Regarding Meteor Lake only the CPU tile is fabbed with intel 4 the remaining tiles (GPU etc.) are fabbed by TSMC. This seems to suggest that yield/capacity at intel 4 fabs is rather low.
Chinese semiconductor thread II
- Thread starter vincent
- Start date
Intel usually operates several fabs to make their CPUs. And right now they only have one large serial fab to make those Intel 4 chips plus the pilot fab which probably moved to make Intel 18A.
Worse of all is that the Intel 4 laptop chips they released just a couple months ago are supremely worse than the TSMC 3nm chips they released this month. Intel just Osborned their Intel 4 chips. Would you want one of those Intel 4 laptop chips knowing the TSMC 3nm ones are available?
Their management of their product lineup and release schedule is just plain retarded. No wonder they are hemorrhaging cash.
What's ridiculous is that you found a lot of western/American self-congratulators just a couple years ago because of China's stumbles with their own semiconductor funding
(This is actually an objective article, but there are others with a lot more back patting and high fives)
Now you have Intel with basically a bunch of bad assets funded by the CHIPS act. The latest news of them wanting to cut them loose probably necessitates another round of government funding for someone to be a caretaker of the bad assets (underutilized fabs and half built mega fab in Ohio) while Intel switches to a fab-light company structure.
It is a tough business, going from market leader to drowning in about 10-15 years. As I mentioned, even if Chinese companies were to perfect EUV overnight, it's not a Messiah. Intel could not even get good yield from their DUV 10nm (Intel 7) for years, so they didn't even make use of EUV even though they had it. In this sense, SMIC is very close behind Intel. You can see very easily why the Taiwan authorities are so fearful with their heavy-handed laws.
(This is actually an objective article, but there are others with a lot more back patting and high fives)
Now you have Intel with basically a bunch of bad assets funded by the CHIPS act. The latest news of them wanting to cut them loose probably necessitates another round of government funding for someone to be a caretaker of the bad assets (underutilized fabs and half built mega fab in Ohio) while Intel switches to a fab-light company structure.
It is a tough business, going from market leader to drowning in about 10-15 years. As I mentioned, even if Chinese companies were to perfect EUV overnight, it's not a Messiah. Intel could not even get good yield from their DUV 10nm (Intel 7) for years, so they didn't even make use of EUV even though they had it. In this sense, SMIC is very close behind Intel. You can see very easily why the Taiwan authorities are so fearful with their heavy-handed laws.
ASML CEO says US desire to restrict exports to China 'economically motivated'
New ASML CEO Christophe Fouquet says:
Translation: You (US) want us (ASML) out of China market just because for your companies the writing is already on the wall. You just don't want US companies to pay the heaviest price...but this thing was started by you, not by EU/Japan, and it started in a fully unilateral way (i.e. you didn't even consult us.....but TBH it has always been like this in US vs EU relations).
Translation: this national security stuff is obvious BS, you can't really think we give up China market just because of this fairy tale.
Translation: this is the most clear sentence. Mainly he says, ok let's negotiate, but don't expect you come back home full handed, we have to find a compromise.
My opinion: I was waiting from the new CEO, the old one was quite combative, the new one (that is French) seems a tough nut to crack too. He doesn't want to be remembered as the one who sunk ASML.
OTH is not up to the ASML CEO to negotiate with US officials, he can do this only indirectly through the Dutch government...so far from clear how all this will go. Personally I am still quite pessimistic on the final outcome.
Last edited:
Intel's Irish fab started delivering "Intel 4" Meteor Lake chips in January this year.
That fab only started manufacturing 4nm chips in August 2023.
And now Meteor Lake was obsoleted by Lunar Lake with TSMC 3nm.
So what is this fab supposed to be making right now? Their press releases claim "Intel 3" server chips. Well good luck with that. Making huge chips on a whole new process.
That fab only started manufacturing 4nm chips in August 2023.
And now Meteor Lake was obsoleted by Lunar Lake with TSMC 3nm.
So what is this fab supposed to be making right now? Their press releases claim "Intel 3" server chips. Well good luck with that. Making huge chips on a whole new process.
Huawei Technologies and its chip making partner Semiconductor Manufacturing International (OTCQX:SIUIF) are struggling with production of the chips used in Huawei's next flagship phone Mate 70, The Information reported, citing people with knowledge of the matter.
The root cause of the issue is the continuing impact of a four-year-old U.S. ban on the supply of chip manufacturing tools to Huawei and SIUIF, the report added.
Huawei is planning to launch new products at an event on Sept. 10, just hours after Apple (NASDAQ:AAPL) unveils the new iPhone 16 on Sept. 9.
Huawei wanted to procure at least 2.5M chips ready before releasing Mate 70, and was planning for a September release. However, due to the current manufacturing capacity and productivity, the company will not meet that target, which is needed to meet expected consumer demand for the phone, the report noted.
The company could still announce the phone next week but manufacture only a small number of phones immediately available for buying, and then let customers pre-order and wait for delivery. Huawei employed a similar tactic last year when it unveiled Mate 60. It is also possible that Huawei could delay the release of the Mate 70 series until November or later, according to the report.
The chip inside the new phone is a new generation of mobile phone processor. The chip is manufactured at the most advanced chipmaking process available in China called N+3, which Huawei and SIUIF hope is comparable to the 5nm process by Taiwan Semiconductor Manufacturing (TSM), the report added.
TSM's 5nm process is about two years behind, being used to manufacture the chips for the iPhone 12, 13, and 14.
The N+3 chips are smaller, faster, and consume less power versus prior generation of chips. Huawei designs its chips while SMIC makes them, as per the report.
SIUIF uses outdated and inefficient tools to make advanced chips because of not being able to access essential U.S. equipment due to the ban. This has resulted in higher manufacturing costs and lower productivity. The Mate 70 chips SIUIF did manage to make had a high defect rates. As many as 60% of the chips on each wafer were considered unusable, the report noted.
Source: TheInformation on SeekingAlpha
----
Thoughts? I trust TheInformation. Nonetheless, initial production of N+3 120sqmm Kirin 9100 chips was known to be delaying Mate 70 release date into Q4 2024. We also should expect initial production of SOTA nodes to have yields around 50%. With more production and knowledge gain, this yield will rise closer to 80% over time. Why? SMIC only began producing Kirin 9100 in Q1 2024. Getting yields on 5nm class chips up to 50% with 2000i ASML machines in such a short time is nothing short of astounding.
This is an enormous success for China and all Chinese.
The root cause of the issue is the continuing impact of a four-year-old U.S. ban on the supply of chip manufacturing tools to Huawei and SIUIF, the report added.
Huawei is planning to launch new products at an event on Sept. 10, just hours after Apple (NASDAQ:AAPL) unveils the new iPhone 16 on Sept. 9.
Huawei wanted to procure at least 2.5M chips ready before releasing Mate 70, and was planning for a September release. However, due to the current manufacturing capacity and productivity, the company will not meet that target, which is needed to meet expected consumer demand for the phone, the report noted.
The company could still announce the phone next week but manufacture only a small number of phones immediately available for buying, and then let customers pre-order and wait for delivery. Huawei employed a similar tactic last year when it unveiled Mate 60. It is also possible that Huawei could delay the release of the Mate 70 series until November or later, according to the report.
The chip inside the new phone is a new generation of mobile phone processor. The chip is manufactured at the most advanced chipmaking process available in China called N+3, which Huawei and SIUIF hope is comparable to the 5nm process by Taiwan Semiconductor Manufacturing (TSM), the report added.
TSM's 5nm process is about two years behind, being used to manufacture the chips for the iPhone 12, 13, and 14.
The N+3 chips are smaller, faster, and consume less power versus prior generation of chips. Huawei designs its chips while SMIC makes them, as per the report.
SIUIF uses outdated and inefficient tools to make advanced chips because of not being able to access essential U.S. equipment due to the ban. This has resulted in higher manufacturing costs and lower productivity. The Mate 70 chips SIUIF did manage to make had a high defect rates. As many as 60% of the chips on each wafer were considered unusable, the report noted.
Source: TheInformation on SeekingAlpha
----
Thoughts? I trust TheInformation. Nonetheless, initial production of N+3 120sqmm Kirin 9100 chips was known to be delaying Mate 70 release date into Q4 2024. We also should expect initial production of SOTA nodes to have yields around 50%. With more production and knowledge gain, this yield will rise closer to 80% over time. Why? SMIC only began producing Kirin 9100 in Q1 2024. Getting yields on 5nm class chips up to 50% with 2000i ASML machines in such a short time is nothing short of astounding.
This is an enormous success for China and all Chinese.
btw, just browsing on this topic, because I hear that the challenge for scaling up clusters is power rather than money.
So this is some decent math out there for everyone on just how much is needed to run a 100k H100 cluster.
America and China has two problems in building large AI clusters to train models. The former needs to find abundant enough power and factory space for these large data centers and the latter doesn't have enough chips.
Another good article here about the multi-datacenter training approach carried about by OpenAI, Google and Anthropic now that there are somewhat limitations of putting more than 100k GPUs per data center. This entire thing honestly starts to sound like China's EDWC project which uses 800G pipes on data.