How will 'chipageddon' affect you?
bbc.com
bbc.com
High end devices have seen a massive bottleneck on TSMC and in some of the required materials like ABF, combined with very high demand (new console cycle, people buying new computers for work/study from home).
The automotive sector suffers from another problem. For decades all manufacturers have converted to JIT (lean manufacturing), which means that most of them counted on almost daily deliveries of parts to keep operating. With the transport restrictions and difficulties on many countries, this has become unmanageable and unwieldy, and factory operators have made the obvious call, increase stock margins. Typically a factory might keep stock enough for a day of operations.
A manager sees that they will have problems getting their deliveries in time, so increases their orders to cover a week. This depletes the local distributor that also tries to avoid keeping large stocks. Now a lot of other managers in other companies see that part vanish, and fearing that they’ll be out of parts, they increase their stock keeping too. With boards having from 100 to 1000 unique references, the effects propagate from one manufacturer to another, while everyone runs to get enough stock to avoid having to close.
There has been no less parts than one year ago, but the massive “de-risking” means that there’s now too much demand to keep everyone happy.
- car makers don't make their electronics. I don't know of exceptions. What they normally do is to design the architecture, the requirements and buy from Tier 1 supppliers: Bosch, Continental, Vitesco, Valeo, Delphi, Magneti Marelli etc.
- the contracts, when awarded, are multi anual and usualy specifiy the units per year during project lifetime production and in serial life (after the product is not produced, the supplier MUST GUARANTEE to be able to provide the part)
- these contracts usualy go in numbers like N million parts (ECUs) for 4-5 year total (SoP date to EoP date) + serial life for X years. Then they get detailed, even the plants where the parts are made ar negotiated. Slips on both sides come with penalty. I don't know all the details.
- lots of electronics are safety relevant (iso 26262) especially in this case, once HW design is frozen, is frozen. You cannot change silicon components easy as this will make the product undergo a long series of product and design validation which take many months => the Tier 1 supplier MUST SECURE it's own supply from chip makers.
- usual "suspects" to supply the Tier 1 are chip makers like Infineon, Renesas, NXP, STM. Probably at some point, maybe TSMC comes into the picture. There are many providers of small electronic parts.
Most automotive electronics is built around COTS (components off the shelf), which means the fab are a long way away. For instance the manufacturers you’ve mention, work with TSMC for some things, but those products are certainly not the ones for automotive. Most automotive still works on 45nm. NXP announced about 18 months that they had a design for a M7 processor on 28nm (i.MX RT1170, M7 1GHz), and for most automotive vendors, this would be considered too expensive (NXP markets it for the edge computing sector).
This is a far cry from the 7nm EUV processes that are slowing down the Ryzen 5000 rollout.
This past month I’ve had to redesign a board myself, because we were using a component from Texas Instruments targeted for the automotive sector (a DC/DC converter). Not only have all stocks vanished, some “big client” (around my area this means automotive) has already bought the whole stock (3 consecutive shipments) until Q3. And that’s from Mouser, a distributor that for a lot of these companies, is a last resort (like Digi-key, mouser provides parts and fast delivery, which of course means that you pay a premium).
Sounds a lot like the bullwhip effect, https://en.wikipedia.org/wiki/Bullwhip_effect
Nope. No one cared. They just decide to make an excuse or an enemy to blame and point fingers at.
No one question why Apple didn't have the same problem, considering they are selling 200M+ iPhone ( Likley 220M+ this year ), 20M Mac, 50M iPad, or together 300M+ of their own silicon?
No one question why Apple don't have the same problem with "their" supplier? i.e Broadcom with their WiFi, Qualcomm with their Modem, Skywork, USB4 Controller, etc etc.
Everyone thought TSMC was 7-11, or Circle K or whatever Corner Shop you have in your region. They somehow would have unlimited amount of shelf space for "you" and for everyone else.
Supply Chain and Logistic Management isn't hard. All it needs is some logic and common sense. But over the past 20 years all I have seen is that Common Sense is rather uncommon.
I find your comment very ironic, since I remember when Apple was notorious for logistical problems that led to shortages of new products. They are very strong in that respect now, and that's how they got to be a >$1T company.
That was because demand always exceed supply, and Steve Jobs refuse to give up on the JIT management system. After iPhone 4 they finally give up. Something just dont scale.
That is not to say Apple doesn't use JIT anymore. They still do it with supply guarantees.
On a tangent: Is there any good place to learn about the current state of supply chains? I searched youtube occasionally, looking for talks / interviews with people who have in depth insight in what is happening right now (/during the last year) but I could barely find anything. Shouldn‘t this be one of the hottest topics right now?!
JIT is a major trend everywhere where logistics are involved, but in car manufacturers it's the most pronounced - simply because the sheer physical size of many components required for a car would mean that car manufacturers would need big warehouses, which they have torn down or sold many decades ago.
Instead, society has picked up the cost of storage - by building ever more and more bigger highways for all the trucks.
It has more to do with Human side of things. No one wants to take the blame. Or similar to the old saying, no one gets fired for buying IBM. The concept is a lot easier to understand once you have worked in a large enterprise and witness it firsthand.
It you put too much inventory as backup. Someone will point the finger at you why are we having so much inventory. The world is at peace. We dont need that many.
Your CFO doesn't like your department? Great!. You have now given them enough ammunition to fire the bullet in your direction.
Your inventory might also depreciate in price. How do you know if your component won't drop price in three months time? Although that is also the case with hiking price. But politics is more about protecting your downside, not upside.
You want to put a contract that keep stable supply for at least a year with an accurate forecast of your sales. Which department is in-charge for that forecast? Let me tell you one little secret. There is no such thing as "forecast", but only educated guess.
You want a stable supply and guarantee? But your manufacture will also have to safe guard their Interest. You cant just forecast your buying volume to be 10M / year and only buying 6M, leaving 4M capacity that were allocated for you but now going empty. If you dont fulfil those requirement, you have to paid a penalty. And If you follow news around Apple, you might have heard Apple paying Samsung this penalty because their OLED panel did not meet the contract's minimum target.
In most circumstance, unless your Vendor absolutely hate you, your penalty will be paid back to you and deducted in your next order over large unit volume. i.e You are paying less for the same component next year. So your manufacture will be happy to keep you as a customer for a longer period of time.
None of these are specific to silicon or tech. Substitute the above with Walmart, Cargrill or any player from any industry. Supply Chain and Logistic are pretty much the same everywhere.
But I also think it's important to admit that JIT actually has real benefits. It's not "just" a finance or paper magic exercise. JIT is an approach to answering to the question of "how much inventory do I hold at each stage of distribution", where inventory held really IS money that you can't spend else where. As in, it's money you can't spend on salary, or R&D or line expansion or anything.
The risk management part is where shit goes astray.
It used to be the case that these companies would build stuff (this doesn’t only apply to cars), and charged for that stuff, and thus had an incentive to have their well guarded IP. This era is long gone.
The manufacturing business has become a race to the bottom because this IP is no longer considered profitable. There’s a mantra in MBA parlance “externalize cost centers, internalize profit centers”. For the manufacturing sector this meant keeping the sales business, while externalizing engineering. This has gone to the point that even the subcontractors have undergone that process (that’s why there’s a “tiered” model). Nowadays most of the time Engineering is 4 subcontractors down. This also means that most manufacturers are buying from the same tech providers, and are sharing most of the IP (and their costs).
The only reason this makes sense is that car companies don’t sell cars anymore, their real business is loaning, because it turns out that manufacturing big things is an amazingly good way of generating free cashflows. Look at the earnings report of any carmaker (or GE for that matter), more than half of their earnings come from finance.
The shift to modern inventory management coincided with offshoring. It became easier to source parts in Northern Mexico and now China as a result.
I'd also been keeping track of used i7-4790k CPU prices, looking to upgrade from my i5-4670k. They started out at $120-$130 earlier in 2020 and are up to $160-$170 now. For an eight-year-old CPU!
It used to be the case that computer components aged like milk, price wise, but with the death of Moore's law in the 2010s that is less and less the case. I remember RAM prices spiked a couple years back too.
But instead of splunking 140 for one i got a used workstation with a Xeon E3-1271 (the 2. fastest Xeon without GPU) for 180. It included other parts, which made the upgrade free (if you don't count the work i did) by selling my old CPU+Mobo+GPU. The GPU (RX570) i had was at the same price level on ebay as when i bought it.
If you badly wan't to upgrade (from a 4670k the delta is way smaller than from the CPU i had), i'd suggest looking for an E3-1281 or 1271 if you have an dGPU, if you need an iGPU replace the last digit with a 6
https://en.wikipedia.org/wiki/List_of_Intel_Haswell-based_Xe...
(edit: clarify, add last paragraph and links)
Little known fact and even less available, but: The best processor for the LGA1150 socket is not the i7-4790K, but the i7-5775C. Faster IPC, significantly bigger cache and a stronger iGPU, and despite not being called a K cpu it's overclockable. https://www.pc-kombo.com/us/benchmark/games/cpu/compare?ids%... shows a comparison with many gaming benchmarks.
The Xeon route can be cheaper though, it's a good alternative.
https://openbenchmarking.org/vs/Processor/Intel%20Core%20i7-...
You can't just swap it out too, you need a mainboard with a 9-Series chipset, which you usually don't if you are running a Haswell.
There are many Z97 boards out there that currently run a Haswell processor, all of them do support this broadwell alternative.
> Well, it has the faster iGPU thats for sure. But if you use it for work it is usually slower:
The pc-kombo collected benchmarks do not use the integrated graphics, that's all cpu performance with usually a high end gpu.
The 5775C not only has a better IPCs, it has 128 MByte eDRAM "L4" cache. That's just cool and helps some applications a lot. That it is slower in some benchmarks is just because of the lower turbo clock. Overclock it and it will always be faster than the 4790K - it likely won't reach the turbo clock of the 4790K, but it just needs to come close, together with the higher IPC that should be enough. Then you have the performance advantages that the gaming benchmarks show and match the performance of the 4790K in the multithreaded application benchmarks.
It's just seldom available, but sometimes when searching for used processors like the 4790K the 5775C turns up and can be had for less.
"The complexity for minimum component costs has increased at a rate of roughly a factor of two per year (see graph on next page). Certainly over the short term this rate can be expected to continue, if not to increase. Over the longer term, the rate of increase is a bit more uncertain, although there is no reason to believe it will not remain nearly constant for at least 10 years. That means by 1975, the number of components per integrated circuit for minimum cost will be 65,000."
Somehow what became known as "Moore's Law" excluded any reference to costs. For a similar distortion, see the "Turing Test".
[1]: https://newsroom.intel.com/wp-content/uploads/sites/11/2018/...
"The complexity for minimum component costs has increased at a rate of roughly a factor of two per year [...].".
When the cost part became tough the "cost side" was progressively dropped by being more and more creative on what cost is considered. First ignoring the NRE part to look only at the marginal cost per transistor. Then even this got problematic and the cost aspect was just ignored.
The "true" Moore low creates a virtuous cycle: everybody move to the next node as it's better and cheaper, so there's no point to stay back. Also, because cost gets down the addressable market gets bigger, sustaining more costly fabs. This virtuous cycle operated for a long while, so much that many takes it for granted now.
But if Moore's law die by not being lower cost (the other way is by not increasing density, and this is still progressing if slower than before), this virtuous cycle is broken. You can get higher density, but at a higher cost. With cost increasing only a subset of the market will move on to higher density. Little by little, the market size supporting the costlier high end will get smaller, while the fab cost is still increasing, contributing to a slower progress too. The virtuous cycle becomes vicious. This is happening now, although it's not too bad yet with enough very large actor able to sustain higher price (thanks Apple, I don't use you products but do benefit from all the money you're pouring into TSMC ;).
And in the end, if this gets to a serious slow down, then there is the specter of commoditization. It won't happen soon, only after a possibly very long transition period as the tech is really hard. But if progress slows a lot, the lagger may eventually catch up. With more competition would come lower margins. Which is definitely bad for stock valuation. We're still far away from this, but it's still a possibility. Look how the process leader has always been the biggest cheerleader for the "Moore low is still alive and well" (and let's not talk cost between friends, shall we?). See how it was Intel when they started selling their 10nm future, and look how more discrete they are now, and how TSMC has replaced them as the "everything is fine" voice.
[1] https://hasler.ece.gatech.edu/Published_papers/Technology_ov...
Cost per transistor is still a relevant property, but it's harder to analyze since cost numbers aren't public, verifiable figures, plus the estimates I've seen were heavily criticized by people with better information, but from what I can tell it would just offset the rate of growth, not negate it.
a 6Gb version of it would generate about $3/day on Ethereum as of today, so it would be sweet, even by cryptocoin standards, ROI of ~70 days. The rise of Ethereum prices makes even the 2-3x GPU price growth of the last couple months still a good deal. If one wonders where all the GPUs are i think that provides pretty good idea https://cdn.videocardz.com/1/2021/01/GeForce-RTX-3080-Mining...
That and it’s already hot enough in my office haha
It stops mining when you use the PC.
Granted, this was a limited edition board, but it is still 10 year old tech!
With RAM it’s a more common problem. There’s even software to check for it (memtest86)
Can you elaborate?
CPUs should go pretty much forever, provided they're run on clean power. They're a little square, entirely encased. Motherboards are bigger and with physical connectors you need to use force on.
https://semiengineering.com/chip-aging-becomes-design-proble...
There's electromigration which is harder and harder to mitigate with smaller process nodes
[1] https://www.newegg.com/p/pl?N=100007671%20601351803%20601321...
Looking on eBay, it looks like I could sell that $290 card for about $500...
Intel is releasing Rocket Lake shortly but it won't have any options with more cores than the 5800X. Zen 4 isn't launching until late fall. Not sure availability will be any better on either of these launches anyways.
GPUs are a little bit more ridiculous. You can get them if you don't need it to show up tomorrow but getting one at a reasonable price any time soon is much more a challenge than doing the same for a CPU. That being said I did manage to get a reasonably priced 3070 in a build early on but I put that to luck more than anything.
Motherboards I haven't had any issues with.
I broke down waiting for a 5950X and bought one off a scalper last week :/ luckily got it before the recent spike in price.
Sour milk is not worth very much but it then becomes yogurt and cheese which are much more expensive. So I guess your analogy still holds? ;)
I paid €128 for it.
That's not going to get you anything today. 3 years ago I paid €350 for a GTX 1050 TI. Now, everything new starts at €500, and everything older is a waste of money, says everybody.
Still not dying, and it was made in enormous quantities because nVidia made a big bet on crypto craze.
This is not a good way to measure things. Did you just exchange 10 years of monthly payments for 30 years of "similar" monthly payments?
My mother protested to me over the cost of her new house, defending the idea that she could afford it by pointing out that monthly payments were similar -- property taxes on the new house were about the same size as mortgage payments on the old one. I was kind of horrified by this; conceptually, when you make a mortgage payment, your net worth goes up. When you make a tax payment, it goes down.
Nonsense. It runs all games on High/60fps or better, with the possible exception of Cyberpunk and RTX marketing gimmicks. That's why it's still a relevant card 4.5 years later and goes for $200+. I own one.
> I'd also been keeping track of used i7-4790k CPU prices, looking to upgrade from my i5-4670k. They started out at $120-$130 earlier in 2020 and are up to $160-$170 now. For an eight-year-old CPU!
Funny, since I own one as well (4670k). And I haven't ever felt CPU-limited. Mine is mildly overclocked to 4GHz, and it's been able to crunch everything I threw at it since I bought it in 2014, again with the possible exception of HEVC encoding. It's still a very relevant CPU and handily beats Ryzen 1st gen for example.
It's also the last generation of Intel CPUs that lets you run a non-spyware version of Windows.
I'm sorry you fell for the marketing.
Now for the current iteration of RT graphics cards, it is still very limited so calling it a gimmick is a bit more warranted. But the hardware will improve (and has already from the first RT generation) which will allow more games and engines built arround RT instead of having it tacked on as an option.
Ray traced reflections are realistic reflections and look outstanding in Control.
Because they're based on the real level geometry.
Calling it a gimmick just means that you personally don't care about graphics.
I've got a £450 PS5. I'm playing Control right now and I choose the 30fps ray tracing mode because it looks much better.
That's probably sacrilegious to you.
Can you elaborate on this please?
Placing Planned Purchase Orders for your current products helps a lot to weather supply problems.
Also tsmc cancelled euv equipment orders due to the huawei ban:
https://semiwiki.com/forum/index.php?threads/asml-reports-20...
Second that linked forum thread is really unclear. Some even thinking that Intel was the customer who delayed EUV machine orders (which would make sense given their 7nm problems).
200mm is what most of automotive chips are made on.
Cheapest 300mm equipment is many times more expensive than 200mm.
> The shortage is amplified by the us china tech war
I keep hearing this vague accusation, over and over, with no evidence provided.
Taiwan is not part of China -- at least as far as the sanctions are concerned.
China exports lots of finished electronics but very, very few semiconductors -- and virtually none for the automotive sector. SMIC is not competitive. They're kept afloat by the Chinese government and supply almost exclusively for domestic products.
> that forced key supplier asml to limit deliveries of chip making equipment.
Source?
> Also tsmc cancelled euv equipment orders due to the huawei ban:
How would that make any sense? TSMC is completely booked to the hilt right now. Would any sane person think "gee, that's a great cancel our expansion plans"
That $150 million ASML EUV scanner order that got cancelled was from SMIC, not TSMC:
https://www.electronicsweekly.com/news/business/asml-delays-...
Again, China not Taiwan.
> https://semiwiki.com/forum/index.php?threads/asml-reports-20...
I'm not going to read the whole thing, but the word "Huawei" appears only once on that webpage, as part of an anonymous blog commenter's assertion. That is not a credible source.
The forum thread links to the conference call transcript in the very top.
Smic got their euv scanner blocked. Tsmc cancelled an order. Both due to the tech sanctions.
https://www.fool.com/earnings/call-transcripts/2021/01/20/as...
> https://www.fool.com/earnings/call-transcripts/2021/01/20/as...
The word Huawei doesn't appear on that page either.
Please stop spreading misinformation.
The sanctions are on Chinese companies, not Taiwanese companies. China's chip exports are somewhere between zero and insignificant.
"... as clearly our key foundry customer came back and said, listen, our key customer for N3 is now blacklisted. So we cannot ship. So we need to adjust our 2021 outlook for EUV systems."
Could sell more had it not been for the US sanctions:
"... but clearly see potential upside to these numbers where we can disregard any further impact of export control regulations resulting from the current geopolitical situation."
The CEO is business person and for him to do any kind of politics or worse geopolitics would hurt his business. So he clearly cannot just say: "Listen, let us sell a bunch of these bad boys to the Chinese and the chip shortage will be a thing of the past in 6 months." Even though that just might be the case.
Also as far as anything else is concerned (except spineless organizations like the UN).
"it's supply chain limitations by design, because the customers told us, we don't need them, and then coming back and said, oops! we might have been wrong."
From the CEO it's due customers thinking it was going to be bad (for various reasons) and then it wasn't. The process have long lead times and when you pause them you can't just expidite them, they have a fixed lead time.
I hope people will learn from this.
As far as I see it, we need more capacity, and for a situation like this it wouldn't matter if that extra capacity was in 50 different countries or if it was all in a 50 km radius in Taiwan.
Edit: just realized that also supply of these ICs stalled. Right now, it's hard to see whether this supply shortages are just due to temporarily increased demand or not. If the latter is the case, it is still easier and cheaper to set up fabs for older nodes.
The current technical situation does make peaceful cross-strait relations critical to global technology, however. A hot war between China and Taiwan would send the silicon state-of-the-art backwards by quite a bit, as would any trade embargo post-annexation if China somehow accomplished that peacefully.
Taiwan's silicon fab prowess is a pretty potent strategic weapon against a much larger adversary!
This line of thinking gets very impractical very quickly. It's also a recipe for war because interdependence makes countries play well together.
Countries don't go to war because they are independent. Countries do have incentive to maintain peace when they are dependent.
Two very different things.
This line of thinking led directly to WWI.
Also, mass sanctions last year doesn't help.
Are people living in smaller countries able to afford redundant fabs?
Another issue more pertinent to automotive manufacturing is certification. For example, car makers once had to wait almost a year for Renesas to rebuild their facilities in Naka following the 2011 earthquake despite the company operating fabs in multiple countries. The latter were perfectly capable of producing the product in demand but the cost of certification is so prohibitive that moving production elsewhere was never brought up as an option.
Profit oriented companies take those cost savings and choose to drive further down the technology stack (to more complex and expensive nodes), where there are fewer competitors that have enough profits to build competing production lines.
Moore's law is almost our friend here. Being at 3nm instead of 16nm is decreasingly important.
Is anyone doing ~16nm microprocessors besides Samsung/TSMC/Intel/GloFo? I wouldn't call it "cheap and easy" if only they can do it.
(After a quick Google, it seems that China's state-owned SMIC is also starting to do ~16nm, but that's all I could find, unless you mean memory and not processors).
Combination of mining craze and massive demand for entertainment electronics and GPUs since everybody has been stuck at home due to pandemic and decided to build/upgrade their computers.
Which also lead to the situation that scalping outfits have extended their services from covering mostly designer appeal to now also include electronics like GPUs and consoles. The popularity of this also attracts a lot more people to use these kinds of services as a form of investment, which most certainly does not help an already supply constrained market.
I think a lot of people are getting into gaming again and a lot are looking for cards that can play their game just well enough.
Due to this, people needing replacements and any in-flight new orders apparently have wait times of up to 9-10 weeks...
Not helping this is that some people try to sell them in the secondary market despite them being premises (or rather segment) locked.
[1] https://www.nbnco.com.au/utility/temporarily-limiting-the-nu...
Come take my money
There are outfits out there that charge a monthly subscription for access to prioritized acquisition channels in combination with bot networks.
This is streamlined to such degree that people don't even need to hold physical inventory themselves: The scalping outlet handles all of that for them, all they need to do is supply the capital for the bots and set the parameters for buying/selling, everything else will be handled by the service.
The situation is far simpler. Consumer demand exceeds supply which drives prices up. Some people figured out where the current equilibrium price is and it turns out that it is really high.
Those high prices encourage consumers with old GPUs to sell them on the market and thus provide more supply instead of throwing their old GPUs away every time they buy a new GPU. If you were to artificially restrict the price it would result in less people selling their GPUs and thus you would end up with less GPUs for everyone.
If the manufacturer is selling a product that is worth X at market rate for a lower price Y then it basically gifted the difference X-Y to the buyer. If the manufacturer is selling the product at market rate prices they will have higher profit margins which encourages them to build a slightly bigger fab next time. It would also ensure that existing manufacturers make enough money to continue running their business despite the paper thin margins outside of a shortage.
I feel like we need to find peaceful ways to resolve our severe differences with China. Either that or stop depending on China's neighbors for critical technology.
I just hope that there can be some kind of progress with the new administration in Washington.
However, the last software update was about 4 years ago.
The tablet has storage expansion, but due to artificial limitations in the Android system, only a fraction of software will utilize the sd card.
These are only minor examples.
I could rant for hours on the ”who cares, it’ll be garbage in a year attitude” that has invaded much of the tech world, but others have done a better job. I my self am some times party to this attitude.
It all results in a lot of waste of all kinds and also bad user experience.
To look at the silver lining, maybe this scarcity and some other trends will result in a small corrective trends in engineering and management.
In the meanwhile I’ll keep dreaming of a new gaming PC for a year or two.
I was finally looking at replacing the 8-ish-year old parts of my gaming PC which has otherwise been running like a champ (Intel 2500K), and now... Well, I guess I'll just focus a lot on pixel art games :P
Life is much calmer in the slow lane.
Particularly the factory-sealed liquid-loop, which has both moving parts in a pump and gradual liquid chemical degradation to consider. On the flip-side, that particular component seems like it would be less sensitive to silicon-fabbing issues.
If anything, such companies show there is demand and money to be made, there.
There is nothing wrong with a solid thinkpad if you can replace the battery so that it still works well, but my days of jumping from outlet to outlet because I got 20min on a full charge are over. So are the days of carrying a heavy laptop with a bulky charger.
I brought a oneplus phone some years ago and I paid too much for it because I wanted a "great phone". I don't regret the ultra-fast charging and day+ battery though, and my next phone is probably also going to be a oneplus, just a much cheaper model.
So if there is something to sum up this rambeling comment, it should probably be this: know what you are missing in your current tech, and upgrade as it makes sense to do so for your personal use case.
There's a lot of noise about the shortage of high end parts from TSMC, but I'm much more concerned about shortages that can impact more common consumer goods. Having issues like this when the economy is trying to recover from a pandemic could have far reaching consequences now that so many products contain semiconductors.
Either due to maxed-out fabs or shifts in uC/uP usage, the shortages are wide-spread.
In many categories, like cheap laptops, webcams, small printers, etc demand exploded. More Chromebook orders came in 2020 Q3 than normally happen in a year. Laptop LCDs are super-constrained as well. The dilemma is that expanding manufacturing capacity now will reduce margins tomorrow, especially since the order books will be empty as COVID fades as schools open again.
My guess is you’ll see prices in free fall as the edu backlog clears and COVID restrictions fade. Probably this time next year.
Sold my old system for around what I bought it for, so if you want to game these times are whack.
BUT... those are just for enthusiast builders. You can find OEM systems with those chips.
So I guess the only lesson is that being a computer manufacturer will give you priority on high-end chips, and in times of short supply, that's where they'll go first, drying up the supply for small buyers.
Big suppliers of ABF substrate includes multiple companies from Taiwan Unimicron, Kinsus Interconnect, and Nan Ya PCB, which all trade in the Taiwan stock market.
I considered investing in those companies...but seems like only Nan Ya has been showing an upward trend. Perhaps we are still early?
Presumably this will be good for the foundries and makers of industrial semiconductor machinery, and bad for fabless companies. Hard to know the time horizon, however.
My 2¢
Though I have to say, for all its faults, CP2077 really blew me away graphically (using the RTX support) like no other game did so far. Also, VR would also really benefit from better GPUs. And 4K gaming is also not quite there yet. So actually, I maybe wouldn't even say that. Performance improvements are still welcome.
The problem is more that there is an unprecedented demand for compute. ML, Crypto mining, Gaming being more popular than ever with Corona, etc. just means that demand is higher than ever. Combine that with the new console launch and exhausted semiconductor manufacturing capabilities and you can easily see why customers are biting the bullet and going for less pricy, older cards.
DLSS is a cute hack. We have better scalers than that in video market, its main advantage is speed. Instead, fixed function scaler should be better than outdated 5-tap filter. Remind me again when it matches basic scalers put in Smart TVs.
The gaming GPUs are hilariously underpowered for compute, even something like 3090, vs MI100 or A100. It's just the latter cost $10k per unit.
https://www.eurogamer.net/articles/digitalfoundry-2020-contr...
What TV upscaling does that?
Also ray tracing produces significantly more realistic lighting. It's not comparable to traditional lighting methods in accuracy.
All the work is going into memory and caching. Fine tuning the arch. Ray tracing and such have also changed the latency tightrope you balance a bit.
For an end user, the difference in visual fidelity will probably be minor. As an GPU researcher myself, this is an extremely exciting period.