Це відео не доступне.
Перепрошуємо.
The Semiconductor Bust. Still Coming?
Вставка
- Опубліковано 14 сер 2024
- A year ago, I posted a video about business cycles in the semiconductor industry. I titled it the “Coming Semiconductor Bust” and people seemed to have enjoyed it.
Now that a year has progressed, I thought some of the topics might be worth revisiting. I have been thinking about a few things recently.
So I want to run some numbers, talk some recent news items, and make a few predictions about what might happen in the coming years.
Go check these amazing resources out:
- Fabricated Knowledge: www.fabricated...
- SemiAnalysis: semianalysis.s...
- Redfire on Twitter: / redfire75369
Links:
- The Asianometry Newsletter: asianometry.com
- Patreon: / asianometry
- The Podcast: anchor.fm/asia...
- Twitter: / asianometry
Enjoy the video.
ok thanks.
thank you Jo(h)n. You are the best youtuber for research, fact-finding, documenting and humour. Really appreciated :)
This is incredible. Yesterday youtube recommended your previous video on this topic and wondered exactly what has changed since.
Thanks for the amazing research and content.
done that, thanks
dont tell me how to live my life!
Just wonder about how ASML saying that they weren't going to get all the machines out they promised until 2024.
Well, that works pretty well then because that's when these fabs being built by Intel and TSMC are due to be fitted with the equipment, so perfect timing huh?
It might cause a delay of 3 - 6 months but in the general scheme of things won't have much of an impact.
@@johndoh5182 Except they promised them for the end of this year. So that doesn't work.
I've just started following Asianometry (the name threw me off when I first ran across it), and I've watched about a dozen of the videos. I find John's descriptions and explanations very useful and his style very approachable. Thanks, John.
I love that the volume is so much louder (more normal) in recent videos. Not a fan of ASMR narration. The content is outstanding as usual.
But why did it take so long?
He is shouting
@@janeblogs324 because not many dare to mix everything to -6dB at first because it seems too loud, hearing their own voice
My voice hasn't been well lately. Sorry if it sounds a little strained.
An episode about RISC-V would be interesting.
Yes. Is it still relevant or not relevant yet?
Agreed. It has potential to unseat ARM down the road in specialist areas like network switches within the next 5-7 years. Don't think it'll make a dent in smartphones unless Apple gets to the point they see the ARM license and ISA as too inflexible for their future SoC designs.
@Zaydan Naufal Indeed. But it can make in ways. I know of one small specialist industrial test equipment manufacturer that's doing R&D with RISCV right now with final product expected in 2024 Q2. So it's starting already at least in the low volume specialist areas.
It's already in some controllers, so it is coming.
Every time I watch your videos, I feel a little smarter about a subject I knew nearly nothing about. I appreciate the lessons, thank you.
I have to buy a new hat after watching one of his videos. 😀
Thanks. BTW the shoe manufacturing episode was enlightening. Keeps things fresh :)
I liked it too!
Hey, I loved your shoe video! Keep doing what you find interesting: your audience will get catch up over time. Think of the semiconductor videos as the gateway into broader economics, manufacturing, and supply-chain videos. Never thought I’d be fascinated by supply-chains (for example), but here we are. 😄
There is a possibility that the extra logic capacity could be taken up by 3D stacking. If there's a need to grow vertically, then your area requirement scales with the number of layers. So this may be the perfect time to be adding logic capacity. And if it's cheap enough, the extra capacity could always be taken up by SRAM, like AMD's V-Cache (again, 3D stacking though).
3D stacking isn't the issue. It's getting to these smaller nodes that use much less energy. Intel's new fabs are supposed to be 2nm. What TSMC is building is supposed to be 5nm, but at the same time they're fitting fabs in Taiwan for 3nm.
TSMC N3 is about triple the density of N7, and that's what AMD is on right now with Zen 3. Stacking is something that can help, but that's not why all these companies are building new fabs. Intel and TSMC want to be part of the BOOM cycles that gets created when they put out these new nodes. It will be unprecedented.
@@johndoh5182 I wasn't suggesting that 3D stacking was an issue, just that it could easily eat up any production surplus. In some cases, it may be more advantageous to use 4 layer 3D stacking at 7nm than move to 3nm, especially if the price is right.
Ev market is growing with Logic and dRam. It takes over 100 GB of Dram per EV. Seems like everything (Logic/Dram) is increasing with faster broadband. Great video !
Here is my obligatory yell at you on the comments. Good update. I feel I’ll have to rewatch as it was late at night for first viewing. Love your style and research you put into all your topics. Cheers!
Nice video - AWS grew 36.5% YoY per their q1-2022 report. MS Azure grew 46%, and Google Cloud revenue grew 43.8% YOY. Small slowdown for each all things considered, and still enviable growth rates given how most other growth stocks and even some big tech are getting slammed. The big cloud providers should provide good business for some time it seems? (And in terms of marketshare: "According to Canalys, in the fourth quarter of 2021, AWS represented 33% of the cloud infrastructure market, Azure held 22%, while just 9% of cloud spending was on GCP.")
The most knowledgeable and informed deer I know of in the semiconductor industry.
Long time viewer and also a big fan of the semi ecosystem. Great video, but I think the one major flaw in your argument is that you argued with the assumption that datacenter growth only received a major boost because of COVID. If you look at the decade prior to COVID, datacenters was the primary growth driver for Amazon, Google, and Microsoft from the onset. If you look at the growth rates from COVID and also compare it to pre COVID, the growth rates have been consistently 30-60% (depending if you're talking about AWS/Azure/GC) for nearly a decade. It COULD slow down coming out of pandemic, but there's no historical evidence that the pandemic even affected it's original trajectory.
I do agree with you that TSMC and Samsung's customer ecosystem will remain strong, while fabs producing more mature nodes might see some commoditization and destruction to their profit margins. Intel is a gamble, they may join the ranks of TSMC and Samsung as a foundry, but I also think they're underestimating the total cost (of producing in western nations) and over-optimistic on their roadmap.
Hi Asianometry
I'm always really impressed when you make a video about a topic that is not your field. In the spirit of teaching a man how to fish, how do you approach new topics, research them and create a knowledge base for yourself so quickly? Maybe an idea for a video?
Always look forward to your videos and commentary. Keep up your fine work! 😃
There's a tidal flood of demand for SSDs being barely held back by price. (For example: some desktop computers are still being sold with 1TB hdds, just waiting for the price of SSDs to reach a tipping point.) Couldn't any glut of memory production capacity be applied to that latent demand for SSDs? (It's NAND vs. DRAM, but still.)
It looks like 1TB SSDs will soon match the price of 1TB HDDs (to OEMs).
@@nathangamble125 honestly man it's pretty really really close even at the retail end. I can buy a 1tb Seagate hdd for $47 and a 1tb SSD for 60$ish... If it comes down 5$ retail price the OEMs has no excuse to continue with hard drives
That's likely more of a marketing issue. For oem and si customers the larger capacity storage is easier to market then the performance of an ssd.
You had put so much soul into the shoe video! It was interesting and insightful.
Its not just cloud demand. Automation, edge computing and electrification of transportation, smart medicine, smart and connected housing would provide extremely strong demands in the next coming decades.
Thanks for the analysis, I am here to say thanks, not yelling at you. TSMC provided good guidance for the remaining of 2022. Next year onwards will be more uncertain, I don't think anyone can see that far away.
I recently saw that before the last memory crash, the cap ex of memory to logic was 60 to 30. Due to consolidation and massive Logic cap ex this trend is now exactly reversed with logic spending 60% of cap ex and memory 30. Could that not make Memory less vulnerable than logic to a supply glut?
Yeah he doesn't understand what's happening. First there's a shortage of memory for what's about to happen because in the world of computing the move is away from DDR4 to DDR5, so the demand for memory is going to be great for the next few years.
Terrific and very honest humour in the opening! And personally I'm a big fan of your narration style.
I regularly watch ur videos, u provide great insight, keep up the good work man 😃
I'm yelling at you now for how amazing your content is!
I really loved your Taiwan making shoes episode!
Great video, I also liked the Taiwan shoe making video. Also love, but don't fight the haters, as it boosts engagement which is good for the algorithm.
An interesting video would be a follow the products from a fab that produced for a long time. E.g. what is produced on them after the leading edge moved somewhere else, and what is produced at the end of the fabs life.
Its always enjoyable to see and hear some sanity in this crazy mixed up world. Thank you.
if u can, do a video on analog computing / RRAM / artificial synapse / analog floating gate transistors / etc hehe thanks
No yelling from me. You did quite a good job again.
How much of the build-out is companies repositioning their fabrication to shorten supply chain and improve security of supply i.e. onshoring? Will the overseas plants that are left behind be able to repurpose themselves to non logic/ memory areas or will they be removed from the market as they lack orders?
That's one side of why companies are building fabs, but if they felt it was a losing proposition they wouldn't have agreed to build these fabs. These are many billions of dollars, EACH. You don't spend that kind of money without understanding what the growth will be, and the guy making the videos isn't comprehending what's going to drive growth. TSMC is going to be the main benefactor though, and if Intel is ready with either their 3nm or 2nm node by the end of 2024 or early 2025 they'll enjoy one of the biggest boom cycles this industry has ever seen.
If you want to understand it, I posted a long comment to the video. Sorry, it has to be long to drive the points. The jist of it though is turnover of current equipment, not new growth. And this will effect the world of server, laptops, PCs, pretty much everything involved in computing.
Love your channel and watch all your vids.
Not sure I agree with your take on Cloud being due to level off - Cloud infrastructure is different to SaaS companies like Netflix that offer services using Cloud. Use of cloud by total addressable market is still astronomically low - most IT is still run more expensive and less efficient on-prem, so unlike mobile, Cloud is a long way from saturation, or even growth leveling out - I think this is why so many players still think it is not too late to get into it (IBM, Oracle, etc). Not saying it will be the hottest or most rapid growing driver for semiconductors but hard to see that growth dropping off for a while.
Still, that is also a prediction so... :)
I really liked the shoe video! I hope it came across in the comments there as well, but I always find it so fascinating when so much of an industry is quietly dominated by one dedicated group that found a niche decades ago and exploited the hell out of it.
Electronics engineer since 2006, seen a few cycles and they all kick my butt
Not only is the cloud growing, but it will all have to be rebuilt.
i appreciate you sharing your knowledge, thank you.
Thanks for being brave enough to share your opinion.
Consider Apple silicon integrating "unified" memory on the SOC, shipping millions of devices that can't expand dram throughout the life of the product. And they are not looking back. Most likely, this design will become the norm for laptops and a lot of desktops even outside of Apple.
there are active projects that aim to counter it though - the framework laptops or fairphones
@@aravindpallippara1577 but the masses wont bother with fairphone or framework laptops.. they'll go to big box store and buy an iphone or an HP laptop or whatever.
God I hope not! Luckily, Apple has been the only large hardware limiter for some time, in various aspects...hopefully others never follow suit! One of the many reasons why I'll never own an Apple product
it does offer some extreme benefits - so its something im far less jaded about. Making better integrated RAM can offset the need of execessive amounts of extra ram less necessary. apple's new SOC smokes a I9 when it comes to things like compiling code using way less power.
I thought it was bad too, but realized phones, video cards, game consoles and maybe 30% of laptops already have soldered memory and no one thinks twice of it (except about laptops). It's PCs and servers that are the exception. Servers will probably keep RAM slots for a long time, but who knows about PCs. If HBM is added to the CPU, will there even be a reason for RAM if storage gets fast enough?
I liked your shoe episode -unfortunate that it didn't do so well.
Hi, can I ask that you to pronounce DRAM as Dee-RAM rather than Dra-am? I'm a DRAM guy, and no one pronounces it as Dra-am. Thanks
Semiconductors is the only human endeavor to suffer a surplus of logic. On the other hand, humanity has constant glut of reminiscence, nostalgia and other cheap error prone memory. ;-)
2 months after this video the world is bracing for a recession, a huuuuge crypto bust, and a super glut of gpu's. Man, the world changes so fast that I'm witnessing things at a pace that I never could imagine.
Great insight! Tks John
the show vidoe was interesting to me, i watch all of your video because of how thoroughly you go into topics and economic behind them, you nade me realise not all government actions are automatically bad
Another amazing episode. Thank you for sharing!
You should do these annually and honestly compare predictions with reality, not in an "I told you so" or "oops, missed that one" kind of way but to fill in the gaps where things diverged.
I like the way he say DRAM. He should have patented it.
Just wanted to say that you've got a shout out in the Aquired podcast, on their latest episode about NVIDIA
Many reasons to believe you, yet few to doubt your reasoning. I don't honestly understand where the data center growth is coming from, unless it is emerging markets. I would expect replacement value in current cloud systems in the USA at any rate would not be many more times the cost, since compute density is also ever increasing.
Please cover what a full scale Taiwan invasion would mean for semiconductor manufacturing and R&D.
Very good. I am subscribing
Aaaw, I loved the shoe video!
Fiiiine. I’ll watch the shoe one 👟 👀
DRAMMM a semiconductor video!
What I always thought was the driver in new production in the USA and the EU was the possible reality that China would invade Taiwan and their ability to supply the West was going to be hijacked or destroyed. That an China as a supplier would be a no go. Kinda like having 2 bathrooms in a 3 bedroom house. It might not be needed but sure is a benefit to have one extra.
2:52 Interesting that 2019 was also "the best time to start PC gaming" as it was also this year when Techtuber videos exploded in popularity (LTT, J2C, GN, HUB, etc.). So when the suits are sad for all the revenue losses, consumers like us rejoice for all the cheap hardware.
I don't think video cards are supposed to be as cheap as the original SRP of the current RTX lineup.
Was it that year? I felt like PC gaming finally got mainstream around 2016 and then Ryzen dropped and basically everyone had a fucking gaming computer 🤣
How do you think the progress in AI will influence the demand for chips?
How come almost all big company's stocks are dropping by quite a lot if what you're saying here is the case? I can't find a singular big high-tech company that didn't drop tens of percentages in the past few months-years.
I loved the Taiwan shoe reading, thank you for that
Sorry to be that guy, but you should know that it's usually pronounced "D Ram", not "Dram" like the unit of water. Dee Ram. D-RAM. Dynamic-RAM. I've never heard it said like you do before. :)
Edit: Also, same is true for other types of ram. Like static ram being called S Ram not Sram.
OK smart guy.
I approve of you being that guy, so we don't have to :-)
I don't think the point is regarding chip supply bust or shortage. It is about profitability. I would definitely buy a 3090 ti if it was 500$ or less. The real cost of such chip is like around 300-400$ or less disregarding the R&D cost. Do the chip companies want to sell at this price? That is the question. Also if the likes of TSMC want to keep price of 5nm and 3nm sufficiently low. That's the real question. I don't see in any way how there can be surplus of 3 or 5 or 7 nm nodes. That is unfathomable as there is demand for high performance always.
When TSMC releases N3 to the world in general and it not be only available to risk partners, the turnover in compute products is going to be on a scale that I don't think the world has ever seen because of how much better it is than N7. It's 3X the transistor density and something like 40% more power efficient. The many worlds of server will be buying parts as fast as they possibly can. I don't know how much N5 will drive demand, but I KNOW N3 is going to. And it also happens to align with development cycles for AMD, Nvidia and a few other companies. The stars are going to align, and demand will be HUGE.
Home users will benefit too. TSMC has said that their N5 and N3 should help bring prices down simply because of transistor density and how many more die can be had from a wafer. But they did raise prices for their advanced nodes. Intel wants to be on 2nm by 2025. If they are this will bring competition back to high performance nodes and will help keep TSMC in check.
A founders 3090ti in total costs in the neighborhood of $700-800 to manufacture and ship to your doorstep. It's not a cheap product by any means. Nvidia for sure loses money on their founders 3060ti and possibly 3070. The next generation will be even more expensive due to the extremely expensive (in comparison to samsung) tsmc node not to mention higher power requirements. New design paradigms are coming due to just how high power consumption for such parallel compute devices is becoming. I predict ~2025 will be a tipping point of converging technologies creating truly astonishing levels of performance.
Please do a video on the economics of food production and supply.
The data center is impacted, but new edge compute
30 years ago semiconductor growth was driven by the personal computer, more recently it has been driven by the smartphone, I think the next massive sector will be automotive. Electric cars probably have 10 times the silicon content as do conventional ICE cars thanks to everything which is required for power management, charging and so on. Most of these functions are not neccessarily in need of the bleeding edge lithographic technologies, but they do eat up a lot of wafer area. That 26% recent growth in automotive is just a preview of things to come. Electric cars are still a very small percentage of the automotive market, but they are highly likely to get to at least 50% market share over the next five to ten years. Another profitable growth market (though primarily for US and European fabs) is going to be military spending. The volume isn't huge, but the profitability can be astonishing. Military and aerospace markets provide a massive portion of the early semiconductor industry's revenue and investment capital. Keep up the great work!
Why are some of the fabs colored in Mondrian designs?
Thanks for the quality content.
The enterprise move to the cloud will certainly continue, BUT in about 3 years there won't be any onprem servers left to move there. A few very big competitors that will either collude or compete strongly on price.
The cloud market can still grow even if no onprem servers are left to convert. The applications that have moved to the cloud, such as AI, can grow in compute demand and in areas in which they are applied.
Metaverse will be an interesting sector to watch out for in the coming years. The headsets need high performance mobile chips but with 5g streaming of cloud rendering we're bound to see some unorthodox demand on the chip market.
100% Both the front and back end are going to need a ton of chips. GPUs to render the space and CPU/AI to give realistic interactions.
@@sin3rgy Meta is even making custom chips to render hi-fi realistic avatars. That's probably some exotic stuff.
How many old fabs are going to be retired or converted and are now being counted as new capacity?
Wish I had watched your video a year ago =p
2009 silicon bust you say? We need video on that!
Your analysis was spot on before. The bust started some 3 months ago and affects logic. Intel has had a massive margin squeeze and AMD has started feeling it in DC GPUs.
On consumers, all prices have come down and low end laptops are now selling 15% below pre-pandemic or at then sale price. With the current inflation environment...
I follow semiconductor companies for two decades and your videos are great research on the matter. I can also tell you they come early and much more complete than what financial analysts do. Take pride, because you should.
I loved the shoe vid!
In the pie charts of TSMC’s revenue, is DCE Desktop Computer Equipment?
the "bust" will be quite slow and gradual though..
LOL ~ Thank you for your work. I'm not about to yell at you.
Enjoyable video.
I enjoyed the shoe show!
That moment you said demand for gaming hardware I knew something is off, look at how the crypto situation have been...
Cloud will keep having to be updated as ai and work loads increase. I have no worries
i am however expecting a big AI bust coming first.
what we have currently are very good pattern recognition - not a true decision making AI, which while exist (the starcraft video game AI made by google) they require very solid rules and boundaries (video game ruleset) and extensive computing power.
real world has strange rules and flexible boundaries unlike video games - and pattern recognition will only get you so far
IDK, But Apple just pushed my New MBP back for the THIRD TIME!! June 23rd now....
The human soul is combined of chaos and cosmos. Digital is on or off. Life is actually analog. Let that sink in.
Today Adafruit release a video on chip shortage in which they say they haven't received the order of 8000 ATSAMD51 they give in 2021-07-07 and if you give order today you will receive in 03-May-2023 direct from Microchip and Digikey do not allow backorder.
The answer can be found in the % of different types of traffic in the internet.
Do you know how the possible bust will affect the big chip companies or their stocks?
I too really enjoyed the Taiwan shoes video. Don't know why it didn't do well. :/
Another good video and no, I'm not to criticize you because there is no visible semibust. Your videos are Very Informative and Well researched so there is really nothing to criticize. Your opinion/conclusion is yours and your are entitled to it based on the data you had at that time, and that's end of conversion. But I would like to add that there is another factor that will impact Specially the Big data multinationals ... CPU increased Efficiency ... or better Computing Density ... Most of these company can replace their Servers from old say Intel E5 Xeon series generations to much more Powerful AMD Epyc Server line ... We tend to forget that there was indeed a jump in core-count And IPC in the last 4 years all lead by AMD in this case (no this is not a fanboy headsup ok ... )
Also servers tend to be replaced Less often then other hardware, with cycles of 5-6 years if not more ... therefore the situation we have right is curious,. .. Once those quad-core/sic-core double CPU E5 series Xeons are replaced companies will experience a Huge jump on Computing power ... Making their previous operated Data Centres cope with a Lot more web-centric activity.
This means that if they upgrade their Servers to say AMD ..remember AMD is clearly Delaying top of the line high-Cache CPU's in order to make those top silicon wafer to ... the Server market (Epyc/Theadripper) ...
This means that a Much smaller number of servers will do the same work that previous generation E5 type of servers would do.
Meaning ... Less top of the line Server CPU's are needed to compensate the increase in web-centric activity ...
If they replace only a small portion of legacy systems the increases in capacity will be huge unlike in previous generations ... let's face it Intel was sleeping on the wheel for about 8 years straight ...
And that is one argument that could also impact all of this "future demand" negatively.
You didn't factor in the move towards EVs, it is a large demand especially when some manufacturers and countries have already set their goals to hit certain targets.
And even ddr5 ram now also needs euv
@@mzamroni Think you are talking about EUV to make chips when the comment was about Electric Vehicles (EVs)?
idea: make a video about the mask writing process and ecosystem. Recently there was a leap forward with the multibeam writing technology. I'm sure you could wrap it all up in a nice video.
Depending on just how much of a factor machine learning plays, I wouldn't be too surprised if it floats memory for a while. There are a bunch of models now that don't run or don't run well with less than 16GB of VRAM. Some NLP models demand far more, requiring cards like the Nvidia a100's 40GB of VRAM just to load the weights.
I'm nowhere close to being an expert in any of that; just stuck feeling like a dingus around people who are, but just as a person who interacts with automated systems, I can see huge potential in advanced natural language processing. If Google "knew" what you were asking for while searching, it may yield better results, for instance.
That's perhaps a bad example, as Google is currently clearly more interested in showing you what it wants you to find rather than what you're looking for, but they might have a lot to work with if they ever pivot back. And if they don't, there's other problems like image recognition that are similarly RAM-hungry. If we ever get self-driving cars that aren't marketing fluff and anyone can buy, they might do a lot of local image processing.
The second models like we currently have find their way out of being unknown to most, a novelty or hobby for some and research for others, I'm guessing it'll set a new floor for memory demand.
And just in case I need to restate it, I'm purely a voice in the peanut gallery and no one should put their marbles behind a random UA-cam comment.
The big question is when/if China is taking the warpath and attack Taiwan. It could also disturb the production in South Korea that is close to China. So that would cause a huge disruption in semiconducting fabrication.
Huawei just posted a video on YT about their 5nm chip: "Huawei's new 5nm chip success: Kirin 9006C" ...how did they manage that?
Perhaps by lying about it
There was a memory crash in 2019 ? Thats funny to hear, SSD prices never went back to even their 2014 levels... price per Go is still a joke since then despite supposedly better nand available.
What is the image at 5:00?
There is not enough silicon ingot production for a bust
Statements for Intel CEO and a likes is for investor and to drive share price, company value up (short term gains). Not surprise if they are BSing and have no idea what's going on for half the time. They (people with decision making power) will miss all busts and booms in foreseeable future, while real engineers, researcher and other people in the know, will be ignored.
I OVED THE SHOE VIDEO. Wait, did I like it? brb.
_"Even if the cloud doesn't grow another 30, 40, 50% year over year however, things will be alright for the big foundries of the world and their new fabs. They can re-adjust their capacity pricing and find some other fabless customers with a cool idea to fill those capacity slots. And like I mentioned in the first video, _*_TSMC and Samsung prioritize market share over profitability._*_ If things cool down, they will get more aggressive."_ [12:05-12:22], Emphasis Added
Does anyone know what business rationale behind this TSMC Samsung's strategy? Is it simply to prevent other competitors from rising and taking their place.
Foundries have huge fixed costs, whether they open for business or not, whether they sell chips or not. As long as you sell something, even at a loss, you lose less. They don't fight for market shares for fame. They fight for market shares so they can keep sales. This is not just for TSMC or Samsung. Everybody needs to do it for their own survival. That's the reason when the semi-industries get into down turn, it's bloody hell. The only difference is that TSMC and Samsung have deeper pockets sustaining the loss.
Maybe it’s because the chips they make has a very high fixed cost (Billions for Fab and Millions for ASML machines) and low variable cost so they decided that better keep the machines churning out chips no matter what.
@@iamsink @John L. Thanks for the explanation. What you said makes sense.
Now do digital vs. analog.
great video