You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Sept 4 (Reuters) - Intel's (INTC.O), opens new tab contract manufacturing business has suffered a setback after tests with chipmaker Broadcom (AVGO.O), opens new tab failed, three sources familiar with the matter told Reuters, dealing a blow to the company's turnaround efforts.
The tests conducted by Broadcom involved sending silicon wafers - the foot-wide discs on which chips are printed - through Intel's most advanced manufacturing process known as 18A, the sources said. Broadcom received the wafers back from Intel last month. After its engineers and executives studied the results, the company concluded the manufacturing process is not yet viable to move to high-volume production.
Sept 4 (Reuters) - Intel's (INTC.O), opens new tab contract manufacturing business has suffered a setback after tests with chipmaker Broadcom (AVGO.O), opens new tab failed, three sources familiar with the matter told Reuters, dealing a blow to the company's turnaround efforts.
Last month, Intel provided an update on its 18A progress, saying it's out of the fab, powered on, and booting an OS (I wonder if booting to BSOD counts)
This Broadcom article says that Broadcom received the 18A wafers last month.
The tests conducted by Broadcom involved sending silicon wafers - the foot-wide discs on which chips are printed - through Intel's most advanced manufacturing process known as 18A, the sources said. Broadcom received the wafers back from Intel last month. After its engineers and executives studied the results, the company concluded the manufacturing process is not yet viable to move to high-volume production.
You do realize that nobody ever claimed that by August 2024 18A would be in HVM or be ready for HVM, right? Now if BCOM was seriously considering 18A and they dropped out because of what they were seeing that would be a completely different and unambiguously bad story. But coming to the conclusion that 18A wasn't HVM ready is something of a nothing statement since even intel says it wouldn't be HVM ready until the end of the year.
"...and we'll start production wafers with Panther Lake before the end of the year. So, we're on track to deliver that." - Pat at Deutsche bank
You do realize that nobody ever claimed that by August 2024 18A would be in HVM or be ready for HVM, right? Now if BCOM was seriously considering 18A and they dropped out because of what they were seeing that would be a completely different and unambiguously different story. But coming to the conclusion that 18A wasn't HVM ready is something of a nothing statement since even intel says it wouldn't be HVM ready until the end of the year.
"...and we'll start production wafers with Panther Lake before the end of the year. So, we're on track to deliver that." - Pat at Deutsche bank
Broadcom certainly knows what to expect from Risk Production wafers, and as the article suggests, Broadcom was not thrilled about these risk production wafers.
Broadcom certainly knows what to expect from Risk Production wafers, and as the article suggests, Broadcom was not thrilled about these risk production wafers.
1) 18A isn't in risk production yet as Dan has said that "HVM ready" correlates to risk production. This is consistent with the time from HVM ready to product release on intel 4 and intel 3.
2) Where in the body of this article did it say BCOM was disappointed with the risk production wafers? The claim is that 18A is not yet ready for mass production. The title is the only place that uses the word "disappointment", and it seems to be speculation from Reuters as no where does that language show up in the findings from the BCOM testimonials they claim to have been hearing.
3) Could they be disappointed with the current state? Maybe, but nothing in this article would definitively give that indication. As an example everyone assumes N2 will be good (and it probably will be) but it is also not ready for HVM and nobody is disappointed with that (it just is TSMC following their roadmap). Concurrently if you looked at intel 10nm in 2015 that was also not "HVM ready" either but you would be very right in being disappointed by that given how far behind schedule intel was and how long it would take to enter a risk production state. The devil lies in which part of those extremes 18A lies.
1) 18A isn't in risk production yet as Dan has said that "HVM ready" correlates to risk production. This is consistent with the time from HVM ready to product release on intel 4 and intel 3.
2) Where in the body of this article did it say BCOM was disappointed with the risk production wafers? The claim is that 18A is not yet ready for mass production. The title is the only place that uses the word "disappointment", and it seems to be speculation from Reuters as no where does that language show up in the findings from the BCOM testimonials they claim to have been hearing.
3) Could they be disappointed with the current state? Maybe, but nothing in this article would definitively give that indication. As an example everyone assumes N2 will be good (and it probably will be) but it is also not ready for HVM and nobody is disappointed with that (it just is TSMC following their roadmap). Concurrently if you looked at intel 10nm in 2015 that was also not "HVM ready" either but you would be very right in being disappointed by that given how far behind schedule intel was and how long it would take to enter a risk production state. The devil lies in which part of those extremes 18A lies.
Risk Production is generally the state when you can manufacture a full wafer of the same part (as opposed to Multi-party shuttles or R&D efforts) for the first time and expect to get reasonable yields out of the fab.
I suspect these were not Multi-party wafers, but I could be wrong, although doubtful.
Further, Intel has stated 18A is currently manufacturing ready, and that manufacturing is to begin in 1H25, but maybe that's also slipping?
Either way Broadcom seems to be dissatisfied with 18A, at this point in time.
Risk production is where the products could be used for production sales. 18A is not in risk production and the process is not qualified yet. Intel has run products through 18A but they are not risk production... they are checkouts. I dont believe Intel stated 18A is manufacturing ready. If they did, they misspoke. Intel also will do at least one more stepping on both new products before they are sold as they do not have ES back yet.
The good news is that a lot of companies and IP vendors tried 18A. The bad news is a lot of companies and IP vendors tried 18A. TSMC N2 wins this round by unanimous decision.
TSMC A16 is the next fight. Let's hope Intel and Samsung do better.
Economic Daily reporter Yin Huizhong/Reporting from Taipei TSMC (2330)’ s most advanced angstrom-level A16 process “made a sensation before mass production.” According to industry reports, not only has major customer Apple booked the first batch of TSMC's A16 production capacity, but OpenAI, a...
It was the PDK.......
The good news is that a lot of companies and IP vendors tried 18A. The bad news is a lot of companies and IP vendors tried 18A. TSMC N2 wins this round by unanimous decision.
Dan -- you are right. If customers designs are to be released in 1H2025 -- IP design & silicon verification should be complete by now. (or near complete)
Pat G. said a few days ago that IP design/verification was a bigger challenge than what Intel had anticipated.
Here is the quote from Intel / Pat G: "The second piece that's been disappointing is just the -- we underestimated, I underestimated the amount of heavy lifting beyond producing good wafers the EDA, the IP ecosystem that
needs to get enabled to bring designs on to the foundry. So those have been the two areas that in this current environment have been a bit harder than I would have expected."
The article also indicates that the issue, as one would expect, is yield. If so your “boot to BSOD” dig is very naive - early stage processes produce good die, even with poor yield. And leaks like this about yield, anre often balled up in price negotiation strategies, especially where Hock Tan is involved.
“Broadcom's engineers had concerns with the viability of the process, the sources said. Typically that refers to the number of defects on each wafer or the quality of the chips fabricated.”
Or we can say it is simply the user experiences. no matter how great you think Intel 18A may be, it is just that the user (Brodcom) who gives it a try feels disappointed, as compared with the service/ product another foundry tsmc can provide
It's possible that it was leaked out by former Intel employees. But typically Routers' policy required to have multiple verifiable sources to check the agreement and disagreement due to the sources don't want to be named in public.
That makes me believe that it's coming from the multiple camps inside the Intel's senior leadership team and Intel's Board of Directors. They are talking to Routers and Bloomberg to promote their own Intel rescue plans.
By the way, it does not say that Broadcom will not use 18A. PDKs are easily improved and I'm sure it is "maturing" as we speak. I have no doubt 18A will be in HVM next year. I just think TSMC N2 will win the node like N3 did, minus a couple of customers namely Intel and possibly some Broadcom and Microsoft products. TSMC N3 was a 99% market win. TSMC N2 may be 90% but the margins will be better due to price hikes.
Not sure this conversation is particularly enlightening for 2 reasons:
We're all over the map on the proximate issue based on the leak:
News article implies yield
Dan: it's the PDK
ChrisGar: Pat G. says it's IP design and validation
Kevin: Perhaps it's part of a price negotiation strategy
Hist: or maybe internal camps within Intel advocating different Plan B's
Second, we're only looking at one point in time. Broadcom also says they have "not concluded their evaluation (or purchase decision)". Historically, new TSMC processes, PDKs and enablement haven't jumped out of the fab completely ready at PDK 1.0. TSMC's strength has been the meticulous and dogged grinding down of the causes of end-customer issues, both individual, and later at the broader specific market level. That has enabled them to build an efficient, customer and ecosystem-engaged 4 year process for rolling out new nodes, that catches and fixes the gotchas earlier than would traditionally be found. So the real question is how well Intel listens to external customers and and how fast they improve, going forward. Upfront, all the things that TSMC needed to do for the PDK and enablement sound like overkill, especially to foundries trying to compete, but long-term, they have refined the lowest-cost path to success to a resilient, but expensive, enablement process.
By the way, it does not say that Broadcom will not use 18A. PDKs are easily improved and I'm sure it is "maturing" as we speak. I have no doubt 18A will be in HVM next year. I just think TSMC N2 will win the node like N3 did, minus a couple of customers namely Intel and possibly some Broadcom and Microsoft products. TSMC N3 was a 99% market win. TSMC N2 may be 90% but the margins will be better due to price hikes.
If tsmc grab 90% share of N2 market, how significant is a 10% share of the N2 market? Given that the N3 family is still ramping up, the N5 family might offer a better reference. Four years after TSMC’s N5 began HVM, the N5 family now accounts for about one-third of TSMC's revenue, which is roughly $25-30B of their projected ~$90B for 2024.
Considering potential price increases for N2, it's reasonable to estimate that the N2 family (N2/N2P/A16) could generate double that revenue—around $50-60B—four years after its launch in Q4 2025. If Intel captures 10% of the N2 market by 2029, that would translate to approximately $5-6B in revenue, aligning with my previous estimate on how Intel could boost its revenue.
If tsmc grab 90% share of N2 market, how significant is a 10% share of the N2 market? Given that the N3 family is still ramping up, the N5 family might offer a better reference. Four years after TSMC’s N5 began HVM, the N5 family now accounts for about one-third of TSMC's revenue, which is roughly $25-30B of their projected ~$90B for 2024.
Considering potential price increases for N2, it's reasonable to estimate that the N2 family (N2/N2P/A16) could generate double that revenue—around $50-60B—four years after its launch in Q4 2025. If Intel captures 10% of the N2 market by 2029, that would translate to approximately $5-6B in revenue, aligning with my previous estimate on how Intel could boost its revenue.
The big change is the moat that is around TSMC. Not just capacity, PDKs, IP etc., but also packaging. The NOT TSMC business is definitely shrinking but it will not go away, my opinion.
Samsung and Intel will chip away at the 5nm, 4nm, and 3nm business over time but the lions share goes to TSMC and that means profits not just revenue. I don't think Samsung has made money on the foundry business since 14nm so how do you think Intel will? Let's get real here. There are three dogs eating out of the same bowl. One great dane and two chihuahuas.
