TSMC reportedly plans to raise 3nm wafer prices 25 percent over 5nm to $20,000

[Daniel Sims]

Why it matters: Prices of computer components and other consumer electronics have recently either risen or stagnated. Manufacturers blame component costs, inflation, supply chain shocks, and other macroeconomic conditions. Reports indicate TSMC's latest node process won't alleviate the problem.

According to DigiTimes, TSMC intends to sell its 3nm N3 semiconductors for $20,000 per wafer. Contract prices for larger companies like Apple, Intel, or Nvidia will likely be lower, but these increased costs will probably filter down to consumers regardless.

DigiTimes has charted the history of wafer prices, starting with TSMC's 90nm process in 2004, which cost a tenth of the N3 wafers at $2,000. Estimates indicate N3 wafer costs will rise 25 percent over TSMC's 5nm N5 process, which was around $16,000 per wafer. DigiTimes didn't specify whether it accounted for inflation.

The 3nm node is seeing a price hike primarily because the tools to make it are still expensive. The scanners TSMC uses for extreme ultraviolet (EUV) lithography cost around $150 million each, and fabs require multiple scanners. The N5 process requires 14 lithography layers, while N3 needs even more.

Samsung's yield problems with its 4nm and 3nm processes also leave TSMC uncontested in the bleeding-edge semiconductor space. Qualcomm initially partnered with Samsung for its new chips but had to rely on TSMC for its Snapdragon 8 Gen 2 mobile platform.

While semiconductors are far from the only element affecting component and electronics prices, they will influence the price tags consumer see. After revealing the shockingly expensive RTX 4000 series graphics cards in September, Nvidia boss Jensen Huang declared (again) that Moore's Law is dead. The Jury is still out on that claim, but TSMC's price hike may lend it some extra strength.

Apple could be the biggest TSMC customer affected by the price increase. The upcoming iPhone 15's A17 Bionic chip will use the Taiwanese company's 3nm process, and Apple has begrudgingly accepted TSMC's price hike. Last year, Apple accounted for a quarter of the company's sales.

Costlier semiconductors could also affect game console pricing. Sony increased the price of the PlayStation 5 in August, blaming supply chain problems, inflation, and manufacturing costs. Microsoft also admitted that it would eventually have to raise the price of its Xbox Series consoles, likely next year.

Permalink to story.

https://www.techspot.com/news/96742-tsmc-reportedly-plans-raise-3nm-wafer-prices-25.html

 

[seeprime]

The price per wafer will be up 25%. Won't there be more chips per wafer? This should mitigate much of the cost increase.

 

[yRaz]

No! Wafers are a fixed diameter (300mm). The only way you get more chips on a wafer is … make physically smaller chips. Sure, the density is increasing as you move from 5nm to 3nm but they won’t use those gains exclusively to “make the same chips but more per wafer.”

They’ll want to use those gains to increase performance (more transistors) or reduce heat too.

In general, chip area will likely stay about the same and we’ll receive a more performant generation.

Check Wikipedia out and look at the area size for each chip generation … A15 vs A14 vs A13 … and you can see what’s going on.

Obiwanbill

300mm wafers are a relatively new thing, so comparing wafer cost instead of cost per mm^2 doesn't really help anything. 90nm wafers were 100mm, or 31,400mm^2. 5 and 3nm waffers are 300mm, which is 282,000mm^2. This means the wafers can produce approximately 9 times the chips a 100mm wafer can

the 90nm, 100mm wafer cost $2,000
the 3nm, 300mm wafer costs $20,000

$20,000/9=~$2,200

adjusted for inflation, a 3nm, 300mm wafer is cheaper than a 90nm, 100mm wafer

actual manufacturing costs are extremely low for chips, most of the costs are in IP, R&D, marketing and corperate overhead. If you look at the actual cost of a product, something like the 4090 on costs about $150 chips and material, that is not the break even price to develop and produce the product.

 

[Burty117]

I was watching a guy work out the Bill of Materials for the 7950X and including packaging its about $70 to make a 7950X.
It's probably fairly accurate as well since TSMC publish a lot of the details (failure rate, distance between chips etc...) and since the 7950X is released, it's easy to find how much silicon is required. The unknown was the packaging so they guestmated $10 for packaging.

The markup on these things are massive, I'm sure they'll try to explain that we the customers will need to pay more but it's all a lie, they just want an excuse to charge more.

And before anyone comments back, yes I'm aware you need to factor in transport, marketing and the actual R&D to design the chip. But still, when you read stories like this "Wafer price increases by 25%" and AMD/Intel/Nvidia trying to tell you the reason they doubled the price of their chip is because of this wafer price, it's a complete lie, the math doesn't work out.

Edit, in-case anyone wanted to see how this was worked out:

 

[brucek]

The current console generation will not be moving to N3, correct? Meaning if a bunch of demand for the current N5 moves to N3, availability for the customers that remain on N5 could actually be better.

We may reach a point where even the raw capability exists to move some devices to better process nodes, economic forces keep them at older process nodes that are sufficient if not optimal.

 

[brucek]

I was watching a guy work out the Bill of Materials for the 7950X and including packaging its about $70

That's all true while also underplaying the magnitude of all those other costs (including distributor / retail cut) by quite a bit. AMD's last reported net margin was just under 10%, down from a recent peak of 25% with their resurgence:

https://www.macrotrends.net/stocks/charts/AMD/amd/profit-margins

Of course the margins on certain SKUs are much better than others, probably including the halo projects (although sometimes they are intentionally money losers for reputational effect.)

 

[DonquixoteIII]

No matter how you slice and dice it, the cost for the wafer is far exceeded by the number of gates you get. Somewhere near 166% more gates at 3nm vs 5. Of course the entire wafer will not be used just for gates, there will be small areas of blank space for packaging, but VERY small. So say a 150% overall gate increase. I think that TSMC's pricing is actually chump change for what you get.

 

[Puiu]

No! Wafers are a fixed diameter (300mm). The only way you get more chips on a wafer is … make physically smaller chips. Sure, the density is increasing as you move from 5nm to 3nm but they won’t use those gains exclusively to “make the same chips but more per wafer.”

They’ll want to use those gains to increase performance (more transistors) or reduce heat too.

In general, chip area will likely stay about the same and we’ll receive a more performant generation.

Check Wikipedia out and look at the area size for each chip generation … A15 vs A14 vs A13 … and you can see what’s going on.

Obiwanbill

If we look at what Nvidia did then we can assume that they'll manage to get more chips out of a smaller process node. The 4080 is almost half the size of the 3080 (379 mm² vs 628 mm²). At 608 mm² even the 4090 is smaller than the 3080-3090ti chips.

This kinda puts into perspective the price of the 4080 and why it is so bad, right?

 

[takaozo]

Tech advancements in last 10 years are a big thing for the industry. Now the raising costs are normal, all new is more expensive. I'm curios to see when this shrink will stop or slow down what will they do next.

 

[winjer]

These are some really high price increases. It was justified during the high demand of Covid restrictions and mining boom.
For example, before the 2020, the cost for a N7 wafer was under 7000US$, but now it's 10000US$.This cost should have gone down, as the process matures and it's development cost is abated.
Now the chip demand has plummeted, so prices should go back to normal.

I can understand that the cost of developing newer nodes is higher than ever, but it seems like a good chunk of these increases in prices is just a result of TSMC having little to no competition at the high end, so they just jack up the prices and force everyone else to pay.

 

[neeyik]

I can understand that the cost of developing newer nodes is higher than ever, but it seems like a good chunk of these increases in prices is just a result of TSMC having little to no competition at the high end, so they just jack up the prices and force everyone else to pay.

Samsung, GlobalFoundries, and UMC have all been increasing their wafer prices, year on year, for a good while now (price hikes were evident before Covid) so it's not just TSMC.

 

[winjer]

Samsung, GlobalFoundries, and UMC have all been increasing their wafer prices, year on year, for a good while now (price hikes were evident before Covid) so it's not just TSMC.

But not to this extent.

 

[neeyik]

But not to this extent.

No, certainly not, but it's not like TSMC are increasing N7 or N5 wafer costs by 25% - it's node-to-node comparison increase. Rumors earlier this year were purporting that TSMC haven't been seeing enough N3 orders to ramp up production, to offset the development and setup costs of N3. With AMD, Intel, and Nvidia using N6, N5, and N4 for the foreseeable future, only Apple seems to be in the order books by any notable quantity.

 

[winjer]

No, certainly not, but it's not like TSMC are increasing N7 or N5 wafer costs by 25% - it's node-to-node comparison increase. Rumors earlier this year were purporting that TSMC haven't been seeing enough N3 orders to ramp up production, to offset the development and setup costs of N3. With AMD, Intel, and Nvidia using N6, N5, and N4 for the foreseeable future, only Apple seems to be in the order books by any notable quantity.

Raising prices during a world wide recession, only reduces the amount of costumers even further.

 

[FuryX]

So RTX 5090(3nm) = $2000+
so much for "RTX ON"

 

[Uncle Al]

Just my opinion but the current group of chips is more than adequate except, perhaps, for military applications and for the ultra gung-ho individuals that simply must have the latest and greatest but for the everyday user this kind of improvement simply isn't necessary ....

 

[vimann]

Yay, phones weren't getting expensive enough...

 

[Watzupken]

If we look at what Nvidia did then we can assume that they'll manage to get more chips out of a smaller process node. The 4080 is almost half the size of the 3080 (379 mm² vs 628 mm²). At 608 mm² even the 4090 is smaller than the 3080-3090ti chips.

This kinda puts into perspective the price of the 4080 and why it is so bad, right?

To be honest, the RTX 4080 offers very poor value simply by comparing the die sizes of the AD102 and 103. The latter is significantly smaller. Furthermore, the performance delta and the marked up AIB cards made the situation worst.

Also while they get less chips per wafer with Samsung, the price of Samsung 10nm is going to be way cheaper than what they get with TSMC's 4/5nm. I am sure Nvidia got some preferential pricing from Samsung to entice Nvidia to move away from TSMC 7nm, and utilize a more matured 10nm (or what they call 8nm).

 

[Puiu]

Just my opinion but the current group of chips is more than adequate except, perhaps, for military applications and for the ultra gung-ho individuals that simply must have the latest and greatest but for the everyday user this kind of improvement simply isn't necessary ....

The military wants nothing to do with such chips. The military values redundancy, stability and longevity just as highly as performance. None of which can be applied to modern consumer GPUs. They get their special chips. (for example, russia gets them from washing machines and sex toys)

It's consumers (businesses too) that need these rapid advancements since we can use them immediately.