Samsung to build 14nm chips next year

[jpiniero]

Based upon the A8's information, the density difference between Samsung 28 and TSMC 20 planar is about 33%... and from that info costs 30% more. So if the yield is similar the cost difference is basically nothing.

 

[Homeles]

Based upon the A8's information, the density difference between Samsung 28 and TSMC 20 planar is about 33%... and from that info costs 30% more. So if the yield is similar the cost difference is basically nothing.

Do we know the A7's xtor count?

 

[witeken]

[...]
Now do the math and see how much each Transistor cost at 28nm and 20nm.

I believe you will find cost per transistor is lower at 20nm

It depends on your source, I guess:

 

[witeken]

Do we know the A7's xtor count?

~1.2B transistors.

(2B / 1.9 density decrease / 89mm²) * 102 mm² = 1.206B transistors

 

[AtenRa]

I'm still not seeing how your singular data point trumps a handful of others, like from Nvidia, ARM, and the graph I linked above.

It depends on your source, I guess:

Heh, the ones i posted are from IBS as well. :whiste:

 

[Homeles]

Heh, the ones i posted are from IBS as well. :whiste:

So TSMC finally got their act together? Too bad they're going to be losing the race in the next few months...

 

[AtenRa]

Great, you are still dodging in your typical fashion.

Did you even bother to look at the numbers ??? i guess not.

Why is your singular piece of evidence better than my numerous pieces of evidence?

Because all those are using the same source, IBS

 

[jpiniero]

Do we know the A7's xtor count?

I based the 33% based upon the difference between the two chips SRAM and Cyclone.

 

[Homeles]

Did you even bother to look at the numbers ??? i guess not.

The numbers are meaningless -- the story behind it is what counts. So TSMC had a worthless 20nm process, had to have everyone yell at them to change it, and now they've finally managed to beat out their 28nm process?

Well, good for them I guess. They're still ~3½ years behind Intel on performance, and they're going to be beaten by GloFo and Samsung with 14nm, and the comparison to Intel will be even worse by then.

Well done, you've managed to prove that TSMC is not vastly inferior to Intel's 22nm -- just mostly inferior.

 

[pw257008]

You keep refuting this, claiming it's basic economics, but the fact is that even the basic tenants of economics are highly debated.

Intel needs the same % margins off their foundries as TSMC et al, accounting for wafer costs and R&D. The advantage I can see for them is they can better segment their pricing to make more of these margins on Xeons that they can charge more for. But that is not a cost advantage across the board. It could be a de facto cost advantage at the low end, and a de facto cost disadvantage at the high end. But at the end of the day, there is no inherent ability to make less off their foundries by being integrated.

 

[jpiniero]

The numbers are meaningless -- the story behind it is what counts. So TSMC had a worthless 20nm process

Well, it can't be completely useless if Apple is using it, even if it's no cheaper.

If Samsung is really saying that it's "14 nm" is only 15% more dense than 20 nm, that's obviously not a huge hill to climb for TSMC's 16FF. I do wonder if Apple going back is more about helping out GloFo to give them more options down the road.

 

[Homeles]

Well, it can't be completely useless if Apple is using it, even if it's no cheaper.

If Samsung is really saying that it's "14 nm" is only 15% more dense than 20 nm, that's obviously not a huge hill to climb for TSMC's 16FF. I do wonder if Apple going back is more about helping out GloFo to give them more options down the road.

No, it's not a huge hill. They'll be fine anyway, as they had a stellar 28nm turnout. It's clear that the competition is heating up, though.

 

[AtenRa]

The numbers are meaningless -- the story behind it is what counts. So TSMC had a worthless 20nm process, had to have everyone yell at them to change it, and now they've finally managed to beat out their 28nm process?

Well, good for them I guess. They're still ~3½ years behind Intel on performance, and they're going to be beaten by GloFo and Samsung with 14nm, and the comparison to Intel will be even worse by then.

Well done, you've managed to prove that TSMC is not vastly inferior to Intel's 22nm -- just mostly inferior.

So we switched from 22nm FF being vastly superior, to the 20nm superiority in density and being cheaper than even 28nm when 22nm FF cost is even higher.
No wonder Intel is focusing on density more than performance with more aggressive density PR graphs lately.

And because we are talking about 14nm, here are the wafer and die cost for 14/16nm for Q4 2015.

With 10-15% higher density over 20nm, it is clear that 16nm will be more expensive in Q4 2015 than 20nm in Q4 2014.
But, 14/16nm products can command a premium higher price due to lower power consumption and higher performance than both 28nm and 20nm products.
OEM and End-buyers will pay more to have the fastest and lower power consumption so the only question is not about if 14/16nm cost higher but when they will be ready for production.
And it seams Samsung will have 14nm ready for mass production at the end of 2014 early 2015.

 

[ShintaiDK]

You forgot gate utilization and design cost. 28nm is still cheaper than 20nm in quite some years ahead. And 14/16nm is even higher cost.

That alone limits the type of products we see quite significantly.

IBS says there is a 90% chance that 10nm will only be high performance CPUs and the equal.

Even Samsung says 28nm is the sweetspot for transistor price.

 

[Homeles]

So we switched from 22nm FF being vastly superior, to the 20nm superiority in density and being cheaper than even 28nm when 22nm FF cost is even higher.
No wonder Intel is focusing on density more than performance with more aggressive density PR graphs lately.

There's been no switching.

As I stated, and have proven, they're 3½ years behind on performance.

You forgot gate utilization and design cost. 28nm is still cheaper than 20nm in quite some years ahead. And 14/16nm is even higher cost.

That alone limits the type of products we see quite significantly.

IBS says there is a 90% chance that 10nm will only be high performance CPUs and the equal.

Even Samsung says 28nm is the sweetspot for transistor price.

Shintai, AtenRa has provided updated numbers from IBS that show that 20nm has improved cost/transistor over 28nm. To allow for this, TSMC changed from using a 90nm gate pitch to using 80nm, making their 20nm actually useful.

 

[ShintaiDK]

Shintai, AtenRa has provided updated numbers from IBS that show that 20nm has improved cost/transistor over 28nm. To allow for this, TSMC changed from using a 90nm gate pitch to using 80nm, making their 20nm actually useful.

The raw die cost without the gate utilization, parametric yield and design cost?

28nm will simply stay the sweet spot for quite some time ahead.

 

[witeken]

Edit:

Just to make it more clear, lets see 100mm and 200mm die costs.

[...]

Now do the math and see how much each Transistor cost at 28nm and 20nm.

I believe you will find cost per transistor is lower at 20nm

Okay, I did the math and 20nm is only a measly 1.23x cheaper (Q4'14).

I can't find it, but I remember a quote posted somewhere that says that 20nm will be more expensive because of the increase in mask cost. Also, here's the analysis from Khato: http://forums.anandtech.com/showpost.php?p=36560660&postcount=264. So if you add those other costs, 20nm is probably not really worth it for a lot of companies compared to 28nm.

BTW, I'm not sure if the wafer price is correct.

 

[witeken]

Because all those are using the same source, IBS

Nope.

 

[Homeles]

The raw die cost without the gate utilization, parametric yield and design cost?

28nm will simply stay the sweet spot for quite some time ahead.

You aren't getting it. The numbers you are citing, from the same company, are outdated compared to his.

Supposedly. I can't find the information he's provided.

 

[Nothingness]

I can't find it, but I remember a quote posted somewhere that says that 20nm will be more expensive because of the increase in mask cost.

Isn't mask cost paid once? So for high volume the impact should be low (unless of course the price is 10x, which is unlikely).

 

[ShintaiDK]

20nm is more expensive due to design cost mainly. Thats why billion $ designs still yield lower transistor cost. Without that huge sum into the design, you end up with higher cost per useable transistor.

 

[ShintaiDK]

You aren't getting it. The numbers you are citing, from the same company, are outdated compared to his.

Supposedly. I can't find the information he's provided.

I havent seen him mentioning any of the metrics I covered.

 

[witeken]

Isn't mask cost paid once? So for high volume the impact should be low (unless of course the price is 10x, which is unlikely).

You need apparently something like 10B revenue.

 

[AtenRa]

The raw die cost without the gate utilization, parametric yield and design cost?

28nm will simply stay the sweet spot for quite some time ahead.

This is another good slide to understand why 20nm and 16nm are doing fine.

28nm full production started sometime in Q2-Q3 2011 if im not mistaken, lets say Q4 2011 to have the same quarter as the rest.
According to the above slide, 20nm in Q4 2016 will be cheaper than 28nm in Q4 2015. That means that 20nm will be cheaper only 2 years after the initial production start when 28nm took 4 years to reach the same price.
And then again 16nm at Q4 2017 will reach almost price parity with 28nm of Q4 2015. That means 16nm will only need a single year to reach price parity with 28nm at Q4 2015.

Also, in Q4 2015 28nm products will no be able to reach 20nm performance and power usage. So, although 20nm in Q4 2015 will cost more, 20nm products will command higher prices due to superior performance and lower power usage and ASP will be higher.
Same apply for Q4 2016 when 14/16nm products will be far more advanced than both 28nm and 20nm. So again those 14/16nm products will command higher ASP than the rest at that time.

This is how the industry is working for the last 20-30 years, Intel always degrease its die size of its CPUs but ASP remains the same. AMD and NVIDIA will have to do the same. NVIDIA already did that with Kepler and GK104 two years ago. They introduced a smaller die at higher ASP than before(GF110). They will do the same again at 20nm, they will introduce a smaller die than GM204 at higher ASP.

 

[ShintaiDK]

This is another good slide to understand why 20nm and 16nm are doing fine.

28nm full production started sometime in Q2-Q3 2011 if im not mistaken, lets say Q4 2011 to have the same quarter as the rest.
According to the above slide, 20nm in Q4 2016 will be cheaper than 28nm in Q4 2015. That means that 20nm will be cheaper only 2 years after the initial production start when 28nm took 4 years to reach the same price.
And then again 16nm at Q4 2017 will reach almost price parity with 28nm of Q4 2015. That means 16nm will only need a single year to reach price parity with 28nm at Q4 2015.

Also, in Q4 2015 28nm products will no be able to reach 20nm performance and power usage. So, although 20nm in Q4 2015 will cost more, 20nm products will command higher prices due to superior performance and lower power usage and ASP will be higher.
Same apply for Q4 2016 when 14/16nm products will be far more advanced than both 28nm and 20nm. So again those 14/16nm products will command higher ASP than the rest at that time.

This is how the industry is working for the last 20-30 years, Intel always degrease its die size of its CPUs but ASP remains the same. AMD and NVIDIA will have to do the same. NVIDIA already did that with Kepler and GK104 two years ago. They introduced a smaller die at higher ASP than before(GF110). They will do the same again at 20nm, they will introduce a smaller die than GM204 at higher ASP.

I dont think you understand the slide at all.

Where do you think 40nm would be on that slide? Even if we scale it back to 40, 28 and 20.