First EUV light marks key milestone for production of Intel 4 in Europe
- Thread starter Daniel Nenni
- Start date
I don't know which layers moved to EUV. If I had to guess Fin patterning and M0-2, but I could be (and probably am) wrong. M0 are just the contacts. I may be misunderstanding what you mean by routable, but no they aren't routable. They are basically just vias that connect M1 to the devices.
Scotten Jones
Moderator
"when one considers that N7+ was just one layer" TSMC 7+ is 5 EUV layers
Scotten Jones
Moderator
7 was Double Diffusion Break, 7+ added a EUV Single Diffusion Break mask and ViaC, Via0, M1 and M2 all went to EUV = 5 layers.
Last edited:
I don't think one process makes you a viable contender in the foundry business. It is a marathon not a sprint. No matter what you think IDM foundries will never have the level of collaboration that a pure-play foundry does. TSMC will never compete with customers but you never know with IDMs. One day they may not, another day they may, you cannot bet your whole business on that type of partnership. The whole concept of frenemies and co-opetition is a fools errand. TSMC will never trust Intel nor will Apple ever trust Samsung like Apple, AMD, and others trust TSMC.
So, in my eyes there are pure-play foundries (TSMC, UMC, SMIC, GF, etc....) and IDM foundries (Intel, Samsung), same zoo, different animals.
Profitable yes, but not with IDM or TSMC type of margins, more like UMC and GF.
Process isn't everything for foundries. Ecosystem is far more important and the ability to deliver what you promise. Customer silicon is they key benchmark here not the PPA nonsense that the media rehashes. Remember, this is not the first time Intel has been in the foundry business and the pervious times they had a process advantage.
Apple had no choice with 14nm, TSMC did not deliver the wafer capacity Apple required. It was a good lesson for all and definitely changed the face of the foundry business.
Nobody likes monopolies which is why there will always be a NOT TSMC market. Just like AMD used to be the NOT Intel supplier. How long was AMD on life support? Now AMD has added BETTER THAN Intel market share. Lisa Su has been with AMD for 10+ years? IFS and SFS have the same opportunity but it will take time. Does Pat have 10 years in him? Samsung Foundry changes executives like socks so the clock keeps getting reset.
Just my opinion of course...
I assume you meant AMD has a higher market capitalization than Intel, which it did for a while. As of this posting that's not the case, but since AMD's is $104B and Intel's $107B, Intel's is still pathetic compared to their revenue advantage.
In market share the two companies aren't even close. In x86 CPUs it's about 70/30 in favor of Intel.
Thanks for answering that long string of questions Dan. When it comes to the wider fabless ecosystem you are the man. With all of the articles about people's add-ins for TSMC design flows I was curious if all of the work SFS and IFS were putting in to catch up to TSMC was for naught (at least for the foundry side). Obviously the pure-play model is a nice selling point. Where I was unsure is if getting a strong enough IP/design ecosystem was even possible, or if we would see a similar issue to Microsoft trying to cut into the mobile phone OS market (nobody bothering to support it resulting in nobody using it). Technical prowess, cadence, and PDKs are purely within Samsung and Intel's power to improve. Ecosystem less so. Based on your answers it seems that the IDM foundries can be reasonably successful if they can catch up on the things that are in their control. In the process of improving those items the ecosystem will slowly build around them, and with that they can overtime grow into being healthy competitors (although obviously never bigger than TSMC).
When you say customer silicon is the most important benchmark, could you elaborate? Apple is very clearly the best at squeezing every drop out of every mm of TSMC silicon. Does that mean that if a 28nm A series SOC beats a 22nm atom SOC that TSMC has the better process? Or do you more so mean that theoretical PPA is meaningless compared to what is easily achievable with the process and PDKs (ie it doesn't matter if in theory intel 14nm smokes N16, only intel is capable of designing a chip with those dumpster fire PDKs). I don't know if I'd agree with the first option since they aren't really like for like comparisons, but the second option makes complete sense to me.
When you say customer silicon is the most important benchmark, could you elaborate? Apple is very clearly the best at squeezing every drop out of every mm of TSMC silicon. Does that mean that if a 28nm A series SOC beats a 22nm atom SOC that TSMC has the better process? Or do you more so mean that theoretical PPA is meaningless compared to what is easily achievable with the process and PDKs (ie it doesn't matter if in theory intel 14nm smokes N16, only intel is capable of designing a chip with those dumpster fire PDKs). I don't know if I'd agree with the first option since they aren't really like for like comparisons, but the second option makes complete sense to me.
Last edited:
No, better silicon. What was AMD’s market share 10 years ago? Single digits? That was the NOT Intel market share. Now AMD has “our silicon is better” market share.
Now I see what you’re saying.
Big customers do a lot of chips so they know how the process reacts to their designs thus the PPA marketing claims are largely ignored. Process is important but design is key and that is why PDKs are critical. TSMC has the best PDKs because of experience. 30+ years making thousands of PDKs for thousands of customers. When TSMC says the trusted foundry the PDKs are a big part of that trust.
New chip design companies are in a different position of course. They do not have a silicon track record but their newly hired chip employees do and you can bet that experience is mostly with TSMC. New chip design companies (fabless systems companies for example) also depend more heavily on the ecosystem and that brings them to TSMC. So PPA numbers will not be the first thought of these new customers. The prevailing thought will be "first time working silicon so I can keep my job".
Interesting discussion, lots of threads: Wither Samsung, pure-play vs. IDM foundry, first spin trust, whether Samsung even has an ecosystem.
With Samsung, there are plans to build a lot of fabs in the US for Foundry, on spec (ie no firm customers). So that’s a risk, but part of the plan to truly go toe-to-toe with TSMC, matching leading-edge capa for capa. No one has ever challenged TSMC like this before, and because of that, I think TSMC clearly has the advantage, but it’s an old advantage.
It’s been pointed out that the dot com bubble produced a glut of fiber capa that lasted for a decade after Worldcom disappeared amid scandal. The company disappeared but the fiber was still lit, and helped launch Facebook and Google and so on with cheap internet service. Samsung (or Intel) could end up like Worldcom, over-building capa that kills them, but spreads cheap silicon capa to the masses.
My guess is if Samsung’s challenging TSMC fails, those fabs being built will have plan B (memory) or plan C (new applications). And some executives will be fired. But it won’t end Samsung. Intel, on the other hand, has bigger challenges, and their new fabs pretty much have to succeed to live. That could be some good motivation or a hopeless hail Mary.
With Samsung, there are plans to build a lot of fabs in the US for Foundry, on spec (ie no firm customers). So that’s a risk, but part of the plan to truly go toe-to-toe with TSMC, matching leading-edge capa for capa. No one has ever challenged TSMC like this before, and because of that, I think TSMC clearly has the advantage, but it’s an old advantage.
It’s been pointed out that the dot com bubble produced a glut of fiber capa that lasted for a decade after Worldcom disappeared amid scandal. The company disappeared but the fiber was still lit, and helped launch Facebook and Google and so on with cheap internet service. Samsung (or Intel) could end up like Worldcom, over-building capa that kills them, but spreads cheap silicon capa to the masses.
My guess is if Samsung’s challenging TSMC fails, those fabs being built will have plan B (memory) or plan C (new applications). And some executives will be fired. But it won’t end Samsung. Intel, on the other hand, has bigger challenges, and their new fabs pretty much have to succeed to live. That could be some good motivation or a hopeless hail Mary.
Xilinx discussed (here: https://www.spiedigitallibrary.org/...ss-for-7nm-FPGA/10.1117/12.2584288.full?SSO=1) how they switched a single layer (Vc) on some lots. I got the impression that the customer would be free to select which layers to switch (up to some number).
Last edited: