Speculation: Intel will become fabless

[senttoschool]

Called it back in Feb 2020. It was quite obvious that Intel would eventually use TSMC or Samsung or allow AMD to chip away at its market share.

One of the interesting things that happened over the last 10 years is TSMC utilizing the capital made from mobile to pull ahead of Intel in process development.

In other words, AMD's recent success owes a lot to Apple pumping money into TSMC over the last 10 years.

As a pure chip manufacturer, TSMC has the same market cap as Intel. All of TSMC's R&D dollars go into manufacturing chips while Intel spends a significant portion of its budget on chip design.

TSMC is the great equalizer that AMD needed going against Intel. Intel will have to win purely on chip architectural design from now on. They might have to compete with a lesser node process for a long time.

I don't see Intel ever having a node advantage ever again over AMD. TSMC will continue to have an advantage from now on because of how much money they get from mobile.

It won't surprise me if, in 5 years, Intel uses TSMC to manufacture their top chips.

Page 6 - Question - How ahead is Intel in CPU design compared to AMD?

Page 6 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Vattila]

The world needs two players, if not Intel then who?

The proposition isn't that Intel's manufacturing — a huge organisation in itself — will disappear off the planet. The issue is whether design and manufacture will continue to be integrated dependent entities at Intel, or whether Intel will spin off manufacturing into an independent entity, e.g. selling it or, more likely, creating a joint venture. As independent entities design and manufacture become more flexible and viable. Design can use the best manufacture available. Manufacture can offer services to anyone, including a lot of global demand for economical manufacture on established nodes, not necessarily on the bleeding edge, thereby keeping the fabs utilised (ref. GlobalFoundries' current strategy).

 

[Vattila]

so moving to a different Fab is not as big of a deal as it used to be

Excellent post. When do you figure Intel must have started on a TSMC design, to be able to have something ready to compete with AMD's Genoa on the N5 process, which according to current roadmaps is arriving before 2022?

 

[Doug S]

Excellent post. When do you figure Intel must have started on a TSMC design, to be able to have something ready to compete with AMD's Genoa on the N5 process, which according to current roadmaps is arriving before 2022?

Designs don't start out as "an Intel design" or a "TSMC design", that type of process specific tailoring can happen fairly late in the game (or at least can now that Intel is using industry standard tools instead of internally developed stuff)

Intel could be in the middle of an Intel 7nm design and still have time to change course towards a 5nm TSMC design if the goal was to ship in Dec. 2021. You don't necessarily have to choose, either. Remember what Apple did with the A9, having it fabbed by both TSMC and Samsung. They certainly didn't start out four years earlier when A9 was a blank sheet of paper with that in mind. Intel can have TSMC N5P as a backup plan while still regarding their own 7nm process as plan A.

 

[Vattila]

in 5 years intel could be ahead again like nothing happened

But will the Intel board and shareholders accept paying for manufacturing facilities that are just limping along for 5 years, including the huge risky investment in research and development needed to maybe regain the lead to produce only your own products, which may not even gain the market share needed to fully utilise the factories? If spun off into a separate independent venture, those facilities could do good business satisfying the enormous global demand for manufacturing on mature and economical nodes (ref. GlobalFoundries). Even at TSMC, most wafer starts are at mature process nodes.

 

[DrMrLordX]

@RasCas99

Thanks for the feedback. From what other insiders have said on these boards for years, Intel takes a very different approach to design process than other major designers, but I don't have the industry knowledge to comment saliently on the subject.

@Doug S

When was the last time that Intel produced a product outside of its own fabs? Honest question, I don't know the answer.

My take on the fab situation:

Intel has massive fab capacity, a lot of cash, and plenty of talent (10nm/7nm failures notwithstanding). That has value to Samsung, TSMC, and anyone else that might hope to crank out wafers. It really makes sense for someone else to swoop in and prop up Intel's fab business just as an eventually-cheap way to expand fab capacity.

TSMC could use the extra capacity. If Samsung stays competent with their processes, they could use it too. Intel needs a process that they can use to maintain IDM status, somewhat. It's a match made in heaven for Samsung and/or TSMC, and one made in hell for Intel. They'll have to mortgage their entire fab business just to get a process that isn't broken (which they will modify to suit their own needs, no doubt), and they'll have to lay out more cash to expand capacity so that the licensee (Samsung/TSMC) gets what they want out of the deal. Best-case, Intel maintains IDM status at the cost of being vassal to someone else's fab business, and their entire R&D effort shifts to adapting someone else's processes to their designs.

 

[jpiniero]

When was the last time that Intel produced a product outside of its own fabs? Honest question, I don't know the answer.

They have several right now actually, but I don't know of any on a leading edge node. Modivius and Mobileye use TSMC 16 nm I think.

 

[DrMrLordX]

They have several right now actually, but I don't know of any on a leading edge node. Modivius and Mobileye use TSMC 16 nm I think.

Were those designed in-house, or is that legacy from acquisitions?

 

[senttoschool]

I find it funny that AMD's stock went up by 16.50% today while Intel's went down by the same percentage.

Intel signaling that they will use a 3rd party fab is literally the worst thing that can happen to AMD. Now AMD's node advantage is gone and both their CPU and GPU margins will take a huge hit as Intel slushes its cash around and buys up TSMC/Samsung wafers which will increase the price for everyone.

In addition, Intel still has its own fabs which means they will always have significantly more capacity. They can use TSMC for high-performance chips and their own fabs for mid-range to low tier chips.

Yes, this is really bad for Intel but it's also bad for AMD.

 

[VirtualLarry]

Honestly, Intel 22nm and 28nm probably have a lot of value, at least 22nm, for things that don't need to be so much on the "bleeding edge", like IoT micro-controller SoCs and such. I think Intel spinning off their fabs, and accepting orders from other (potentially, non-competing) companies MIGHT be a good thing, and MIGHT be even "necessary", if things get worse, before they get better for Intel.

 

[chrisjames61]

Unless Intel's revenue hits heavy decline in next couple of years, I don't think that will be the case yet. Don't be so overdramatic.

You are whistling past the graveyard dude.

 

[Doug S]

When was the last time that Intel produced a product outside of its own fabs? Honest question, I don't know the answer.

During Intel's shortages a few years ago it was reported they were outsourcing chipset production to TSMC.

 

[senttoschool]

Can we just take a moment and appreciate the foresight the OP had? This dude nailed it in 2018, way before Zen2 and before more Intel node issues. Reading his original argument makes all the sense at this moment. While I predicted it in Feb 2020, it was fairly obvious by that point.

Everyone who said he was an idiot or was wrong needs to apologize.

The second post said it was "nonsense". A few posters said OP needs to take his meds.

The majority opinion in this forum has been wrong so often. It makes you wonder.

First, most posters here thought Apple would never switch Macs to ARM or that Apple would only use ARM in their entry-level Macs. Apple just came out said they're doing the whole shebang within 2 years.

Second, most posters here went against the OP. Some called him crazy. The poll is completely lopsided.

What's next for this forum to be wrong about? ARM chips having no chance against x86 in high-performance environments such as the server? China won't ever be able to compete in semiconductors?

In hindsight, a lot of these things are quite obvious but it can be hard to see the trend line when you're standing right in front of it.

 

[Vattila]

This dude nailed it in 2018

Thanks for the recognition. But among other long-term AMD investors, already back in 2008, the idea that the fabless model would win out was discussed as inevitable, despite Mark Bohr at Intel predicting the death of the foundries. This was when AMD spun off their fabs into GlobalFoundries, due to the escalating cost of building leading edge fabs. It was clear back then that the pure IDM model only worked if you (a) managed to stay at pole position in process development and (b) your market share (product success) was sufficient to fully utilise the fabs, thus recoup the huge investment. With a more level playing field in x86, which we all hoped for, with leadership not guaranteed at every product generation, as well as increasingly capable foundries, that would become a challenge for anyone, including Intel. It took a while, but it has been playing out for some years now.

The separation of design and manufacture is more robust and flexible, at the cost of some gross margin.

 

[JasonLD]

You are whistling past the graveyard dude.

Eh, maybe. Though all the superior products AMD had for this past year and half, market share shift has been anything but dramatic. I guess we'll just wait and see. Lot of interesting times ahead in next couple of years.

 

[Kenmitch]

It's not just the market share they will lose, it's also that they will be forced to cut prices pretty significantly on the share they do keep.

Zen's double edged sword in action.

 

[Doug S]

Can we just take a moment and appreciate the foresight the OP had? This dude nailed it in 2018, way before Zen2 and before more Intel node issues. Reading his original argument makes all the sense at this moment. While I predicted it in Feb 2020, it was fairly obvious by that point.

I wasn't reading this forum until recently, but what he was arguing was somewhat similar to what some of us argued years earlier when 450mm fabs were discussed on RWT.

I said that more than doubling the number of chips coming off a wafer would be disastrous for Intel's economics - they had (and still have) about four large fabs making x86 CPUs. That would put them down to two fabs - but even that assumed that the economics of 450mm wafers wouldn't be such that you'd need a larger fab for it to make sense (as happened in the 200mm to 300mm transition) at which point they'd be down to needing only a single fab's capacity to supply all x86 CPUs. They'd need a second fab for geographic redundancy and to allow them to upgrade one to a new process which would mean an average fab utilization of only 50%.

While the PC market has been shrinking, it is only shrinking on the client side. Servers - which are much larger chips - are still growing, and they've developed a presence in HPC and own the majority of cloud. As he correctly argued, they were/are pretty vulnerable to AMD taking share from them - the economics of operating a fab don't work nearly as well when they are at 50% utilization versus 100%. They were actually over 100% (needing to farm some chipsets off to TSMC) when they were making modems for Apple a couple years ago. The same fate that would have awaited them with 450mm wafers awaits them if AMD is able to take significant server share from Intel. People buying servers move slowly, it won't happen quickly but it will happen eventually if Intel continues to step on its own dick.

Posters here have been annoyed that AMD isn't moving faster on laptop CPUs and think that's how they stick it to Intel but that's not how they hurt Intel. They hurt Intel by taking server CPUs away - those have the fattest margins, and leave them with the largest amount of unused capacity, which increases the share of fixed fab investment cost attached to each chip they sell. That's where TSMC has a massive advantage - they have customers for all their old processes so they continue depreciating them forever, they are still running .35u wafers that were leading edge in the mid 90s!

 

[naukkis]

I find it funny that AMD's stock went up by 16.50% today while Intel's went down by the same percentage.

Intel signaling that they will use a 3rd party fab is literally the worst thing that can happen to AMD. Now AMD's node advantage is gone and both their CPU and GPU margins will take a huge hit as Intel slushes its cash around and buys up TSMC/Samsung wafers which will increase the price for everyone.

Intel does need competitive products with equal manufacturing possibilities - which they don't have. They have been relied to their superior manufacturing capability so their products are bloated and way inferior. Their cpu archs are similar like Samsung - too bulky to be compete with smaller cores that perform just as well. Samsung did axe their cpu line, Intel has to do something too.

 

[RasCas99]

Designs don't start out as "an Intel design" or a "TSMC design", that type of process specific tailoring can happen fairly late in the game (or at least can now that Intel is using industry standard tools instead of internally developed stuff)

Intel could be in the middle of an Intel 7nm design and still have time to change course towards a 5nm TSMC design if the goal was to ship in Dec. 2021. You don't necessarily have to choose, either. Remember what Apple did with the A9, having it fabbed by both TSMC and Samsung. They certainly didn't start out four years earlier when A9 was a blank sheet of paper with that in mind. Intel can have TSMC N5P as a backup plan while still regarding their own 7nm process as plan A.

Excellent post. When do you figure Intel must have started on a TSMC design, to be able to have something ready to compete with AMD's Genoa on the N5 process, which according to current roadmaps is arriving before 2022?

Agree with Doug here on most of the things including a late decision that can be made , but it cant be terribly late either , a chip design process is structured the same way for basically everyone in the industry , you have a Uarch definition of the CPU, which usually has some inputs from Marketing/Product teams (depending on the company) , so lets say Marketing/product tells you , listen you need to have PCIE4 support and DDR5 support in the next CPU , then comes inputs from the Uarch team itself which have their own ideas for changes/improvements, this one can be really big changes (ala core2duo) , or minor changes (KBL) , you also get a budget for Die size and some target power / performance and a timeline for TO , and many more inputs that i am not aware of (note die size , power , performance are ALL relied on process characteristics , so a change in process will create different CPU on all of those accounts , but it wont impact the fact you will have memory controller that supports DDR5 designed as its process agnostic).
Then the design teams provide feedback on feasibility and after some iterations with project management there is an agreement and off to the races we go.

The bulk of the chip is designed in RTL (system-verilog) which is for the most part process agnostic , until this point we can define and design our chip without process dependency (from a logic POV that is ,but we will have different CPU if we switch in regards to area/power/performance).

But once we start integrating memories (SRAM for example) , start using standard/custom libs , and mainly starting with synth to get some timing/size/power numbers then it becomes very important to have the right process kit available , also as i mentioned the Analog portion needs to know whats the process characterization to make sure their design is working.

We are now in execution - verification , bug fixes and synth convergence until TO.

I cant say exactly when in the game Intel can switch over , but if they knew in advance that they might need to switch , i would say its basically the same timeline as they will do it in parallel (as they will plan accordingly) , if the request to do it lands today , I would say its about 1 year delay until they get it up to speed without hurting current on going projects, but thats a WILD guess on my side with a lot of grain and salt , I am trying to account for the 2x work for the synth folks , different memories integration (i.e SRAM), changing all the custom designs ,layout impact , different Analog all the while not being able to get a influx of engineers (recruiting and integrating engineers takes time) and supporting both data bases of design , if they abandon current Intel process and focus everyone on lets say TSMC , the penalty will be lower , again pure speculation here , I am doing design of logic parts (RTL writing) and my interaction with the process/backend is only when i get feedback that I have timing issues or area violations in my design block so this (changing to TSMC) might have bigger impact then i can guesstimate on the backend teams.

Where I disagree with Doug is that i believe (believe != know) , that if they want to ship in 2021 (i.e in clients hands not OEM`s) on TSMC process it means that the change has already been done , as a TO should be imminent or already happened , i guess in some crazy expensive world they can TO Intel first and flush out the bugs so a TSMC silicon will be clean from bugs POV and might be mass produced on A0 stepping , but post silicon work will be needed even if its bug free and a dual cpu validation sounds crazy complicated from a post silicon POV.

 

[Doug S]

I cant say exactly when in the game Intel can switch over , but if they knew in advance that they might need to switch , i would say its basically the same timeline as they will do it in parallel (as they will plan accordingly) , if the request to do it lands today , I would say its about 1 year delay until they get it up to speed without hurting current on going projects, but thats a WILD guess on my side with a lot of grain and salt , I am trying to account for the 2x work for the synth folks , different memories integration (i.e SRAM), changing all the custom designs ,layout impact , different Analog all the while not being able to get a influx of engineers (recruiting and integrating engineers takes time) and supporting both data bases of design , if they abandon current Intel process and focus everyone on lets say TSMC , the penalty will be lower , again pure speculation here , I am doing design of logic parts (RTL writing) and my interaction with the process/backend is only when i get feedback that I have timing issues or area violations in my design block so this (changing to TSMC) might have bigger impact then i can guesstimate on the backend teams.

Where I disagree with Doug is that i believe (believe != know) , that if they want to ship in 2021 (i.e in clients hands not OEM`s) on TSMC process it means that the change has already been done , as a TO should be imminent or already happened , i guess in some crazy expensive world they can TO Intel first and flush out the bugs so a TSMC silicon will be clean from bugs POV and might be mass produced on A0 stepping , but post silicon work will be needed even if its bug free and a dual cpu validation sounds crazy complicated from a post silicon POV.

Apple provides proof that this can be done, and Intel has much greater resources in terms of design teams and engineers. What we don't know is whether Apple knew well in advance they were going to dual source from TSMC and Samsung so they could pursue parallel tracks, or the second source was added later in the game. The rumor mill at the time kind of suggested it was the latter.

That said I will defer to you as the expert since you do this for a living, while I'm going on what I've read or been told by people like you do who actually do it.

 

[Hitman928]

Apple provides proof that this can be done, and Intel has much greater resources in terms of design teams and engineers. What we don't know is whether Apple knew well in advance they were going to dual source from TSMC and Samsung so they could pursue parallel tracks, or the second source was added later in the game. The rumor mill at the time kind of suggested it was the latter.

That said I will defer to you as the expert since you do this for a living, while I'm going on what I've read or been told by people like you do who actually do it.

The biggest challenge I would see to switching from Intel 7 nm to TSMC 7 nm late in the game is that they are not very comparable nodes. When Apple/Nvidia/AMD were sourcing from Samsung and TSMC at the same time, the nodes, while obviously not identical, were very similar at least in terms of density, fmax, power, etc. What Intel would be doing would be more like moving from TSMC 5 nm to TSMC 7 nm. Sure it's doable but there's a lot of implications in terms of chip architecture that will be disrupted because you were targeting a more dense, higher efficiency process to begin with. This is a much bigger deal in terms of what the final product will look like versus jumping to a comparable node and having to re-do your physical design to meet timing and maybe take a 10% frequency hit or something like that.

 

[Gideon]

The biggest challenge I would see to switching from Intel 7 nm to TSMC 7 nm late in the game is that they are not very comparable nodes. When Apple/Nvidia/AMD were sourcing from Samsung and TSMC at the same time, the nodes, while obviously not identical, were very similar at least in terms of density, fmax, power, etc. What Intel would be doing would be more like moving from TSMC 5 nm to TSMC 7 nm. Sure it's doable but there's a lot of implications in terms of chip architecture that will be disrupted because you were targeting a more dense, higher efficiency process to begin with. This is a much bigger deal in terms of what the final product will look like versus jumping to a comparable node and having to re-do your physical design to meet timing and maybe take a 10% frequency hit or something like that.

What's stopping them from moving to TSMC 5nm? The product has insignificant volume so competition for wafers shouldn't be a concern

 

[RasCas99]

Apple provides proof that this can be done, and Intel has much greater resources in terms of design teams and engineers. What we don't know is whether Apple knew well in advance they were going to dual source from TSMC and Samsung so they could pursue parallel tracks, or the second source was added later in the game. The rumor mill at the time kind of suggested it was the latter.

That said I will defer to you as the expert since you do this for a living, while I'm going on what I've read or been told by people like you do who actually do it.

Reading your posts in this forum you are pretty much on point and i do like reading them , as i put you in the "read" group of posters in this forum ! (already identify some of the posters here hhhhhh).
I have no idea about the apple situation to be honest , and it is of course possible to do , I think i read that Nvidia are doing the same now with going both TSMC and Samsung and picking Samsung at the end ? not sure i got that part right in the rumor mill.

If the discussion was about chip microarch and design (in the logic side of things ) I wouldve had more confidence in my statement , as it is more backend/manufacturing , there is a possibility that your guesstimate is more accurate then mine , but to my basic knowledge of that part of the CPU design world i feel like my guesstimates are probably in the ballpark give or take , and there is little chance (albeit possible) they can decide on going TSMC and then have a TO in few months time , once silicon is out on TSMC ill ask my friends and let you know the general timeline as they wont be sharing it exactly no matter how much i poke them .
If you remember this discussion in due time , remind me about it, so we can score our guesstimates!

 

[Hitman928]

What's stopping them from moving to TSMC 5nm? The product has insignificant volume so competition for wafers shouldn't be a concern

It just depends on timing. Trying to secure 5 nm volume to any significant degree is probably pretty tough right now with Apple usually taking a ton of the wafers on the bleeding edge node. Even if this is the plan, I don't know how comparable the current 5nm optimized for mobile is to what Intel's 7 nm is supposed to be. Target frequencies, power, density will all effect the end product and if these things are too far removed from what you thought they were going to be when designing the CPU, your end product is going to be underwhelming.

 

[Vattila]

On the likelihood that Intel could catch up to AMD's Genoa chip, scheduled to arrive before the end of next year, by porting their own server chip design to TSMC's N5, I am doubtful.

I suspect that if Intel had the foresight to go with TSMC in parallel, Jim Keller would still be at Intel. According to rumour, he promoted use of TSMC but was voted down. While at Intel, it seems his main task was to lay out a product and process roadmap that the engineers could get excited about (see his presentations "Moore's Law is not dead", as well as his excellent interview with Lex Fridman, on YouTube). With the realisation of the 7nm delay, the morale must have really deflated, undermining a lot of his efforts.

Meanwhile, Intel CEO Bob Swan explicitly emphasises the 2023 time-frame for his "contingency plans", while talking up his belief in their "strong" product roadmap from here through 2022. It seems the next couple of years are going to be rough for Intel, and then we'll see whether they have a competitive process in 2023 or go into full contingency mode. If the latter, the cost of their fabs will come into focus, if not already, I would think.