Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/chinas-smic-shipping-7nm-chips-reportedly-copied-tsmcs-tech-fake-news.16421/
        )

    [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
)

China's SMIC Shipping 7nm Chips, Reportedly Copied TSMC's Tech (fake news?)

soAsian

Active member
Last edited:
Taiwan News also use TechInsights as source. I'm not sure if this is fake news or not since all reporting use the same source.
However, the past lawsuit between two parties shown there is a possibility.
what do you guys think?

It would be more interesting if Samsung 7nm and Intel 7 were included in this reverse engineer. TechInsights is a reputable firm, absolutely. The other sites who have jumped on this not so much.
 
what do you guys think?
From a CDs perspective, it is very similar based on what is available today (this is a Cryptomining ASIC, so there is no SRAM, there is only one library, there is very limited IO), but if you go back and look at what GF planned for their 7nm, it was also designed on paper to be a second-source to TSMC. SMIC 14nm is very clearly meant to be a "T-like" equivalent to N16P, and this node is "T-like" equivalent to N7 (not 7+, not N6). What is interesting (for process development) is how they achieve the metal double patterning, and whether they achieve anything close to the equivalent performance (fin drive, wire RCs) besides just being able to pattern something at equivalent CDs.

The parts that are produced may never get shipped outside of China based on these similarities though, unless there is some license agreement with TSMC.
 
Out of curiosity, how does one reverse engineer a fabrication process? Stealing IP and hiring people from the source company?
 
Out of curiosity, how does one reverse engineer a fabrication process? Stealing IP and hiring people from the source company?
For sure SMIC hired people with Foundry experience elsewhere, but I think there are limits on what you can mean by reverse engineer. You can take a shipping product apart, do SEM/TEM/EELS analysis on it, and get measurements - this is what Tech Insights and others do for a living. You can probe individual transistors and generate the I-V sweep of them to get performance, and you can use EELs on the gate stack to get some views of the WFMs used for each VT, but that assumes you can find them all on the die. TI does this work as well.

If you are SMIC or someone else, you can just buy TI reports and see what other Foundries are doing with shipping in market, and use that to influence your designs. But that data has limitations, and knowing "what" to do doesn't mean knowing "how" to do it. For example, you can't today get EELS sufficient to tell you Boron concentrations within an Epi, but you can measure strain and then work your own process of implants until you get something similar. If your gate stack is using a 'zero thickness' WFM solution, then again hard to identify exact concentrations, but you can read up on state-of-the-art literature published in IEEE journals and again, use that to inform your own implementation.

For multi-patterned metals (below the limit of 193i LE) you have options: multi-patterning (LELE/LE^n), pitch division (SAnP), or some hybrid (SA-LELE). Samsung / Intel / TSMC all employ different versions of these in their shipping processes, and imec publishes a ton of work on how to do it, but the Foundry still has to make a choice of what to do themselves. That was to my earlier comment that it is clear they support pitches in the ~40-44nm regime, but less clear what process they used to make that, though all options have 'signatures'. FEOL basically everyone is doing SAQP for fins and SADP for gates, but that still leaves the door open to the actual CDs, cell termination style, how your STI is formed / etched, support for fin depopulation, etc.

All of the above is "proper" RE work, and should not be considered IP theft per se, however if you are implementing a process that has been patented by another Foundry (and absolutely everything they use has been patented), then you should be receiving a license to employ the patented process. All the leading Foundries have a MAD stack of patents, and so they generally play nice but newcomers cannot avoid violation, and don't typically have their own patents to assert instead or cross-license.

UMC has a 14nm node which is their first FinFET option - they had limited initial success with this but currently is contributes $0 to their revenue, partially because their bring-up was very late but I would wager mostly because it isn't "T-like" or "S-like", which means customers who have already designed their products elsewhere are going to have to re-design to use UMC process. GF licensed the 14mn process from Samsung and certain designs can be multi-sourced between those two; their planned 7LP node (as published at VLSI) was similar to N7 but had a few differences (like CPP) and ultimately they never commercialized that. SMIC has opted to avoid this business pitfall by making it easier for Fabless customers to multi-source TSMC, but it is questionable on the legality of it (and IANAL, so I won't comment). Without access to EUV tools, they cannot produce anything equivalent to S5/4 or N5/3 so this may be the end of line for their process development anyhow.
 
For sure SMIC hired people with Foundry experience elsewhere, but I think there are limits on what you can mean by reverse engineer. You can take a shipping product apart, do SEM/TEM/EELS analysis on it, and get measurements - this is what Tech Insights and others do for a living. You can probe individual transistors and generate the I-V sweep of them to get performance, and you can use EELs on the gate stack to get some views of the WFMs used for each VT, but that assumes you can find them all on the die. TI does this work as well.

If you are SMIC or someone else, you can just buy TI reports and see what other Foundries are doing with shipping in market, and use that to influence your designs. But that data has limitations, and knowing "what" to do doesn't mean knowing "how" to do it. For example, you can't today get EELS sufficient to tell you Boron concentrations within an Epi, but you can measure strain and then work your own process of implants until you get something similar. If your gate stack is using a 'zero thickness' WFM solution, then again hard to identify exact concentrations, but you can read up on state-of-the-art literature published in IEEE journals and again, use that to inform your own implementation.

For multi-patterned metals (below the limit of 193i LE) you have options: multi-patterning (LELE/LE^n), pitch division (SAnP), or some hybrid (SA-LELE). Samsung / Intel / TSMC all employ different versions of these in their shipping processes, and imec publishes a ton of work on how to do it, but the Foundry still has to make a choice of what to do themselves. That was to my earlier comment that it is clear they support pitches in the ~40-44nm regime, but less clear what process they used to make that, though all options have 'signatures'. FEOL basically everyone is doing SAQP for fins and SADP for gates, but that still leaves the door open to the actual CDs, cell termination style, how your STI is formed / etched, support for fin depopulation, etc.

All of the above is "proper" RE work, and should not be considered IP theft per se, however if you are implementing a process that has been patented by another Foundry (and absolutely everything they use has been patented), then you should be receiving a license to employ the patented process. All the leading Foundries have a MAD stack of patents, and so they generally play nice but newcomers cannot avoid violation, and don't typically have their own patents to assert instead or cross-license.

UMC has a 14nm node which is their first FinFET option - they had limited initial success with this but currently is contributes $0 to their revenue, partially because their bring-up was very late but I would wager mostly because it isn't "T-like" or "S-like", which means customers who have already designed their products elsewhere are going to have to re-design to use UMC process. GF licensed the 14mn process from Samsung and certain designs can be multi-sourced between those two; their planned 7LP node (as published at VLSI) was similar to N7 but had a few differences (like CPP) and ultimately they never commercialized that. SMIC has opted to avoid this business pitfall by making it easier for Fabless customers to multi-source TSMC, but it is questionable on the legality of it (and IANAL, so I won't comment). Without access to EUV tools, they cannot produce anything equivalent to S5/4 or N5/3 so this may be the end of line for their process development anyhow.
So, if I can take the first derivative of what I think you said, SMIC didn't really reverse engineer anything significant. They probably hired some people from other foundries with interesting similar process experience, essentially invented their own 7nm-equivalent process, and in doing that almost certainly infringed on a bunch of patents from various other companies. And if they're just selling to Chinese private and government customers, well, they're in so much hot water with IP-related problems that a little more is not significant. Am I approximately correct? Perhaps this (secret?) 7nm process could also help explain how China reportedly produced multiple exascale-class supercomputers which they are keeping under wraps rather than bragging about.
 
Key message is "SMIC broke the barrier of not using EUV to manufacture yielded 7nm node". DUV 193i is used since 28nm. SMIC and other China foundries have lots of immersion tools which could be capable for 7nm and possible 5nm manufacturing. Will it become a reason to restrict 193i to China?
 
So, if I can take the first derivative of what I think you said, SMIC didn't really reverse engineer anything significant. They probably hired some people from other foundries with interesting similar process experience, essentially invented their own 7nm-equivalent process, and in doing that almost certainly infringed on a bunch of patents from various other companies.
I think this gets to the heart of what you consider IP theft. If "Great Wall motors" is manufacturing a SUV that very much looks like a BMW or Toyota, but has a different grille and badge, is it stolen, copied, or a 'grounds up design' that is heavily inspired by something that already sells well in the marketplace? Common language usage, people would say it is 'copied' because the design aesthetics are similar, and I'd posit the headlines for SMIC "copying" TSMC here are used in a similar vein.

I don't think anyone has claimed SMIC conducted industrial espionage against TSMC to determine how to make this process technology (yet), but it is hard to think you can make something "T-like" without infringing on TSMC patents. My point though was there are several ways you can get to a similar looking final result while employing different process modules to get there, so it is not (yet) clear if what they are doing is "very similar" to TSMC, or "verbatim" to TSMC, but they developed this process with a clear goal of making it "close enough" that designs could be second-sourced from TSMC. This could be to mitigate supply disruptions or for larger supply volume, or for cost pressure - all typical reasons a customer might pursue a second source option. Ultimately, SMIC is a foundry making a technology for fabless customers to use.

And lest it be unclear, I'm trying to be balanced between what is clear they have done / are doing, and what is speculative about why, their customers, or whether what they have done is legal, because again, IANAL. If TSMC (or Samsung, or GF) thinks this process is violating their patents, that will come up shortly, I'm sure.
 
Key message is "SMIC broke the barrier of not using EUV to manufacture yielded 7nm node". DUV 193i is used since 28nm. SMIC and other China foundries have lots of immersion tools which could be capable for 7nm and possible 5nm manufacturing. Will it become a reason to restrict 193i to China?
Well let's be abundantly clear, N7 does not make use of EUV tools, and has been in mass production since H2'2018. It is not a requirement to use EUV to make a "7nm" node, but I struggle to see how they could develop anything contemporary to N5 or S5/4 without EUV - the minimum metal pitches are very fine and would require LE^4 or SAQP.

But the other view of this is... they are commercializing a contemporary technology only 4 years after TSMC. GF hasn't done that. UMC hasn't done that. Depending on your view of IFS this makes them either 3rd or 4th most advanced logic foundry on the planet.
 
You should also know that announcing a new process technology and having revenue associated with high volume manufacturing is two very different things for SMIC. They announced 14nm in 2015 and didn't record HVM revenue for many years later. On the SMIC website they now claim 14nm HVM in Q4 2019. So let's see what percentage of SMIC's revenue is 14nm and 7nm on the next report. Bottom line: I will believe it when I see it on their financials.
 
You should also know that announcing a new process technology and having revenue associated with high volume manufacturing is two very different things for SMIC. They announced 14nm in 2015 and didn't record HVM revenue for many years later. On the SMIC website they now claim 14nm HVM in Q4 2019. So let's see what percentage of SMIC's revenue is 14nm and 7nm on the next report. Bottom line: I will believe it when I see it on their financials.

SMIC put 28nm/FinFet in one category on financial report, so you cannot tell the revenue from 28nm/14nm/7nm apart
 
I think this gets to the heart of what you consider IP theft. If "Great Wall motors" is manufacturing a SUV that very much looks like a BMW or Toyota, but has a different grille and badge, is it stolen, copied, or a 'grounds up design' that is heavily inspired by something that already sells well in the marketplace? Common language usage, people would say it is 'copied' because the design aesthetics are similar, and I'd posit the headlines for SMIC "copying" TSMC here are used in a similar vein.

I don't think anyone has claimed SMIC conducted industrial espionage against TSMC to determine how to make this process technology (yet), but it is hard to think you can make something "T-like" without infringing on TSMC patents. My point though was there are several ways you can get to a similar looking final result while employing different process modules to get there, so it is not (yet) clear if what they are doing is "very similar" to TSMC, or "verbatim" to TSMC, but they developed this process with a clear goal of making it "close enough" that designs could be second-sourced from TSMC. This could be to mitigate supply disruptions or for larger supply volume, or for cost pressure - all typical reasons a customer might pursue a second source option. Ultimately, SMIC is a foundry making a technology for fabless customers to use.

And lest it be unclear, I'm trying to be balanced between what is clear they have done / are doing, and what is speculative about why, their customers, or whether what they have done is legal, because again, IANAL. If TSMC (or Samsung, or GF) thinks this process is violating their patents, that will come up shortly, I'm sure.
I'm hard-pressed to see how SMIC stole anything specific, and hiring some experts from TSMC isn't compelling as proof of anything. Doing a group hire of an entire team might be, but I can't believe that ever occurred. It would be too obvious.

As many of us know, determination of patent infringement can be very difficult. Enforcing a patent for an instruction set is one thing, something highly technical and obscure like a step in a fab process... it is difficult to imagine. This is normally settled with cross-licensing agreements and settlements, which is obviously why most big companies want to hold as many patents as possible. More leverage at license negotiations.

But China and SMIC is a different case, isn't it? Americans paid a lot of money and took risks to drink alcoholic beverages during Prohibition. I'm guessing the cost and risk tolerance level of Chinese commercial and government customers will be high too.
 
SMIC put 28nm/FinFet in one category on financial report, so you cannot tell the revenue from 28nm/14nm/7nm apart

Why would SMIC not separate 28nm and 14nm? They separate all of the other nodes? Clearly SMIC is hiding something.

SMIC 2018 2019 Wafer Sales.jpg
 
There is power struggle inside of communist party right now... Crazy "health protection" restrictions, huge money/debt crisis, protests - even tanks on streets...

In my opinion this announcement is just to show, that they are making some progress, so they will not be victims of some new regulations. This is because there are so much questionable companies inside china collecting lot of money while not having any expertise in area... HSMC, Tsinghua Unigroup and countess other mainly memory and foundry oriented...

SMIC is in dangerous area because there are questions on how can company with stagnating revenue around $3B (now around $5B thanks to sanctions), how can they afford developing and building leading edge nodes...
 
You should also know that announcing a new process technology and having revenue associated with high volume manufacturing is two very different things for SMIC. They announced 14nm in 2915 and didn't record HVM revenue for many years later. On the SMIC website they now claim 14nm HVM in Q4 2019. So let's see what percentage of SMIC's revenue is 14nm and 7nm on the next report. Bottom line: I will believe it when I see it on their financials.
Dan - that would require you to believe the financials reported by a Chinese company ... I'm not sure I would take too much such data at face value without very careful scrutiny. UK-audited financials aren't always 100% reliable (auditors KPMG fined over some cases only a few days ago).
 
However, the past lawsuit between both parties shown there is a possibility.

What's the big deal of historical lawsuits?Tech firms sue each other all the time,can you name any tech firms that has never been sued?

Dr Liang,who was TSMC's director responsible for Finfet development went to Samsung,and helpled Samsung to develop 14nm process. TSMC then sued Samsung for IP infringement, so I guess Samsung's Finfet technology also "inspired" from TSMC
 
If you are SMIC or someone else, you can just buy TI reports and see what other Foundries are doing with shipping in market, and use that to influence your designs. But that data has limitations, and knowing "what" to do doesn't mean knowing "how" to do it.

Exactly,node process is very much like materials,the know-how is in a "blackbox". Meaning you cannot reverse engineer it by just analyze final products.

You can analyze the chemical composition of a material,but you wouldn't know how it is being made. The right temperature,the ratio of different chemicals,the right parameters of machines etc. A tiny change could make huge difference in final products,that's why you just cannot make 7nm process by simply look at X-ray photos of TSMC products.
 
Dan - that would require you to believe the financials reported by a Chinese company ... I'm not sure I would take too much such data at face value without very careful scrutiny. UK-audited financials aren't always 100% reliable (auditors KPMG fined over some cases only a few days ago).

SMIC sales by region 2022.jpg

Why is north America so big for SMIC? Who is still using SMIC?
 
Back
Top