Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/rumor-tsmc-expected-to-produce-5nm-intel-cpus-in-h1-2022.13006/page-2
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Rumor: TSMC Expected to Produce 5nm Intel CPUs in H1 2022

Definitely sounds like Intel is going fabless. It's not going to be an easy transition but I think it's the only path forward. They simply cannot afford to keep building cutting edge fabs without the chip volumes. They will fall behind AMD for a few cycles, sure it won't kill Intel as long as they can recover architectural leadership.
 
Intel 7nm is TSMC 5nm. Porting designs from one foundry to another is not going to be competitive with chips from AMD that are designed only for TSMC. Hopefully he is misspeaking.

He knew what he's talking about. He just wanted to be polite and didn't want to hurt people's feeling. In his mind, Intel must go big with third party foundries in order to compete with everyone else in the semiconductor industry. He believes there is a short time widow allows him to grab. If he misses it, or if Intel misses it, the consequence will be very dangerous.

The outsourcing decision has been made. And it probably was decided even before last Q2 2020 earning conference call.
 
To port CPU to foundry would not be a just-kick action. Before executives making decision in early 2021, some IP validation tape-out could be launched 3-4 quarters ahead and be finished to wait green light of HVM in foundry then. It will be interesting to see the final decision and balance of IDM, Fablite and fabless scenarios in intel transition.
 
Ironically the people that hired swan and krazinich were let go because of them.
 
Everyone wants to idm system to continue. Continuing to manufacture in Oregon and Arizona would be great for everyone. There's a people problem here.
 
Ironically the people that hired swan and krazinich were let go because of them.

That is not true.

Andy Bryant and the Intel board hired BK which was clearly a mistake. Looking back, BK was just an Andy Bryant puppet. They fired BK for having a inappropriate relationship which was complete nonsense. The Intel board knew about that relationship before he was hired. Bottom line: BK took the fall for Andy's incompetence.

Intel board member (appointed in 2017) Omar Ishrak used that failure to force Andy's retirement, get the CoB position, and have Bob Swan to run things for him. Omar has an EE PhD so he is the technical brains behind Intel and Bob is his puppet per say. The rest of the Intel board is no match for Omar so this really is his show.
 
There's a rumor going on that Raja Koduri's (Intel Senior Vice President & Chief Architect) appearance at Samsung Foundry's SAFE Forum on October 28, 2020 could signal that Intel maybe moving to Samsung Foundry rather than TSMC. That would certainly be a surprise as:
1. I didn't view Samsung's process to be superior than Intel's. If Intel does use Samsung...does this signal Intel tacit admission that BOTH TSMC AND Samsung has surpassed its process technology?

2. Swan mentioned the importance of providing a "predictable cadence of leadership products" for Intel customers. I'm assuming Swan is also balancing wafer pricing and foundry availability here as TSMC currently has the best process technology. The question for Intel is...does it really want to compete with AMD using Samsung's 4nm (or possibly 3nm GAAFET process) vs TSMC's 3nm?

Getting Intel's business would be a HUGE win for Samsung. However using Samsung may be a HUGE risk for Intel...imagine telling your customers that demand exceeds supply and that most of them wouldn't be able to get their hands on your product until next year (like what Jensen Huang/Nvidia is saying right now!). That's almost as bad as telling everyone that your advanced process node just got delayed AGAIN (and again)! ;)
 
Last edited:
That is a juicy piece of gossip. It really is bad form in my opinion. I said the same thing about the Renee James (former Intel President) keynote at the TSMC Symposium. She really stuck it to Intel on that one. Maybe this is payback for that?

As I mentioned before, If you use both TSMC and Samsung you will not be in the inner circle of TSMC which would put Intel at a disadvantage competing with AMD. You also would not get the best wafer prices from TSMC.

On the technical side, yes Samsung has had delivery problems. At 14nm Samsung did a great job, even better than TSMC 16nm at the start but at 10nm Samsung did not yield and their solution was to start selling good die versus wafers which left QCOM with shortages. Samsung is so big they can absorb the yield losses but TSMC cannot. TSMC investors would not allow it.

I would call Samsung's processes comparable to TSMC's but remember process is just the start, you must have the design and IP ecosystem to support it and that is where TSMC dominates.

And I agree that Intel should be risk adverse when looking at 2nd sourcing since they have had such a problem with their first source. :rolleyes:


There's a rumor going on that Raja Koduri's (Intel Senior Vice President & Chief Architect) appearance at Samsung Foundry's SAFE Forum on October 22, 2020 could signal that Intel maybe moving to Samsung Foundry rather than TSMC. That would certainly be a surprise as:
1. I didn't view Samsung's process to be superior than Intel's. If Intel does use Samsung...does this signal Intel tacit admission that BOTH TSMC AND Samsung has surpassed its process technology?

2. Swan mentioned the importance of providing a "predictable cadence of leadership products" for Intel customers. I'm assuming Swan is also balancing wafer pricing and foundry availability here as TSMC currently has the best process technology. The question for Intel is...does it really want to compete with AMD using Samsung's 4nm (or possibly 3nm GAAFET process) vs TSMC's 3nm?

Getting Intel's business would be a HUGE win for Samsung. However using Samsung may be a HUGE risk for Intel...imagine telling your customers that demand exceeds supply and that most of them wouldn't be able to get their hands on your product until next year (like what Jensen Huang/Nvidia is saying right now!). That's almost as bad as telling everyone that your advanced process node just got delayed AGAIN (and again)! ;)
 
Intel 7nm is TSMC 5nm. Porting designs from one foundry to another is not going to be competitive with chips from AMD that are designed only for TSMC. Hopefully he is misspeaking.

If designs have been done with foundry-style libraries and DKs (like AMD at TSMC) then it's not actually that difficult porting them between foundries (I've done it), you do have to do re-optimisation and won't end up with identical performance, but this is usually due to one process/library being better than the other not the porting itself -- which still does take a lot of time and effort for a big SoC which is what CPUs are nowadays.

The problem is if your chip has a lot of "CPU-style" hand-crafted custom process-tweaked stuff in it (like domino logic or dynamic latches or non-CMOS data buses) or your libraries and design flow are not "foundry-style", and having had inhouse fabs for years with customised CPU designs tightly tied to them this is the biggest problem for Intel in moving their high-end CPUs to a foundry.
 
I could see Intel suits deciding to go with Samsung because "AMD is on TSMC and we need to differentiate" or "We want leverage over our suppliers and TSMC is already too powerful" or some other similarly backwards business logic. It would be a disastrously wrong decision, but those are the types of stupid decisions the MBAs at Intel are trained to make.
 
I think parts of Intel will exist 5 years from now and parts won't.
 
I could see Intel suits deciding to go with Samsung because "AMD is on TSMC and we need to differentiate" or "We want leverage over our suppliers and TSMC is already too powerful" or some other similarly backwards business logic. It would be a disastrously wrong decision, but those are the types of stupid decisions the MBAs at Intel are trained to make.
Difficult to say for sure if it's stupid or not...

If Intel go to TSMC then they will get what is probably the best process, but they won't have a controlling influence over process/libraries because TSMC have many other big existing customers (Apple, Qualcomm, AMD, Xilinx...) who are already "insiders" -- yes Intel is big, but the others added together are much bigger for wafer volumes and won't let TSMC be dragged around by Intel (and TSMC won't want to give Intel anything better than Apple and hack off their biggest customer).

If they go to Samsung then maybe the process isn't quite as good (or high-yielding) as TSMC but Intel would have much more clout to get the process/libraries tweaked to optimise for their needs -- Samsung Foundry has far fewer big powerful external customers than TSMC and Intel would be a hugely prestigious catch for them, they'd probably bend over backwards (or forwards...) to be helpful which TSMC are much less likely to do.

So it depends which is best for Intel, a slightly better process they can't tweak or a slightly worse one that they can. Yield is another area where TSMC are ahead of Samsung but Intel/AMD CPUs are high-margin products which aren't so sensitive to this -- and let's face it, Samsung are unlikely to be as bad for yield as Intel's own fabs have been with 10nm ;-)

It's not as cut and dried as count makes out...
 
Difficult to say for sure if it's stupid or not...

If Intel go to TSMC then they will get what is probably the best process, but they won't have a controlling influence over process/libraries because TSMC have many other big existing customers (Apple, Qualcomm, AMD, Xilinx...) who are already "insiders" -- yes Intel is big, but the others added together are much bigger for wafer volumes and won't let TSMC be dragged around by Intel (and TSMC won't want to give Intel anything better than Apple and hack off their biggest customer).

If they go to Samsung then maybe the process isn't quite as good (or high-yielding) as TSMC but Intel would have much more clout to get the process/libraries tweaked to optimise for their needs -- Samsung Foundry has far fewer big powerful external customers than TSMC and Intel would be a hugely prestigious catch for them, they'd probably bend over backwards (or forwards...) to be helpful which TSMC are much less likely to do.

So it depends which is best for Intel, a slightly better process they can't tweak or a slightly worse one that they can. Yield is another area where TSMC are ahead of Samsung but Intel/AMD CPUs are high-margin products which aren't so sensitive to this -- and let's face it, Samsung are unlikely to be as bad for yield as Intel's own fabs have been with 10nm ;-)

It's not as cut and dried as count makes out...
There are many customer engagement models in foundry field, like JDP, Phase-in and others, especially for "Specialty" technologies. It is give-and-take and how much you will pay for the outsourcing and how much foundry to keep their margin. Will that be wafer-buy or Die-buy? Premium for capacity guarantee.... Intel have tsmc as foundry for more than 30 years and I expect they are experts in foundry playground,
 
Ian,

We are talking about Intel porting designs from Intel fabs to TSMC at the advanced nodes which now include EUV. Do you believe that moving a CPU/GPU design from Intel 7nm to TSMC 5nm will be technically feasible? Or do you think Intel will have to design directly to TSMC like AMD does to be competitive?

D.A.N.

QUOTE="IanD, post: 44295, member: 6522"]
If designs have been done with foundry-style libraries and DKs (like AMD at TSMC) then it's not actually that difficult porting them between foundries (I've done it), you do have to do re-optimisation and won't end up with identical performance, but this is usually due to one process/library being better than the other not the porting itself -- which still does take a lot of time and effort for a big SoC which is what CPUs are nowadays.

The problem is if your chip has a lot of "CPU-style" hand-crafted custom process-tweaked stuff in it (like domino logic or dynamic latches or non-CMOS data buses) or your libraries and design flow are not "foundry-style", and having had inhouse fabs for years with customised CPU designs tightly tied to them this is the biggest problem for Intel in moving their high-end CPUs to a foundry.
[/QUOTE]
 
Ian,

We are talking about Intel porting designs from Intel fabs to TSMC at the advanced nodes which now include EUV. Do you believe that moving a CPU/GPU design from Intel 7nm to TSMC 5nm will be technically feasible? Or do you think Intel will have to design directly to TSMC like AMD does to be competitive?

D.A.N.

QUOTE="IanD, post: 44295, member: 6522"]
If designs have been done with foundry-style libraries and DKs (like AMD at TSMC) then it's not actually that difficult porting them between foundries (I've done it), you do have to do re-optimisation and won't end up with identical performance, but this is usually due to one process/library being better than the other not the porting itself -- which still does take a lot of time and effort for a big SoC which is what CPUs are nowadays.

The problem is if your chip has a lot of "CPU-style" hand-crafted custom process-tweaked stuff in it (like domino logic or dynamic latches or non-CMOS data buses) or your libraries and design flow are not "foundry-style", and having had inhouse fabs for years with customised CPU designs tightly tied to them this is the biggest problem for Intel in moving their high-end CPUs to a foundry.
[/QUOTE]

Moving a CPU design between "similar" nodes (e.g. Intel 7nm to TSMC 5nm) is always technically feasible, it all depends how much time/money/effort you're willing to spend re-optimising the design. Done properly there's no reason the end result should be inferior than designing for TSMC in the first place, but this also means selecting the correct process options and deciding on this can be a lot of work in itself -- it's what AMD would have done designing into TSMC, Intel will have to do it all over again to port from their own fab to TSMC.

Multi-patterned non-EUV to EUV porting can be a real pain because the design rules change in non-obvious ways, and not always for the better. Probably not as much of a pain as a complete full-custom analogue macro library port from 14nm FDSOI to 16nm FinFET though, and I've been there...

The end result will depend on just what the process differences are and whether the TSMC process makes the same CPU-centric choices as the Intel one, for example number and thickness of metal layers, number of different Vth available and so on. TSMC processes used to be more focused at mobile SoC but nowadays (7nm, 5nm...) they have features available which are more targeted at CPUs -- actually there isn't one process at a given node, there are multiple variants with different cost/performance/density targets, as well as different cell libraries targeted at mobile/HPC.

What could be a bigger problem for Intel is that they've traditionally done a lot of full-custom circuit design to squeeze every last drop of performance out of their CPUs (the "IBM approach"), and this needs very close interaction with the process team if any non-standard circuits/layouts/structures are used, especially with the super-restrictive design rules nowadays, I don't know for sure if this is still the case but talking to ex-Intel designers in the past it certainly used to be.

Allowing layout changes right down to the design rule level even for Intel is not something TSMC are likely to be keen on because it can break the fab -- OK for Intel when it's their own fab and all their problem (and they can try and hide it like for 10nm) if it goes wrong, much more difficult with a foundry like TSMC who play it safe and will be less keen to play with the process recipe and design rules just for Intel. If multiple big customers want a new process feature/rule then this is different, some of the 5nm process options are targeted at applications like high-speed SERDES and CPU but they still take TSMC a massive effort to bring up/characterize/debug.

This is where Intel might have more clout with Samsung, more chance of getting Intel-specific process/design rule tweaks than with TSMC. Or maybe not, Samsung could always say "Take it or leave it"...
 
Difficult to say for sure if it's stupid or not...

If Intel go to TSMC then they will get what is probably the best process, but they won't have a controlling influence over process/libraries because TSMC have many other big existing customers (Apple, Qualcomm, AMD, Xilinx...) who are already "insiders" -- yes Intel is big, but the others added together are much bigger for wafer volumes and won't let TSMC be dragged around by Intel (and TSMC won't want to give Intel anything better than Apple and hack off their biggest customer).

If they go to Samsung then maybe the process isn't quite as good (or high-yielding) as TSMC but Intel would have much more clout to get the process/libraries tweaked to optimise for their needs -- Samsung Foundry has far fewer big powerful external customers than TSMC and Intel would be a hugely prestigious catch for them, they'd probably bend over backwards (or forwards...) to be helpful which TSMC are much less likely to do.

So it depends which is best for Intel, a slightly better process they can't tweak or a slightly worse one that they can. Yield is another area where TSMC are ahead of Samsung but Intel/AMD CPUs are high-margin products which aren't so sensitive to this -- and let's face it, Samsung are unlikely to be as bad for yield as Intel's own fabs have been with 10nm ;-)

It's not as cut and dried as count makes out...

This is exactly the type of logic that will backfire for Intel. Intel is a big enough customer that the assumption that TSMC wouldn't develop a custom process for them is probably wrong. I believe TSMC has developed custom processes for NVidia, which is smaller than Intel. The assumption that TSMC won't provide a "better" process to Intel than to Apple is also probably incorrect. They compete in different segments, TSMC will almost certainly be happy to tune or customize 5nm process to Intel's needs and it won't compete with Apple because they are designing chips for entirely different purposes. However, the idea that Intel won't have the "clout" to push around TSMC, you're right, they won't. But Intel has developed a monopolistic corporate mindset where they feel like the way to win is through exercising leverage over their suppliers and their customers and for years that's what they have been doing instead of innovating.

Trading of performance for "clout" will guarantee Intel ends up with a worse process than AMD. Intel can win by having the best architecture on a similar process as AMD. If they have a worse process, it's that much more of a gap they would need to overcome in order to compete. Furthermore, Intel trying to exercise clout over Samsung, for example prioritizing Intel orders, could push Samsung's other foundry customers over to TSMC, so Samsung is likely to push back on that. And Samsung is never going to prioritize Intel over it's internal needs.
 
I think we're agreeing here, I didn't mean that Intel would get to push Samsung around (and certainly not TSMC), I meant that they've got more chance of doing it -- but as you say Intel's attitude is also the problem, they're used to being in control of the fab and doing exactly what they want with design and process but this isn't going to work with the foundries, they're in control of the fabs -- and to some extent the libraries and tools, becuase everything is extremely tightly coupled nowadays.

When people talk about "custom processes" nowadays it's not *usually* fundamentally different/new transistors because the cost/time/effort of qualifying these is now so enormous -- which doesn't mean that Apple doesn't get the "turbo" version first, but then this typically becomes part of the standard offering or is given a new name like N5+. It's more usually different libraries or metal stacks optimised for a particular customer/application, which can give just as big or bigger performance improvements than transistor changes. TSMC N5 has 7 different Vth transistors available, 3 or 4 different height libraries, 3 different CPP, and I can't even remember how many different metal stack options, which gives a huge range for "customisation" to suit a particular customer and application -- it's like having a (very expensive!) custom dinner made from a special selection of ingredients from a *very* well-stocked larder.

So Samsung might not prioritise Intel over its internal needs (e.g. mobile SoC) but there's not really a conflict here, the process options you need for ASIC/SoC and HPC/CPU are quite different and Samsung don't have any internal HPC business or external HPC customers to conflict with Intel, so they might see cosying up with them to target an HPC/CPU process as a good thing, as well as good business and a feather in the foundry's cap.

Either way, I'm sure there's plenty of chance for Intel arrogance to screw this up, like they did with their "world's best 10nm process" and "AMD will never amount to anything, we're far superior to them" attitude...
 
You send a few "test engineers", mechanics, technicians or what ever the term is from Israel to Arizona and Oregon and get 10 nm going. It should be a simple thing. There's a people problem here and not everyone is happy.
 
Globalfoundaries are having an IPO. Selling that they may be back in the node size game.
 
Globalfoundaries are having an IPO. Selling that they may be back in the node size game.
Doesn't seem likely, all the reasons they dropped out of 7nm are still there (actually, even bigger) for 5nm and below. The only thing that might change that would be a *massive* cash injection from the US government to back a bleeding-edge US fab and get security of supply, but I'd have thought that if it ever happened it would go to Intel not GF...
 
Back
Top