the high cost of a lot more voltage.

Other interesting data points you can find there are a delta power consumption of 3.4W for the LG G2 and its Snapdragon 800 that didn't perform anywhere near the level of the MDP and then 7.1W for the Toshiba Excite Pro tablet with a Tegra 4.

Note that I'm not suggesting that the entirety of the delta power consumption between maximum idle and maximum load is due to the SoC... but I'd expect it to typically be pretty close. There's a good reason why none of the ARM players are willing to give power usage figures.

 

[jdubs03]

From those data points, it does look pretty favorable compared to at least Tegra 4. I'm anticipating a device review from notebookcheck for Bay Trail so we can get a concrete comparison between the two.

Its pretty crazy to think that Intel is trying to get ~10W for Broadwell, I think ARM may be becoming outclassed from above and in their market segment.

 

[Ajay]

Out of curiosity, where did AT estimate A7 power consumption?

Read that on the RWT forums, I'd have to search, not tonight though.

 

[N-A-N-0]

No, Broadwell will be "well into production" by Q4 2013. Intel is right on target.

http://www.theregister.co.uk/2013/09/10/intel_reveals_14nm_pc_declares_moores_law_alive_and_well/

Exact quote from CEO Brian Krzanic:

"This is it, folks," he said. "Fourteen nanometers is here, it's working, and will be shipping by the end of this year."

If it doesn't ship this year then it's not on their new target in September. Otherwise it looks like the first consumer 14nm is happening 2013, not 2014. More than "well into production." :thumbsup:

 

[Khato]

Read that on the RWT forums, I'd have to search, not tonight though.

No problem, I was just curious as I haven't seen anything concrete regarding it. I'm eagerly anticipating notebookcheck's iPhone 5s review to at least get a rough idea - should be able to compare it to the review they already have up for the iPhone 5c to see just how much power A7 adds over A6.

 

[Fjodor2001]

Update from Oct 30, 2013:

http://semiaccurate.com/2013/10/30/tsmc-shows-production-20nm-16nm-dev-wafers/

TSMC shows off production 20nm and 16nm dev wafers

TSMC was showing off two wafers at Techcon, a production 20nm SoC and a pre-production 16nm without any qualifiers. Not many more details so for now just enjoy the pictures [see at link above].

On the 20nm front the wafers are not just in production, there are devices on the market using it that you can buy. No word on the higher performance variants but the SoC process is good to go. 16nm as the label suggests is still in development and that is on schedule. What is that schedule? Risk production in late 2014 so you should have 16nm FinFETs from TSMC in early 2015 if all goes well.

So by how much does TSMC's 16 nm trail Intel's 14 nm? About 1 year or so?

Also, shouldn't the sentence:

"On the 20nm front the wafers are not just in production"
be
"On the 20nm front the wafers are just in production"

otherwise I think it doesn't make sense, since after that it says "...there are devices on the market using it that you can buy. No word on the higher performance variants but the SoC process is good to go."?

 

[Erenhardt]

http://semiaccurate.com/2013/10/30/tsmc-shows-production-20nm-16nm-dev-wafers/

So by how much does TSMC's 16 nm trail Intel's 14 nm? About 1 year or so?

Also, shouldn't the sentence:

"On the 20nm front the wafers are not just in production"
be
"On the 20nm front the wafers are just in production"

otherwise I think it doesn't make sense?

key-word: "just"
They could change it to "only", but I guess that is clear if one doesn't take part of the sentence out of context...

 

[Fjodor2001]

key-word: "just"
They could change it to "only", but I guess that is clear if one doesn't take part of the sentence out of context...

Yes, you're right. I read it too quickly.

So what they are saying is that TSMC's 20 nm wafers are already in production and there are devices on the market using SoC's based on that which you can buy.

 

[Homeles]

Risk production in late 2014 points to 16nm devices in 2H 2015, no? Early 2015 would be for volume production... and you won't see the devices until a few months later. I applaud Charlie for bringing us information earlier than everyone else, but it would be really helpful to his audience and to himself if he were a tad more educated on the subjects he talks about.

Also, just because there are devices on the market utilizing 20nm does not mean that it's in volume production. I'm pretty sure it's still in risk production. I have no doubt that we'll see 20nm soon, but it's too early for meaningful volumes.

 

[Ajay]

Update from Oct 30, 2013:

http://semiaccurate.com/2013/10/30/tsmc-shows-production-20nm-16nm-dev-wafers/

So by how much does TSMC's 16 nm trail Intel's 14 nm? About 1 year or so?

No, from what I've read - TSMC's 16nm will be no where near Intel's 14nm in terms of density and electrostatics. Of course, TSMC will do just fine - their real competition is Samsung and GFL, who are pretty much following the same path but still up river compared to TSMC.

 

[lefty2]

Presumably, the 20nm products on market are from Xilinx, because they were the first to tape-out

 

[Erenhardt]

No, from what I've read - TSMC's 16nm will be no where near Intel's 14nm in terms of density and electrostatics. Of course, TSMC will do just fine - their real competition is Samsung and GFL, who are pretty much following the same path but still up river compared to TSMC.

Can we read it too?

 

[Homeles]

Can we read it too?

I'm not sure what sources he's read, but it's essentially common knowledge. TSMC and the other foundries can't keep up with Intel's metal pitches. GloFo and TSMC will be using their 20nm BEOL on their 14nm and 16nm processes, respectively (likely to keep costs down).