Discussion What if Intel skipped their internal shrink node factories and went with TSMC?

[CentroX]

Is this possible? Just to catch up with AMD. How long can their internal process struggle before they take action?

 

[Topweasel]

Is this possible? Just to catch up with AMD. How long can their internal process struggle before they take action?

Not realistically. TSMC has fantastic capacity but it still doesn't really compete with Intel, not in a way that would allow TSMC swallow up demand of Intel on top of other clients. Best case scenario is that Intel has a die or two they really need to offload production to (which apparently they are doing with Samsung). The problem you run into is that even more then the 3 fabs or so AMD had when they spun of GloFo. Intel is dependent on their fabs as source of profit. They can't move production away from them without taking a horrible hit on their bottom line and unless they can simultaneously sell off in parts (because as a whole its an albatross) all at once then the impact would be too great. On top of all of that, think of the time it took for Intel to actually back port 10nm designs to 14nm (Rocket Lake). That's when moving from an Intel process to an Intel process.

On top of all of that is that Intel is currently using all their fab space and then some and still is coming up short. They invested extra billions into a tech they want to move away from to be able to I think next year start to catch up with demand a little bit. Meaning as bad as it looks, Intel can still just go about their business with little financial impact without making any major moves like that.

 

[moinmoin]

TSMC has fantastic capacity but it still doesn't really compete with Intel, not in a way that would allow TSMC swallow up demand of Intel on top of other clients.

First mover clients (like e.g. Apple) essentially pay for the capacity TSMC builds up each gen to be sufficient for their purpose. If Intel were to use TSMC they'd definitely try to be a first mover. Thus the capacity issue is a completely non-issue in that context.

Other than that I agree, Intel as of its internal structure right now would simply not be prepared for such a step even if they wanted to.

 

[yeshua]

Is this possible? Just to catch up with AMD. How long can their internal process struggle before they take action?

Reason number one is that TSMS/Samsung will have to build a dozen of factories to satisfy Intel's needs/demand and if Intel changes its mind, those investments in factories (to the tune of a dozen of billion dollars ... each) will be wasted.

Reason number two is that CPUs blueprints are huge trade secrets and there's no way Intel will share them with anyone, especially the Chinese.

Reason number three, you can develop a new node a lot more effectively if you can experiment on your own factory. Imagine how much time will be wasted each time a new specification is sent overseas, then shipped back, then tested (that's actually how NVIDIA/AMD/ARM operate) - we are talking about months for each new design/validation.

See this video for insight (a mist see for everyone who loves technology):

 

[teejee]

Is this possible? Just to catch up with AMD. How long can their internal process struggle before they take action?

I would expect TSMC to require a contract with approximate volumes at least three years ahead to handle Intels volumes (if at all possible for TSMC to handle). So TSMC will never be a quick fix. I’m pretty sure Intel will continue with their own production for at least a decade or more.

 

[moinmoin]

Reason number two is that CPUs blueprints are huge trade secrets and there's no way Intel will share them with anyone, especially the Chinese.

You may want to look up what the T in TSMC stands for.

Reason number three, you can develop a new node a lot more effectively if you can experiment on your own factory. Imagine how much time will be wasted each time a new specification is sent overseas, then shipped back, then tested (that's actually how NVIDIA/AMD/ARM operate) - we are talking about months for each new design/validation.

Is the current Intel a proof of this? 🤔

 

[Topweasel]

First mover clients (like e.g. Apple) essentially pay for the capacity TSMC builds up each gen to be sufficient for their purpose. If Intel were to use TSMC they'd definitely try to be a first mover. Thus the capacity issue is a completely non-issue in that context.

Other than that I agree, Intel as of its internal structure right now would simply not be prepared for such a step even if they wanted to.

Yeah but you are talking about a demand that might not be enough with their manufacturing capability. Then you are basically pushing out your entire customer base for a single contract. Then thirdendarily you have the Intel issue. Fab development and CPU development basically going hand in hand. TSMC would basically have to hit the brakes on process development and implementation because Intel would need to be ready for it, because as their only real customer no use in starting 3nm testing and high risk limited production if they weren't ready to accept the die losses. Least with the current status quo you have AMD/Apple/Qualcomm/Nvidia to sell the new process to, if one or two decide they can or want to move forward with it they can start production and utilize their new Fabs asap.

 

[yeshua]

You may want to look up what the T in TSMC stands for.

You may want to educate yourself first: "Apart from its main base of operations in Hsinchu in Northern Taiwan, where several of its fab facilities are located, it also has leading-edge fabs in Southern Taiwan and Central Taiwan, with other fabs located at its subsidiaries TSMC China in Shanghai, China, WaferTech in Washington State, USA, and SSMC in Singapore,[42] and it has offices in China, Europe, India, Japan, North America, and South Korea.[43]"

 

[moinmoin]

Yeah but you are talking about a demand that might not be enough with their manufacturing capability. Then you are basically pushing out your entire customer base for a single contract. Then thirdendarily you have the Intel issue. Fab development and CPU development basically going hand in hand. TSMC would basically have to hit the brakes on process development and implementation because Intel would need to be ready for it, because as their only real customer no use in starting 3nm testing and high risk limited production if they weren't ready to accept the die losses. Least with the current status quo you have AMD/Apple/Qualcomm/Nvidia to sell the new process to, if one or two decide they can or want to move forward with it they can start production and utilize their new Fabs asap.

TSMC doesn't convert existing fabs to newer nodes the way Intel does. TSMC mostly builds new fabs and expands existing ones for all their new nodes, increasing overall capacity all the time. So the "pushing out your entire customer base" is not what would happen if Intel were to approach TSMC. Instead an insane amount of new fabs would be built, mostly paid by Intel as a first mover. And all the other customers would get access to those fabs as well whenever Intel then doesn't happen to manage to fill those fabs on its own.

You may want to educate yourself first: "Apart from its main base of operations in Hsinchu in Northern Taiwan, where several of its fab facilities are located, it also has leading-edge fabs in Southern Taiwan and Central Taiwan, with other fabs located at its subsidiaries TSMC China in Shanghai, China, WaferTech in Washington State, USA, and SSMC in Singapore,[42] and it has offices in China, Europe, India, Japan, North America, and South Korea.[43]"

Wow, TSMC is an international company that happens to have a Chinese subsidiary. Still doesn't make them Chinese. Or maybe you meant to say they are also American since they have a subsidiary in Washington State as well?

 

[yeshua]

Wow, TSMC is an international company that happens to have a Chinese subsidiary. Still doesn't make them Chinese. Or maybe you meant to say they are also American since they have a subsidiary in Washington State as well?

There are certainly a Taiwanese company first and most but that doesn't mean they are immune to not leaking trade secrets specially if they operate in China and have staff in the country. I have a suspicion multinational companies are easier to exploit.

 

[Topweasel]

TSMC doesn't convert existing fabs to newer nodes the way Intel does. TSMC mostly builds new fabs and expands existing ones for all their new nodes, increasing overall capacity all the time. So the "pushing out your entire customer base" is not what would happen if Intel were to approach TSMC. Instead an insane amount of new fabs would be built, mostly paid by Intel as a first mover. And all the other customers would get access to those fabs as well whenever Intel then doesn't happen to manage to fill those fabs on its own.

I get what you are saying there. But they are already 3 major nodes deeper then Intel with the Fab either built (5nm) or being built (3nm). If Intel did go to TMSC and they took them on it would be at the cost of their current clients. The Scenario where Intel gives up will not give them time to wait the 3-5 years for TSMC to expand capacity to keep Intel/Apple/AMD and eventually Nvidia on the same node.

 

[Atari2600]

Would it not be more likely Intel would just license the node from TSMC? (Or Samsung)

Then they can go and build their CPUs in their own fabs.

 

[Topweasel]

Would it not be more likely Intel would just license the node from TSMC? (Or Samsung)

Then they can go and build their CPUs in their own fabs.

Thinking that falls in line with by the time they make the move it would take to long. It might not be from scratch, but most of the other fabs have offshoots of IBM processes from one point or another. So a cross license from one to another might not involve to much investment. Not sure that applies to Intel their setup is much much different from TSMC.

 

[moinmoin]

I get what you are saying there. But they are already 3 major nodes deeper then Intel with the Fab either built (5nm) or being built (3nm). If Intel did go to TMSC and they took them on it would be at the cost of their current clients. The Scenario where Intel gives up will not give them time to wait the 3-5 years for TSMC to expand capacity to keep Intel/Apple/AMD and eventually Nvidia on the same node.

We are talking pure fantasy hypotheticals anyway. Sure TSMC would need time to build the capacity, though Intel would need to prepare for TSMC's nodes as well so I feel like it's a toss up who will take longer on their job.

But it seems Intel found its magic fairy solutions to everything already, they are again talking about going back to creating a new node every two years. *shrugs*

(source)

 

[jamescox]

I get what you are saying there. But they are already 3 major nodes deeper then Intel with the Fab either built (5nm) or being built (3nm). If Intel did go to TMSC and they took them on it would be at the cost of their current clients. The Scenario where Intel gives up will not give them time to wait the 3-5 years for TSMC to expand capacity to keep Intel/Apple/AMD and eventually Nvidia on the same node.

It isn’t necessarily the case that TSMC is that many nodes ahead of Intel since the numbers (like 14 nm, 7 nm, etc) are completely meaningless from both companies; they are really just names. You would need to compare specific specs from both processes to see how they compare. intel had the process lead for a long time when competing with AMD/Global Foundries. They aren’t going to have that going forward since TSMC has a massive portion of the market. They have the money and resources to push the process tech, so I don’t really expect intel to take the lead again. They may have some advantages in doing being both the designer and the fab, but TSMC has just a massive portion of the market, so that may not be a big enough advantage going forward.

Intel is unlikely to try to have their processors made at TSMC. It would take a long time to port their design, probably longer than a fabless company since they would be trying to move from an internal process to working with an external company. AMD had their own fab, but they had some experience using an external fab from ATI, since they were fabless.

In addition to possibly years to port the design to an external fab, they would also then be paying for and competing with other vendors for building out more fab space. It isn’t a quick fix at all, but at this point, I have to wonder what is going on in intel. Even if they just abandoned 10 nm, their 7 nm process efforts should have been going on in parallel, so where is it? I guess they may have been converting fab space rather than expanding. In that case, they may have an lot of space dedicated to 10 and 7 nm, which isn’t yielding yet. They also need to make a bunch of higher core count processors (larger die) on 14 nm that they probably never planned to do. The back port of a 10 nm processor design to 14 nm seems to indicate that things still aren’t going well unless that is just a backup plan.

 

[scannall]

We are talking pure fantasy hypotheticals anyway. Sure TSMC would need time to build the capacity, though Intel would need to prepare for TSMC's nodes as well so I feel like it's a toss up who will take longer on their job.

But it seems Intel found its magic fairy solutions to everything already, they are again talking about going back to creating a new node every two years. *shrugs*

(source)

I think the last node that was actually on time from them was 32nm. Or was it 45? Like all marketing slides, it can be amusing for discussion but not to be taken seriously. It only counts when parts are out the door, on shelves and in reviewers hands. (For any company, not just Intel)

 

[moinmoin]

I think the last node that was actually on time from them was 32nm. Or was it 45? Like all marketing slides, it can be amusing for discussion but not to be taken seriously. It only counts when parts are out the door, on shelves and in reviewers hands. (For any company, not just Intel)

Year wise even 22nm was on time.
I don't know. I honestly feel trolled by Intel. Technically we are 4 years into the 10nm node and desktop and server chips are still nowhere to see. 7nm is supposedly the magical unicorn that will help Intel match the frequencies of 14nm++++++ again. And here is Intel announcing the return of their traditional biannual new node introduction, as if all the previous troubles are already solved and never happened.

 

[jpiniero]

It's possible, and if Intel falls far enough behind you will see it. Intel is already a customer of TSMC's already, so they know what it takes to port, although I don't think they have anything on a bleeding edge node there right now.

Intel is headed to using chiplets for all of it's products like AMD so it's not like 100% of the product would need to be at TSMC (or Samsung), they could fab some chiplets at their existing fabs on old nodes if they thought they could get away with it.

Agree it's very unlikely Intel returns to a two year cadence. I don't even think Intel really believes that but they need to try to convince Wall Street.

 

[Roland00Address]

First mover clients (like e.g. Apple) essentially pay for the capacity TSMC builds up each gen to be sufficient for their purpose. If Intel were to use TSMC they'd definitely try to be a first mover. Thus the capacity issue is a completely non-issue in that context.

Apple has way more money and can win a bidding war. But do not take this literally what I just said, I want you to unfocus.

Intel has money, but it must constantly sink money in its Fabs to preserve existing assests. If Intel Fabs are not being constantly updated they lose value. Now in theory Intel can suddenly have all these customers for existing fabs and thus they can make money being a foundry that other people use but Intel's business is not set up that way with creating the tools for other people to use, the sales plus engineers that help 3rd parties actually use their foundries and so on.

Thus I am saying Intel has money, but having Fabs and also going 3rd party Fabs is wasteful spending. And if Intel is going to switch its existing Fabs to 3rd party it needs to "waste" money hiring all these people and all that money is wasted till you actually get the customers who spend the money in order to turn the liability into a revenue generator. Intel is not structured currently to pivot this way in a way that they could thrive.

------

Note AMD was not structured this way to suddenly thrive when it and Global Foundries did its first break up in 2009. It took 10 years for AMD to get healthy on its side (though it came close a couple years earlier) and it is not clear if Global Foundries is thriving or just treading water. [And compared to TSMC and Samsung, GF is definitely not thriving if you do a comparison of similar people it its fields, which is a different way to compare than straight profits vs losses.]

------

Long term it may be in Intels interest to abandon its Fabs, but in the short and medium term it would be PAIN, lots and lots of PAIN. If Intel knew in 2015 what its 2019 situation was going to be then maybe pivoting then would have made sense, then again maybe they could have not be in this mess in the first place with a better foundry process if they had this future knowledge, or they would have backported some designs way earlier, and so on.

 

[Adonisds]

We are talking pure fantasy hypotheticals anyway. Sure TSMC would need time to build the capacity, though Intel would need to prepare for TSMC's nodes as well so I feel like it's a toss up who will take longer on their job.

But it seems Intel found its magic fairy solutions to everything already, they are again talking about going back to creating a new node every two years. *shrugs*

(source)

My roadmap shows I'll become a trillionaire by age 30

 

[VirtualLarry]

Intel is unlikely to try to have their processors made at TSMC. It would take a long time to port their design, probably longer than a fabless company since they would be trying to move from an internal process to working with an external company. AMD had their own fab, but they had some experience using an external fab from ATI, since they were fabless.

Wasn't Brazos (for AMD) a synthesized design, that was produced on external fabs at the time? I seem to recall those chips being dual-sourced?

 

[ThatBuzzkiller]

Fab capacity isn't going to be much of a concern especially not if TSMC has a far more advanced logic fabrication technology. Right now TSMC's 7nm process is twice as dense as Intel's 14nm and if Intel are still producing the vast majority of it's highest end parts on 14nm while TSMC and are imminently transitioning to 5nm then the latter will be 4 times as dense as Intel in that scenario!

What Intel are producing now on 14nm would only take half as much die space on TSMC's 7nm and thus a fraction of the wafer capacity as well but doubly so for TSMC's 5nm ...

If Intel doesn't show some results soon for 7nm then it's only pragmatic that Intel should start transitioning their designs to TSMC because their main competitor will happily take advantage of a larger transistor budget ...

 

[CHADBOGA]

Fab capacity isn't going to be much of a concern especially not if TSMC has a far more advanced logic fabrication technology. Right now TSMC's 7nm process is twice as dense as Intel's 14nm and if Intel are still producing the vast majority of it's highest end parts on 14nm while TSMC and are imminently transitioning to 5nm then the latter will be 4 times as dense as Intel in that scenario!

What Intel are producing now on 14nm would only take half as much die space on TSMC's 7nm and thus a fraction of the wafer capacity as well but doubly so for TSMC's 5nm ...

If Intel doesn't show some results soon for 7nm then it's only pragmatic that Intel should start transitioning their designs to TSMC because their main competitor will happily take advantage of a larger transistor budget ...

How long would it take for TSMC to increase their capacity enough to handle what Intel would require?

12 months? 18 months? 24 months? 36 months?

 

[NostaSeronx]

With dGP at a place with good enough Capex, I doubt anyone would be competing with Intel 5nm and beyond.

 

[ThatBuzzkiller]

How long would it take for TSMC to increase their capacity enough to handle what Intel would require?

12 months? 18 months? 24 months? 36 months?

I cannot say with absolute certainty since demand for high-end logic fabrication can change for specific markets. It could be faster if there's a slowdown in high-end chips or it could very well be slower if demand for high-end chips increases. TSMC extending their technological lead just makes things easier ...

The beauty of a 'transition' is that Intel doesn't have to immediately shutdown all of their facilities, they can instead make a gradual move to fabless by shutting down their plants on a periodic basis once their ready. Intel also doesn't have to solely source from TSMC, if Samsung also manages to beat Intel at their own game then they can also source from them as well. Intel can always decide to manufacture their higher-+end designs at TSMC or Samsung while Intel's own fabs can manufacture their lower-end designs ...