Chinese semiconductor thread II

[tphuang]

I have a question. Does the report explicitly say the dies were fabbed last year and how did they assert that. For example does it use process nodes or features that were not available at TSMC at the time of cut off. I am saying this because it is very common to use dies that were fabbed in earlier years in chips launched much later. As we have seen with several Kirin chips Huawei launched after getting cut off they used stockpiled dies that were fabbed with TSMC pre-sanction.

Even Nvidia does that, at least for the consumer GPUs. They just bin the dies and use the lowest binned dies in the first year and save the better binned dies from the same batch to use in subsequent years as new products. For example Nvidia 3080 ( launched in 2020), 3080ti (2020), 3090 (2021), 3090ti (2022) are all from the same wafer batch of the same die.

I think there is a good chance that the die in the report were fabb'd a while back.

But I will say again, TSMC fabb'd huge quantity of Ascend dies this year. Remember seeing those report of Chinese AI chip giving last minute rush orders to TSMC in Q2/Q3? Most of that was probably Ascend.

There is a reason TSMC Q3 numbers looked so good. It had a lot of 7nm capacity that probably wasn't doing anything other than fabbing Chinese AI chips.

 

[tphuang]

Interesting to see that China has localized most of their FPD manufacturing equipment. So export controls have no effect on them now.

View attachment 141212

I think the same will happen in Wafer fabrication Equipment, China tool companies are going to eat from the market share of US and most foreign tool companies restricting them to the cutting edge or nothing at all. And That will happen, is probably almost inevitable at this moment with these sanctions.

This is a really good chart for the future. IIRC, NAURA is a major supplier for this?
Who are the other major players?

 

[huemens]

I think there is a good chance that the die in the report were fabb'd a while back.

But I will say again, TSMC fabb'd huge quantity of Ascend dies this year. Remember seeing those report of Chinese AI chip giving last minute rush orders to TSMC in Q2/Q3? Most of that was probably Ascend.

There is a reason TSMC Q3 numbers looked so good. It had a lot of 7nm capacity that probably wasn't doing anything other than fabbing Chinese AI chips.

May be. But the problem I have with that is if they were indeed still fabbing with TSMC they shouldn't have had the yield and volume issues that are known about it. For example it was reported earlier that Bytedance ordered 100k 910B this year but they were only able to deliver 30k. There were also earlier reports of a 910B card with less active cores than they used to have when using TSMC dies, which indicates low yield. So if the report doesn't specifically say that they were made by TSMC this year (only just that they were fabbed by TSMC) then it makes more sense to me that these are just stockpiled dies.
And as for the TSMC volume usage there's plenty of other Chinese companies fabbing with TSMC. All I am saying is if Huawei was still activly fabbing with them OR those other companies where doing it on behalf of Huawei where's all those chips, why Huawei is not able to do high volume.

 

[sunnymaxi]

Interesting to see that China has localized most of their FPD manufacturing equipment. So export controls have no effect on them now.

View attachment 141212

I think the same will happen in Wafer fabrication Equipment, China tool companies are going to eat from the market share of US and most foreign tool companies restricting them to the cutting edge or nothing at all. And That will happen, is probably almost inevitable at this moment with these sanctions.

Display Equipment supply chain entirely was in Japan they had full monopoly. Koreans have also using this supply chain but sanctions on chips forced China to develop indigenous supply chain and look where we are today. BOE now mostly using domestic equipment.

This is a really good chart for the future. IIRC, NAURA is a major supplier for this?
Who are the other major players?

sir that's display supply chain and NAURA has little presence in this. there are different companies like Piotech, major supplier of CVD. Leibo Microelectronics is the supplier of flat-panel display mask coating machine.. For a long time, this equipment relied on import but they successfully broke Japanese monopoly in this. and many more suppliers but NAURA/AMEC mostly in wafer fabrication equipment.

OLED main material FMM was so hard. they also broke Japan's monopoly in display material.

Please, Log in or Register to view URLs content!

Ink jet printer , CVD , PVD , large photomask , 365nm photoresist , mask coating machine..

 

[tokenanalyst]

This is a really good chart for the future. IIRC, NAURA is a major supplier for this?
Who are the other major players?

FPD market is pretty unknown for me but from my posts:

AMEC in the MOCVD market.

Please, Log in or Register to view URLs content!

SMEE in FPD lithography.

Please, Log in or Register to view URLs content!

SVG Optronics in lithography.

Please, Log in or Register to view URLs content!

Liande a lot FPD manufacturing tools.

Please, Log in or Register to view URLs content!

Sineva Evaporation Equipment

Please, Log in or Register to view URLs content!

Delphi laser

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

parts and components

Please, Log in or Register to view URLs content!

Materials

Please, Log in or Register to view URLs content!

FPD Metrology
Wuhan Jingce

Please, Log in or Register to view URLs content!

Here is more companies:

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

 

[tokenanalyst]

EFEM for lithography machines.

 

[JPaladin32]

I have a question. Does the report explicitly say the dies were fabbed last year and how did they assert that. For example does it use process nodes or features that were not available at TSMC at the time of cut off. I am saying this because it is very common to use dies that were fabbed in earlier years in chips launched much later. As we have seen with several Kirin chips Huawei launched after getting cut off they used stockpiled dies that were fabbed with TSMC pre-sanction.

Even Nvidia does that, at least for the consumer GPUs. They just bin the dies and use the lowest binned dies in the first year and save the better binned dies from the same batch to use in subsequent years as new products. For example Nvidia 3080 ( launched in 2020), 3080ti (2020), 3090 (2021), 3090ti (2022) are all from the same wafer batch of the same die.

I tend to agree with you and share some suspicions, but my argument is less on yield but more on that I think it's impossible that TSMC didn't know who the customer was. First, the claim is that the process is N7+, which went into mass production in 2019, so this is a pre-sanction process. Then, 910B is a humongous and monolithic die of 660mm2. This is so big that even Intel and AMD server chips are much smaller than it, and the only die I can think of that is even larger is from nvidia H100 and B100, of 800mm2. In fact, I cannot think of a third company that mass produces dies with this size, so it's hard to imagine that TSMC just fabbed those in large numbers without being suspicious. Or, TSMC fabbed the chips, knowing Huawei was the customer? That's an even more interesting story.

But anyway, now that 910C is out, it's probably not super important to know the answer.

 

[gelgoog]

I do wonder if these report miss on some of the smaller Chinese fabs, which probably add together to more than HH and Nexchip's total foundry revenue. Like for example, SMEC, CanSemi, Shanghai Jita, CR Micro, Sanan and a bunch more.

In a lot of cases those smaller players are still using 8" wafers or even smaller. Quite often they make discretes or something similar for which 12" wafers make less economic sense. At least not in large wafer volumes.

I do know that CanSemi, CRMicro, GTA Semiconductor (aka Shanghai Jita), SiEn, had plans to make 12" wafer lines for ICs. These were likely meant for power or automotive applications. Of those CanSemi's 12" Phase I, II lines should be operational. That should be like 40,000 wpm in 12" wafers. Phase III should double that to 80,000 wpm by adding another 40,000 wpm capacity.

While CanSemi will still have less capacity than HH or Nexchip if they reach 80,000 wpm in 12" wafers, that will be more capacity in terms of area than smaller players like Vanguard Semiconductor.

Those 8" wafer lines are often still relevant. They are important for certain specialty processes. MEMS, SAW filters, discretes, etc.

 

[tphuang]

May be. But the problem I have with that is if they were indeed still fabbing with TSMC they shouldn't have had the yield and volume issues that are known about it. For example it was reported earlier that Bytedance ordered 100k 910B this year but they were only able to deliver 30k. There were also earlier reports of a 910B card with less active cores than they used to have when using TSMC dies, which indicates low yield. So if the report doesn't specifically say that they were made by TSMC this year (only just that they were fabbed by TSMC) then it makes more sense to me that these are just stockpiled dies.
And as for the TSMC volume usage there's plenty of other Chinese companies fabbing with TSMC. All I am saying is if Huawei was still activly fabbing with them OR those other companies where doing it on behalf of Huawei where's all those chips, why Huawei is not able to do high volume.

don't believe all the report. They got plenty of Ascend dies.

I tend to agree with you and share some suspicions, but my argument is less on yield but more on that I think it's impossible that TSMC didn't know who the customer was. First, the claim is that the process is N7+, which went into mass production in 2019, so this is a pre-sanction process. Then, 910B is a humongous and monolithic die of 660mm2. This is so big that even Intel and AMD server chips are much smaller than it, and the only die I can think of that is even larger is from nvidia H100 and B100, of 800mm2. In fact, I cannot think of a third company that mass produces dies with this size, so it's hard to imagine that TSMC just fabbed those in large numbers without being suspicious. Or, TSMC fabbed the chips, knowing Huawei was the customer? That's an even more interesting story.

But anyway, now that 910C is out, it's probably not super important to know the answer.

Of course they knew who it was. They wanted to make a quick buck and they made a lot of money off that.

In a lot of cases those smaller players are still using 8" wafers or even smaller. Quite often they make discretes or something similar for which 12" wafers make less economic sense. At least not in large wafer volumes.

I do know that CanSemi, CRMicro, GTA Semiconductor (aka Shanghai Jita), SiEn, had plans to make 12" wafer lines for ICs. These were likely meant for power or automotive applications. Of those CanSemi's 12" Phase I, II lines should be operational. That should be like 40,000 wpm in 12" wafers. Phase III should double that.

Those 8" wafer lines are often still relevant. They are important for certain specialty processes. MEMS, SAW filters, discretes, etc.

right, my point is that there are a lot of smaller foundries in China that if added together, I'm surprised the total isn't higher.

 

[tphuang]

Please, Log in or Register to view URLs content!

YASC's 8-inch SiC are expecting to go into mass production in May 2025. Expected to produce 30k wpm of substrate/epitaxy, 61m power modules.

Silan Micro added a new SiC production line, expecting to produce 288k MOSFET chip and 288k SBD chip