Apple-Intel chip deal makes strategic sense but production is years away

[Daniel Nenni]

Intel 14A vs. TSMC A14 — public PPA comparison​

 

[siliconbruh999]

14A is de-risked because PDF Solutions is partnering with Intel on yield. Still, Intel does not have the volume to support Apple's full business at 18A or 14A. Remember, 18A is behind TSMC N2 in PPAC and yield/risk which is important to the top tier. Apple is full on N2 this year and next and may stick with N2 variations for an extra year or two. I do not see Apple using TSMC A16.

Intel needs to add few more plus to make it PPA competitive like they did with 14nm/10nm

 

[Daniel Nenni]

Intel needs to add few more plus to make it PPA competitive like they did with 14nm/10nm

I think Intel 14A and TSMC A14 are comparable enough for the NOT TSMC Market. My concern is the release date. TSMC A14 is scheduled for HVM in 2028 and Intel 14A is 2029. The other issue is trust. When TSMC says HVM in 2028 they mean HVM in 2028. Hopefully Lip-Bu can pull 14A in and customers will trust him to deliver.

 

[siliconbruh999]

I think Intel 14A and TSMC A14 are comparable enough for the NOT TSMC Market. My concern is the release date. TSMC A14 is scheduled for HVM in 2028 and Intel 14A is 2029. The other issue is trust. When TSMC says HVM in 2028 they mean HVM in 2028. Hopefully Lip-Bu can pull 14A in and customers will trust him to deliver.

for Intel Product most of the stuff is either 18A or it's variant or N2. 14A has to be carried by external it seems for Intel also DZ/LBT have said multiple times it's 28-29 time frame time would tell where 14A HVM will land.

 

[IanD]

Apple will not use A16 for the reasons previously discussed, which is that the PPA improvements quoted by Daniel (compared to N2) are for high-power chips with a dense power grid, and that active (liquid?) cooling is recommended to keep die temperature down (because of worse SHE) -- which is why TSMC recommend its use for "HPC applications with active cooling", which fits with Nvidia (only A16 customer?) but not Apple... ;-)

Oh yes, don't forget higher die cost too because A16 wafer price increase is a lot bigger than die size decrease. Which doesn't really bother Nvidia given their super-high margins, but would bother Apple.

 

[my_wing]

You're missing the point here -- the "derisking" was for foundry customers, not the raw BSPD process/transistors.

Customers have many years of experience with FSPD process and tools and libraries and how to use/optimise a design with them. With BSPD the layout and tools and libraries are new and unfamiliar, and there are new/different things to worry about like worse SHE. And the other key issue is that a BSPD design is completely incompatible with a FSPD process, so once you've gone down that route for a particular device there's no going back -- the effort to switch BSPD to FSPD would be at least as big (probably bigger) than moving from FSPD to FSPD+1 nodes.

So you're putting all your eggs into the BSPD basket at one particular foundry, and if there are any problems -- whether it's with the process or tools/timing closure or SHE or anything else -- which lead to delays or reduced yield, then you've got no way out and your product could be too late to market or uncompetitive, and with absolutely no way to even switch to another foundry.

People seem to think that the technology is everything and better PPA is all that matters, but in the real world that isn't the case, there are many other factors that increase risk with either a radically new process (which BSPD is, and don't let anyone tell you otherwise!) or the tools/libraries needed to support it, including delay/yield problems. Those are the risks I was talking about, not "my transistor's better than your transistor"... ;-)

And right now there's close to zero real-world experience with BSPD outside Intel 18A (who are not exactly transparent about process problems, are they?) and Nvidia (the only A16 customer, presumably with no MP yet).

Intel's problem is they're caught between a rock and a hard place -- if they just do something similar to TSMC (FSPD) then why would anyone use them, given that they'll have no differentiator except being American -- and almost certainly later and more expensive and with hardly any ecosystem and no convincing high-yield record.

So they had to do something different/better to stand a chance of attracting customers, and they chose BSPD -- which also fits in with their traditional "we have the best technology" theme.

Same with high-NA EUV -- Intel are pushing it because it's the latest sexy technology and they can use it to promote the same theme, TSMC are holding back because it doesn't make commercial sense *for them*.

TSMC in itself is a big risk, price increase is the number 1 risk that customer has, and the risk is not from somewhere that you are thinking, and TSMC is focus to transfer this risk to customer as well.

The actual location of Risk :
The risk actually lives with me everyday, inside our petrol station that I use everyday, i.e. Australian Coal and LNG --- Energy. While Intel risk in this area is not as much, maybe the AI data center but not as much.
Let say Ireland, with 40% Wind 40 LNG (plus close to production i.e. North Sea -- if UK can be wake up by reform party), wind is the 2nd cheapest form of electricity (just behind solar - if you are located in Australia but it is Ireland)
Let say USA, major LNG producer, with Solar, Wind and good mix of electricity, increase capacity is not as expensive as Taiwan.

To de-risk from a customer point of view, how many times TSMC increase its price for the last few years (I remember that N3E is 30% more expensive than N4/5 then now another ~10%, N2 is reported 30% more than N3E, so in general, 1 * 1.3 * 1.3 * 1.1 = 1.86 times at least more expensive than N5/4 for N2 class, it is also report that Intel 18AP can be as much as 25% discount on N2 Class price.

Don't mix Left and Right here, the republicans and it's heart land (Texas) is around 35% solar and wind vs Taiwan (mostly believed to be left) ~10% (mostly offshore wind i.e. expensive). It is all about cost, nothing about the environment, the fact that the US is so large also invite innovative solutions i.e. HVDC to the picture to further reduce cost and improve stability.

Cost is actually Intel big advantage and TSMC biggest risk. If you put all the eggs into one basket, if the oil price and if Australia fail to provide (lower risk if that $20B offshore facility is blow up by terrorist, but price risk is high) meaning that depends on Australian Energy sector, this is also a problem. Correct me if I am wrong, but as an expert in costing / cost accounting I can tell you this even as an outsider, electricity is like 50% of OPEX, and ~25% of IFRS accepted cost allocation (ABC, Standard cost, etc.)

Intel's do not have a problem, it has an image problem, in fact why AMD existed is that there is an initial agreement that AMD can produce X86, they think Intel can steal their tech. But that is the image problem.
High NA is really a problem, but this is a symptom, not the actual problem, in my point of view, with all the Quad patterning issues we face, it must be something that is hidden becoming a problem, intel 14nm tried it, TSMC N3B tried it all not okay, I think CAR photoresist is a bottleneck, together with things like pattern sharpening, none of these, none got implemented into TSMC system. That is the true problem, the true problem is not High NA, it is the inability or complexity of Quad patterning.

BSPD, pmos FINEFT, 3 layers GAA, are another point closely related to the fact of lack of innovation. That is the true risk of TSMC. You still want to do CFET or not TSMC ??? If you are not implementing all GAA now, how are you going to gain experience and jump to CFET, 3 layers GAA, why 3, it is not balance, on the other hand intel is doing 4 which is logical 2-2 nmos-pmos (seamless transfer to CFET). The lack of / refuse to innovate is TSMC major problems, in that sense, I will trust Intel to provide me with that innovation, since 8086, there are countless tech that Intel is successfully did. TSMC innovation is just very slow, what they did is pure foundry model, implemented EUV, that is it, the major of the semi-industry life is Intel is always on top, I think people should realised that letting Intel do it's job and at least you can be the AMD, to stay with TSMC, is like a dead end.

 

[hist78]

I think Intel 14A and TSMC A14 are comparable enough for the NOT TSMC Market. My concern is the release date. TSMC A14 is scheduled for HVM in 2028 and Intel 14A is 2029. The other issue is trust. When TSMC says HVM in 2028 they mean HVM in 2028. Hopefully Lip-Bu can pull 14A in and customers will trust him to deliver.

Another challenge is capacity and financial limitations. Assuming 14A is good, Intel must buy those expensive equipment and build/complete fab facilities as soon as possible in order to meet the 2028/2029 target. Without building enough capacity, Intel Foundry's potential customers, either internal or external, can't afford to make a commitment.

Unfortunately, Intel's financial position is not strong enough to take on more risks. I'm thinking that, in order to manage the risks, Lip-Bu Tan is following the same approach that Pat Gelsinger and Bob Swan had taken before him: outsourcing significant amounts of Intel Products' wafer needs to TSMC. Effectively, it could allow Intel Foundry to take on more external customers without jeopardizing Intel Products' on-time product delivery.