> Slashing orders to a contract manufacturer is not trivial since fabless chip designers are obliged to procure a fixed number of wafers in certain quarters. Nevertheless, TSMC is reportedly willing to accept compensation (as it will hold wafers with chips from AMD, Intel, Nvidia, etc., before they are ready to buy them) and even renegotiate deals on long-term supply contracts (i.e., increase the number of wafers that a company is committed to buying in the future) in exchange.
Sounds like it might not be such a big deal for TSMC, at least in the short term.
They are forecasting a lower demand in the short term future. I’d guess they ran the numbers and producing fewer high cost items is more cost effective for them.
Unfortunately, this likely means more shortages and costs stay high for the consumer.
The fact that the market can still do this (reduce supply to operate in the high margin quadrant) is a temporary blip due to the x86 duopoly. Soon anyone will be able to supply cycles/$ at a better rate than the entrenched players.
It would take a lot of time and money to create anything competetive with Nvidia and AMD's offerings. It took Intel over a decade and their dGPUs still aren't very good.
GPUs are insanely complex and the drivers are full of hacks (even per game basis, a driver could swap shaders with ones written by Nvidia/AMD if it detects you're playing the newest popular title). The threshold to enter the market is incredibly high.
This is far from the truth; oftentimes the joke with GPUs is that you're buying a hardware driver with a device to unlock it.
Intel is hitting that problem head-on; no game is written to take advantage of its new GPUs, and has to suffer a ton for it. It's not a guarantee Intel will ever be able to escape this problem.
From my limited outsider perspective, it appears like there are way more CPU and FPGA startups than GPU startups - why is that when GPUs seemingly have lower barriers?
Most of the related startups are targeting the much higher-margin deep learning training accelerator portion of the GPU system market, and there are a lot there:
It's also a quite explored and already deeply optimized area.
Companies like TensTorrent, Graphcore and Lightmatter (for example) try to hit different spots of the architectural landscape (seemingly all optimising for 'cheap to tapeout' in all sorts of clever ways, but you still have to code for it (even though you often get python APIs and deep learning tensorflow/torch support, but that's probably not getting you to to max perf). Very interesting to watch, and hopefully one can get their hands on that kind of hardware and build a community around it.
This is all pretty accurate. The only real surprise is that 40um tech and above are still running full steam and even the 20+ year old tech has only dropped a few percent for next year.
This is entirely unsurprising if you actually look at things other than big CPU/GPU/IO.
Power regulators doesn't need small transistors, it wants big nice and fat transistors. You'd get at least one of them in every device and a bunch in any bigger device. Same with any power device, LEDs, and really anything that passes some current.
Any microcontroller smaller than "runs full fledged linux" often cares about cost of production first and foremost, hell, STM32 only recently [1] got on 90nm node!
Interestingly enough, bigger node might get you worse uA/MHz but often lower idle current (the bigger things are the lower the leakage) so they might even be desirable. And the chip might want to have some mA current output on each pin (so it doesn't require external drivers to control stuff), which again means some transistor size in your I/O port is fixed and can't be made smaller so your savings from going to smaller process are not linear.
For vast majority of non CPU/GPU chips it comes to cost to manufacture and those 20+ years old lines not only paid for themselves multiple times, they also have great yields.
> Power regulators doesn't need small transistors, it wants big nice and fat transistors.
Is it possible to make bigger transistors using a smaller process? Just because you are using saying a 7nm process, can't you still make a transistor (like if you needed) that was the same size as it was on 90nm? I think of newer processes as merely being more precise so you could still do the old designs on them, you are just unecessarily tying up the newer machines... but concievably in the future you could just start using 5nm process to keep producing 30 nm and 90 nm chips just to consolidate production lines, non?
Only to a point. Certainly not at 5nm. That's too advanced and power-inefficient.
5nm is extremely inefficient to make: IIRC, you need to submerge the parts in ultra-pure water and blast it with a ultraviolet laser to get down to 5nm. This is because even ultra-pure filtered air has too many particles that disperses the laser that you're off by a few nanometers... ruining the design. Filtering all that water and powering ultraviolet-spectrum lasers is extremely high power / electricity usage.
In contrast, a 40nm process is still airborne. You can largely do 40nm with "more standard" equipment. No ultra-pure water needed. No ultraviolet lasers. No quad-patterning. You can use "regular" light, with "regular cleanroom air", and "regular" processes to make the design at far, far lower costs (or at least, with less electricity).
---------
IIRC, the "ideal" cost-efficiency moving forward is to rebuild our older fabs for 300mm wafers, standardize upon that wafer size. Older fabs were standardized on smaller 200mm wafers with only ~45% area of the bigger wafers (and less in practice, because the edge of a wafer has all sorts of inefficient issues involved).
There's also the issue that a lot of these designs, such as Power MOSFETs, are... just one, maybe two, transistors. If your transistor is like a centimeter in size, there's no point using 5nm or 40nm or even 90nm nodes on it. There's just much much cheaper ways to fabricate a "centimeter-sized" MOSFET. There's all sorts of application of "chips", from 1-transistor Power MOSFETs to ~100 transistor op-amps, to ~10,000 transistor transceivers, to 10-million+ transistor NAND-Flash arrays. (1 Megabyte isn't much these days though, lol)
It's hard to convey just how ridiculously complex and expensive 5nm is compared with 90nm.
5nm is a multi-billion dollar fab, absurdly high running costs, and very expensive wafers.
As a data point: 5nm lithography equipment is > $100M/unit, needs megawatts of power to run, and they handle about 60-80 wafers per hour.
90nm lithography equipment is ~ $1M/unit, needs kilowatts of power to run, and they handle about 60-80 wafers per hour.
So a wafer scanned on 5nm equipment that could have been scanned on 9nm equipment is carrying a very large capital and operational cost burden that's pointless.
Yeah, but that's no advantage if the device you're making can't be made any smaller due to non-manufacturing constraints, which is the case for power electronics, a lot of analog electronics, and also a suprising amount of digital electronics as well (where the whole device will fit in the spaces between the bond pads needed to connect the chip to anything else).
Yes, some power transistors place many smaller transistors or mosfets effectively in parallel. An early example AFAIK was the "hexfet" geometry (achieving a very wide width of short-length channels). I'd guess that precisely the same process would not be optimal.
Just to add color to the discussion above, what distinguishes a microcontroller from other chips is that the volatile memory is on the same piece of silicon as the processing cores.
This allows for power saving because you don’t have to condition signals as they jump from silicon piece to silicon piece (oversimplification).
It is somewhat harder to make the above on smaller nodes than it is to make just processing cores or just memory.
That fact (in addition to a number of other cost and technical factors) is why microcontrollers lag process nodes.
> It is somewhat harder to make the above on smaller nodes than it is to make just processing cores or just memory.
I'm not convinced this is the reason because the microcontrollers I know of use SRAM not DRAM and SRAM is what CPU caches are (plus of course some extra logic for the cache management aspect). I am aware that designing dense and power efficient SRAM for a cell library is considered a pretty hard thing to do, but it seemingly has already been for modern processes.
3D NAND, which basically underpins all of our SSD tech, is hard enough to make at 40nm. They basically have to fill a tiny hole really precisely with layers of five different materials to create a vertical stack that works. It's really hard to go down node sizes plus once you hit 28nm the price of transistors stops decreasing and starts increasing instead with 22nm and smaller.
This idea that the price per transistor gets higher after 28nm needs to die. Right now the cheapest node for logic dominated chips is N6 (an optical shrink of N7). And even then, N5 is still cheaper than N10.
None of that is really relevant, because TSMC isn't in the business of making flash memory. The fab processes for logic and for flash memory are so different that there's no point to making comparisons between their respective nodes.
I don't know about "no point to making comparisons", but they are definitely different animals.
When I worked at Samsung, flash & LSI were like 2 different companies. Lots of reuse at the systems engineering level though. At a certain point of abstraction you can't tell what is on the wafer anymore so you might as well track it in the same database.
Completely different animals. I know that when TSMC has looked at acquiring used fabs before and in at least one case because the fab was DRAM it was more costly and time consuming to convert it to Logic fab than just building a whole new fab.
> E.g. what is preventing a RISC-V company from leap-frogging the competition?
They would need to have already taped out on N7 and be ready to start mass production now. That doesn't happen overnight. Even if you have taped out a ready-to-fab design, you usually want to start small and run a few rounds of QA on small batches of chips before you start ordering enough wafers to make a dent in TSMC's schedule.
N7 is a rough node to target. You have the high capital costs for NRE being an EUV node, but N5 is supposed to be cheaper to manufacturer chips on pretty much as soon as N3 is in full swing.
So those companies would probably be targeting something in the range of 14nm to 10nm in order to take advantage of the fire sale on late DUV tooling to prove that they can make a working chip to investors, and then leapfrog to N5 after the giants have jumped to N3.
Is this because of the economic slowdown as the article suggests, or simply because 2020-2021 covid lockdowns lead to a big surge in buying of computer hardware, and therefore a big drop now given that modern computers last a lot longer and everyone is equipped?
> 2020-2021 covid lockdowns lead to a big surge in buying of computer hardware, and therefore a big drop now given that modern computers last a lot longer and everyone is equipped
The problem is that the higher end GPUs are actually the same chip as the lower end GPUs. The lower end GPUs are where the manufacturing process had an error and thus they fuse off a core. All that to say that they cant manufacture a high end GPU without also manufacturing a low end one (or multiple depending on the manufacturing error rate). So, if they cant sell the low end one, it may not make sense to sell the high end one even if there is a buyer.
>The problem is that the higher end GPUs are actually the same chip as the lower end GPUs.
Don't know where you got that info but that's definitely false.
If you look at Nvidia's lineup, each GPU has a die that's unique to that product in both size and markings. Sure, there have been/are a few products that make use of the same die but with defects depending on yelds, but those are usually in the same performance class (1070, 1070TI and 1080 had the same die), and not the case you're talking about of entry level GPUs being high end GPUs dies with defects fused off. That's never the case. An RTX 3050 die is a completely different part than a 3080/3090 die.
I think it applies for the datacenter/gaming distinction, where the chip of the A40 (ga102) is supposedly the same as the RTX3090 (Ti?) and they play on frequency, memory sizes and other differentiators? And on the license of course...
Yes, though also only to a point. You still see multiple dies in a product generation for CPUs, but you do get binning into multiple SKUs for each of those dies.
Maybe this isn't strictly about GPUs. Or maybe it's just FUD designed to signal that no, video cards aren't getting cheaper and they'd rather cut supply than lower their margins. Who knows with this industry.
If I were Intel, I would see this as an opportunity to break into the GPU market by releasing their GPU's at very competitive prices and ensuring enough supply.
Arc GPUs already sell at a significant discount and are competitive on performance. The problem is that the drivers aren't mature enough and you can't count on that performance. It's way too inconsistent to pay that much for.
Their drivers mean their performance is an order of magnitude worse than nvidia or even amd. I would not buy an intel gpu until it can give at least 6700 xt perf with rock solid stability for less money.
It's not always that bad, the main issue is it's inconsistent. Some games run great, others run slow, some are very buggy, and some won't run at all. It's gonna take a while to catch up.
This is my hope for the Arch GPUs as I'm in the market for a cheaper--if less powerful card. Sadly their performance with OpenGL isn't great, apparently because of an emulation layer.
Despite popular belief, most computers are used for business and unless there's an entire multi industry collapse, the potential demand destruction probably won't be that high outside of the prosumer PC gaming markets.
I do think crypto is demand destroyed due to their own collapse unrelated but related to the economy in general. My expectation is that the next crypto bubble will take longer to inflate this time vs the last, but hell what's the point reading tea leaves. This could be the end of crypto. All we need is a new and exciting meme way for early adopters to sell losing bets into ignorant followers.
It’s not just about the 4080 though (though, you are right). AMDs new cards are much better value for gamers, but they are in this list too.
I think this is less about consumer GPUs and more about cost cutting in big corps slowing expansions, and the reduction in crypto demand.
Think about every tech company that says “oops we overbuilt for covid, whelp let’s lay everyone off” well they could be buying fewer servers if they over-anticipated their needs in that department too.
TPUs have been all over the shelf, and adoption was snubbed. More graphics cards than ever before! Real-time raytracing available on your cell phone, laptop, and all 5 smart TVs in your home and backyard porch!!
You were unable to convince your executives to adopt huggingface, IOT, and instead they opted to employ phone calls and Excel spreadsheets.
You, consumer, failed to accept Stadia into your heart,... and now you shall pay for your snide indifference!!
Colab and huggingface are fishing for customers by handing out compute time in these things for free, and getting a militia's worth of tensor-connected memory isn't too expensive if you get provisioned to spin up a p2 on EC2, or presumably azure or gcp.
Of course real consumers can't buy the hardware, real consumers don't have AAA credit ratings.
You can always eBay a K80 as well if your motherboard supports the >4G decoding switch, but my primary point was about the insane purchasing power that affords large CommSvc sector company the ability to crash the market for a product before anyone even has logistics in place, apparently wishing to gatekeep any applications.
- Game devs haven't made any games that stress a cloud gaming environment, they didn't even push VR.
- Corporate consumers are hardly adopting AI at all, because there's still barely even a workforce.
- Some very vocal people find proof of work a distasteful mechanism to support logistical metacommodity growth.
Does a 4xxx card make sense for any pedestrian gamer? Clocking in at >$1200, it is a hard swallow to say that the extra FPS is worth it. Unless the price comes down massively, the high sticker price (plus insane power draw), really leave me wondering who is purchasing these things for anything but machine learning.
They're not worth it at all. I have a 3080 for 4k/60fps gaming without issues and I got it for $500 used. If a game struggles to hold a consistent framerate it's usually an issue with the game or certain settings, tweaking settings is free so throwing hardware at it doesn't make sense.
Its worse, TSMC is the single biggest semiconductor manufacturer for general logic chips. Samsung has their own fabs for ram and ssd's specialised for density. All the litography equipment they use come from a single company called ASML. They also have complete market domination for the past few generations of litography machines and are the first and only producers of EUV litography.
Semiconductors are so complex and capital intensive that even a duopoly couln't balance the expense.
Sounds like it might not be such a big deal for TSMC, at least in the short term.