News TSMC Expected to Charge $25,000 per 2nm Wafer

[Admin]

As process technologies get more sophisticated, TSMC charges more for wafers.

TSMC Expected to Charge $25,000 per 2nm Wafer : Read more

 

[hannibal]

So more expensive GPUs and CPUs in the future. Quite expected…
Interesting. About douple the price of 5nm at 2022.
Maybe 20% in total cost of GPUs, but almost full towards the CPUs.

 

[InvalidError]

With ribbon GAA FETs, the transistor layer counts will be going up by 3+X, 6+X for designs that may start using stacked CMOS, sky is the limit if companies start embedding multi-layer SRAM and DRAM.

 

[Deleted member 2731765]

This is actually OLD news. 'SeekingAlpha' predicted and made that table and chart WAY back in Jul. 14, 2022. So nothing new here. "Back to the Future".

 

[usertests]

It could be bad, but I think you're still only looking at $100 max of TSMC N2 silicon for a 7950X style chip with 2 chiplets (I just assumed 500 chiplets from 1 wafer). Maybe the chiplet size will grow to accommodate more cores, and maybe most of the L3 can be made on a different node and stacked, shrinking the 2nm chiplets.

 

[JamesJones44]

When your the only player in town for "leading edge" silicon you can charge whatever you want. I keep hoping Samsung, Intel or someone else will at least start to challenge TSMC. Otherwise prices are only going to go up at an increasing rate if they can't.

 

[evdjj3j]

If you can get more than 25% more dice per wafer then wouldn't this make the price per die cheaper than the old node?

 

[ajs electronics]

So more expensive GPUs and CPUs in the future. Quite expected…
Interesting. About douple the price of 5nm at 2022.

BTW a 300 mm wafer gives 500cm2 of chips for 5 dollars per cm2. the die on an apple chip is 1.2 cm2. What's the maths on the transistor cost? 5 squared is 25 and 2 squared is 4, so you get i.e. 3-4 times more transistors for the same surface area, at twice the price.

 

[PlaneInTheSky]

AMD and Apple are getting royally screwed by TSMC.

TMSC and Taiwan know they can ask whatever they want since they have no competition on leading edge nodes.

 

[Amdlova]

still using my 14+++++ node (10nm) aka 7nm from intel. With some more nodes will be at perfection...
500W or more Single die.

 

[helper800]

Cannot wait for the 500 dollar 8700x and 650 dollar 1X700k... If Nvidia and AMD also use TSMC for their GPUs we are probably looking at 500+ dollar X060 class cards from Nvidia and similar for AMD X700x class cards...

 

[watzupken]

There will always be a tipping point in anything that you keep stressing. And as you can tell, demand of gadgets are not on the rise, but on a fast decline. What used to be an annual or bi-annual device upgrade is becoming 3 years or more for some consumers due to lack of any meaningful improvements and increasing cost. There will still be people with deep pockets that can continue paying for it, but most people don’t have deep pockets, but a load of debts instead.

 

[umeng2002_2]

Alright, I'm done with computers. I'll think I'll take up surfing or something...

 

[usertests]

Alright, I'm done with computers. I'll think I'll take up surfing or something...

Just buy discounted systems or parts. Every CPU from the latest and previous gens is discounted, 5-10 year old CPUs are often perfectly fine, and other components such as DRAM and SSDs have fallen in price precipitously. GPU pricing is still elevated but getting good 1080p performance is fairly cheap, e.g. as low as 180 USD for RX 6600 right now and Phoenix APUs probably coming to AM5 soon.

If the hobby gets really bad again, it will be because of another catastrophic global event, not an 80% rise in bleeding edge wafer costs.

 

[InvalidError]

If the hobby gets really bad again, it will be because of another catastrophic global event, not an 80% rise in bleeding edge wafer costs.

Now that we have practically reached the end of the road for conventional silicon short of stacking transistors, performance is about to become a direct function of how much money you are willing to spend on a CPU/GPU chip package. The only performance per dollar improvements we'll see will be coming directly out of manufacturers' profit margins.

 

[neojack]

well anyways, in just a few generations, we will hit a wall : it will not be physically possible to further reduce the size of the electric paths. I have read somewhere that right now the electric circuits inside a modern CPU are only distanced by 20 atoms from each other. there is not much more to trim before to encounter quantic effects between the paths.

This means that every generation will cost more and more, for fewer gains. It also means that we could keep our hardware for a very long time. (maybe a 4090 will still be relevant in 10 years. who knows. a 10 years old 4790k is still good nowadays depending of the usage)

 

[ajs electronics]

if the process width is 50% smaller, then the number of transistors per wafer is 400% higher, and the price is going up by 200%.

 

[usertests]

Now that we have practically reached the end of the road for conventional silicon short of stacking transistors, performance is about to become a direct function of how much money you are willing to spend on a CPU/GPU chip package. The only performance per dollar improvements we'll see will be coming directly out of manufacturers' profit margins.

We should give monolithic 3D processors built using chemical vapor deposition, nanotubes, etc. a decade to materialize before we go full doom and gloom.

Meanwhile, GAAFET/CFET nodes are going to be dropped on the market through the mid-2030s, according to IMEC at least. Even if wafer cost increases canceled out the density increases, performance and efficiency could go up and improve perf/$ slightly.

 

[TerryLaze]

if the process width is 50% smaller, then the number of transistors per wafer is 400% higher, and the price is going up by 200%.

No, both the width and the hight (x and y) of the process needs to be 50% smaller for the number to increase by 400% that's not how process nm numbers work, they often don't correspond to any physical size at all and are just an representation of how much better the node is.
It's a maximum of 70% better density for 100% more money, TSMC is still in it for the money, selling transistors at half the price would not make them enough money.

Directly from TSMC:

N3 technology will offer up to 70% logic density gain, up to 15% speed improvement at the same power and up to 30% power reduction at the same speed as compared with N5 technology.

 

[TerryLaze]

well anyways, in just a few generations, we will hit a wall : it will not be physically possible to further reduce the size of the electric paths. I have read somewhere that right now the electric circuits inside a modern CPU are only distanced by 20 atoms from each other. there is not much more to trim before to encounter quantic effects between the paths.

It's not about making things smaller, that is not what makes a CPU faster, if you need more cores or wider cores on a CPU you can do that no matter the size, well...within reason. It makes your dies larger but that is not the problem, we have all seen the pics of humongous server CPUs.

The problem with faster CPUs is making them run/switch faster which is the Ghz of the CPU and that has very little to do with the process size.

 

[kyzarvs]

So more expensive GPUs and CPUs in the future. Quite expected…
Interesting. About douple the price of 5nm at 2022.
Maybe 20% in total cost of GPUs, but almost full towards the CPUs.

But the same chip would take up much less space at 2nm compared to 5nm, so you'd get more yield per wafer?

 

[adam25185]

This article misses the point completely.

N5 price is $13500/300mm
N2 price is $24000/300mm

(2/5)^2 = 0.16

0.16*$24000 = $3840

Prices are going way down per area.

 

[evdjj3j]

This article misses the point completely.

N5 price is $13500/300mm
N2 price is $24000/300mm

(2/5)^2 = 0.16

0.16*$24000 = $3840

Prices are going way down per area.

It's clickbait, this site is getting bad about it. I was thinking last night about how much the quality of content has gone down since the early 2000s.

 

[InvalidError]

This article misses the point completely.

N5 price is $13500/300mm
N2 price is $24000/300mm

(2/5)^2 = 0.16

0.16*$24000 = $3840

Prices are going way down per area.

That math only works if process node names scaled linearly with actual feature sizes. The reality is that some dimensions are coming down much slower. For example, "minimum metal pitch" has only come down from 64nm to 24nm between TSMC's 16nm and 3nm nodes. The node name suggests a 5X increase in dimensional density while that part of the actual process only scales half as much.

Node names are just that, names. Any sort of mathematical relationship they may once have had with physical feature sizes has been steadily loosening since the 90nm days if not longer.

Edit: and in the new Samsung fab update, going from SF3 to SF2 only yields 5% higher density, nowhere near the 125% intuitive areal increase based on process name numbers. That is before we even consider the analog, power and other bits that don't scale much if at all.

 

[laxman10100]

I'm almost positive that TSMC had announced their differing process prices around a year ago... As a result, I've been telling people for 6 months now how the upcoming RTX 5090 FE will cost between 2000-2400 USD. These already announced TSMC production prices along with the 5090's leaked PCB & chip layout have proved my point, haha.