Discussion Leading Edge Foundry Node advances (TSMC, Samsung Foundry, Intel) - [2020 - 2025]

[DisEnchantment]

TSMC's N7 EUV is now in its second year of production and N5 is contributing to revenue for TSMC this quarter. N3 is scheduled for 2022 and I believe they have a good chance to reach that target.


N7 performance is more or less understood.


This year and next year TSMC is mainly increasing capacity to meet demands.

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

For Samsung the nodes are basically the same from 7LPP to 4 LPE, they just add incremental scaling boosters while the bulk of the tech is the same.

Samsung is already shipping 7LPP and will ship 6LPP in H2. Hopefully they fix any issues if at all.
They have two more intermediate nodes in between before going to 3GAE, most likely 5LPE will ship next year but for 4LPE it will probably be back to back with 3GAA since 3GAA is a parallel development with 7LPP enhancements.

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

Samsung's 3GAA will go for HVM in 2022 most likely, similar timeframe to TSMC's N3.
There are major differences in how the transistor will be fabricated due to the GAA but density for sure Samsung will be behind N3.
But there might be advantages for Samsung with regards to power and performance, so it may be better suited for some applications.
But for now we don't know how much of this is true and we can only rely on the marketing material.

This year there should be a lot more available wafers due to lack of demand from Smartphone vendors and increased capacity from TSMC and Samsung.
Lots of SoCs which dont need to be top end will be fabbed with N7 or 7LPP/6LPP instead of N5, so there will be lots of wafers around.

Most of the current 7nm designs are far from the advertized density from TSMC and Samsung. There is still potential for density increase compared to currently shipping products.
N5 is going to be the leading foundry node for the next couple of years.

For a lot of fabless companies out there, the processes and capacity available are quite good.

---------------------------------------------------------------------------------------------------------------------------------------------------


FEEL FREE TO CREATE A NEW THREAD FOR 2025+ OUTLOOK, I WILL LINK IT HERE

 

[jpiniero]

The enormous difference in transistor count between Power10 and Power11 is really odd. I wonder if it's mostly due to IBM deciding to count their transistors in a different way.

That's literally what the twitter post says.

 

[511]

The process is likely the same

 

[gdansk]

Samsung got saved by Tesla

Hmm? What saves Samsung is being part of Samsung Electronics.

It seems backward to credit Tesla for Samsung's perseverance.

 

[Hitman928]

The enormous difference in transistor count between Power10 and Power11 is really odd. I wonder if it's mostly due to IBM deciding to count their transistors in a different way.


I didn't know IBM was doing 7nm. Is it EUV? These aren't coming from Globalfoundries, right?

GF doesn’t have a 7nm process. IBM is partnering with Samsung for production AFAIK.

 

[Joe NYC]

than you are right but that was cancelled with 20A to produce ARL there in 2025 they are currently running Wafers in Arizona Intel ain't got customer beside Intel Products

They are running test wafers, while TSMC is running production wafers in Arizona.

Fabs take a lot of time to build (and that time will be shorter for TSMC subsequent fabs on the same site) but in 4-5 years, TSMC output in the US will exceed Intel's output.

I think it is better first comparison before comparing it to Taiwan.

 

[511]

Fabs take a lot of time to build (and that time will be shorter for TSMC subsequent fabs on the same site) but in 4-5 years, TSMC output in the US will exceed Intel's output.

Not happening tbh intel has UMC and it's other fabs as well. TSMC will surpass Intel post Build out of its 3 or 4 th phase.

 

[Joe NYC]

Not happening tbh intel has UMC and it's other fabs as well. TSMC will surpass Intel post Build out of its 3 or 4 th phase.

In Phase 2, they are already installing utilities, and Phase 3 broke ground in April of this year. So, first 3 phases will be in production in 4-5 years.

Possibly Phase 4 and 5 already under construction in the same timeframe.

Also, Advanced Packaging plant will break ground next year and will also be in production in the same timeframe.

 

[511]

Advanced packing is not wafer fabrication count that out Phase 2 is not gonna come online until the end of next year and it will be similarish capacity as N4 or slightly more.

 

[Joe NYC]

Advanced packing is not wafer fabrication count that out Phase 2 is not gonna come online until the end of next year and it will be similarish capacity as N4 or slightly more.

I mentioned 4-5 year timeline, in which TSMC will likely exceed Intel in US based production.

There are no solid numbers, but according to estimates, Phase 1 will be 20k wafers per month, so perhaps we can project the same for subsequent phases.

I think Phase 2 will be N3 and Phase 3 will be N2.

 

[511]

That would be 60K WSPM less than Intel capacity combined still

 

[Joe NYC]

That would be 60K WSPM less than Intel capacity combined still

You have to wonder what percentage of It's own capacity Intel is using... But TSMC said its US capacity is completely sold out for some time into the future...

 

[511]

You have to wonder what percentage of It's own capacity Intel is using... But TSMC said its US capacity is completely sold out for some time into the future...

On Intel 7 fabs they are at full capacity and last year they shipped 83.3K WPM even with dropped utilization it's going to be higher from US.

 

[Joe NYC]

On Intel 7 fabs they are at full capacity and last year they shipped 83.3K WPM even with dropped utilization it's going to be higher from US.

It is hard to square this level capacity with Intel's declining output and Intel's increase in outsourcing.

 

[511]

It is hard to square this level capacity with Intel's declining output and Intel's increase in outsourcing.

Yeah but it's difficult to say cause 18A is getting milked and so is I7 along with UMC 12

 

[MasterofZen]

https://www.reuters.com/business/mu...sign-165-billion-chip-supply-deal-2025-07-28/
$16.5 Billion, that's definitely puffed up number, like Trump math.

 

[Joe NYC]

https://www.reuters.com/business/mu...sign-165-billion-chip-supply-deal-2025-07-28/
$16.5 Billion, that's definitely puffed up number, like Trump math.

What basis do you make that ass-umption? Because Musk said that is the minimum number, he expects it to be much higher.

The next big thing for Tesla, after robo-taxis are humanoid robots of which Musk thinks he can sell millions of. Which is where the chips would go to.

 

[511]

I need to see the chip size to make any estimate regarding the amount of chips and SF2 is not in HVM it's HVM in H2 26 and even if we assume everything goes right there will not be Chips before 2027 at minimum in these robots.

 

[mikegg]

he expects it to be much higher.

I don't know if the $16.5b is made up. I think it is to be honest.

Let's assume $20k/wafer. $16b can buy 800k wafers. Assume 200 chips per wafer. That's 160 million chips. Tesla sold 1.8m cars in 2024. They put two chips per car and let's assume that Tesla doubles their sales by 2028 (big assumption given they're selling less in 2025 than 2023).

1.8m * 2 * 2 = 7.2m chips.

The deal runs through end of 2033. So 5 years of production.

7.2m/year * 5 years = 36m chips needed.

Let's increase the growth a bit more and assume that Tesla will need 50m chips just for the cars from 2028 - 2033.

So the Optimus Robots will use 110m chips from 2028 - 2033? How many Optimus Robots do you think they can sell?

Musk saying that it'll be "multiples" is likely bullshit. He doesn't know that. He's just pumping Tesla stocks. I expect BYD to leave Tesla in the dust and Chinese robot companies to do the same to Optimus.

Given that Samsung 2nm is likely in same class as TSMC N3P, Tesla's products may be behind companies that use more advanced TSMC nodes.

 

[Doug S]

What basis do you make that ass-umption? Because Musk said that is the minimum number, he expects it to be much higher.

The next big thing for Tesla, after robo-taxis are humanoid robots of which Musk thinks he can sell millions of. Which is where the chips would go to.

Musk will sell millions of robots the same way a Tesla drove from SF to NYC unassisted in 2017 like he promised. Oh wait, a Tesla STILL isn't capable of that!

He has to keep pushing the idea of a "next big thing" to keep the stock bubble inflated because his lies about the car business have caught up to him. Everyone can see Tesla's sales declining, and the gravy train of billions in government subsidies ends in a few months. No one who owns a Tesla is "making $100K a year as an unattended taxi" like he promised for 2020. Over the years Tesla has gone from #1 in the world in autonomous driving technology to no longer ranking in the top 10.

If you want to believe his ridiculous and unfounded claims, you do you, but the number is clearly pulled from his ass if this is about Tesla alone. He's got incentive to make people believe all that volume is coming from Tesla - gotta keep that stock bubble inflated long enough to screw Tesla shareholders by having it purchase xAI at a greatly inflated price just like when he had xAI buy Twitter. Then he could promote Tesla as an "AI stock" just like it went from electric car stock to clean energy stock, and most recently robot stock. Pump pump pump!

In order to make that number a reality Samsung is probably going to be making custom chips for xAI's clusters, rather than xAI continuing to buy Nvidia. They already had some custom stuff for the car, they've probably been working for a while on scaling that up. If it isn't as good as Nvidia at least it will be fabbed cheaper than Nvidia's stuff and not bear their massive markup.

 

[MasterofZen]

What basis do you make that ass-umption? Because Musk said that is the minimum number, he expects it to be much higher.

The next big thing for Tesla, after robo-taxis are humanoid robots of which Musk thinks he can sell millions of. Which is where the chips would go to.

Simple math and using your rusty brain.
The cost of a modern chip varies, primarily depending on die size/yield and specific node used. Apple A series chip using the latest node provides a good reference, and latest Apple A series chip costs about $50 dollars each. Tesla AI (previously known as HW) chips are bigger, but it is definitely smaller than reticle limit. Let's assume a $500 safe guess ($500 for one chip in automotive industry is very expensive, since this price do not include packaging, memory and other peripherals).

Even under this conservative estimate, Tesla needs to sell 33 MILLION cars/robots to justify this $16.5 billion. There is no way Tesla will sell these many cars/robots equipped with AI6 chips.

 

[Joe NYC]

Even under this conservative estimate, Tesla needs to sell 33 MILLION cars/robots to justify this $16.5 billion. There is no way Tesla will sell these many cars/robots equipped with AI6 chips.

The contract covers 8 years. Currently Tesla sells 1.8 million cars, when the contract starts, it may be selling 2 million and the end of the contract 4 million.

2 million cars * 8 years = 16 million cars
4 million cars * 8 years = 32 million cars

It does not cover any other needs. Suppose AMD and Tesla agree to make the infotainment systems at Samsung fab (since Musk is getting dirt cheap prices) rather than TSMC.

Suppose Tesla starts to sell robots.

 

[Doug S]

The contract covers 8 years. Currently Tesla sells 1.8 million cars, when the contract starts, it may be selling 2 million and the end of the contract 4 million.

Maybe you missed it but Tesla is selling FEWER cars this year. Musk alienating the people most likely to buy his cars by going all-in on Trump, then doubling down by alienating his new friends on the right when he directly threatened Trump on his way out the door has badly damaged Tesla's brand not just in the US and but the EU. Losing the $7500 tax credit in the US will hurt sales of EVs, which hurts Tesla more than any other automaker since that's all they make. In addition, the traditional automakers have been getting better at EVs - and are surpassing Tesla on battery technology. The biggest problem though may be the pleathora of Chinese EV makers, who are MASSIVELY undercutting Tesla's pricing in any country that doesn't have a domestic auto manufacturing industry to protect.

The odds clearly favor Tesla falling below 1 million cars a year vs ever exceeding 2 million. Tesla is no longer the only choice for an EV, and many people who would have previously considered a Tesla the car to beat if they were shopping for an EV now won't even give them a look.

 

[mikegg]

If you want to believe his ridiculous and unfounded claims, you do you, but the number is clearly pulled from his ass if this is about Tesla alone. He's got incentive to make people believe all that volume is coming from Tesla - gotta keep that stock bubble inflated long enough to screw Tesla shareholders by having it purchase xAI at a greatly inflated price just like when he had xAI buy Twitter. Then he could promote Tesla as an "AI stock" just like it went from electric car stock to clean energy stock, and most recently robot stock. Pump pump pump!

It's a pump like you said. He has to sell a vision of Tesla selling so many cars and robots in order to justify TSLA's 180+ PE ratio.

"$16.5b is just the start! We're going to be so successful that we're going to need many more chips than just $16.5b can buy".

I bet the $16.5b has plenty of clause in it to protect Tesla. I wouldn't be shocked if the real value is something like $5b with up to $16.5b in order as long as Samsung meets certain specs and Tesla sells enough cars and robots.

I do agree that he'll look to eventually merge xAI and Tesla together.

 

[LightningZ71]

Rapidus claims to be on track, producing 2nm GAA test wafers now, volume manufacturing due in 2027...

Link to Serve-The-Home article...

 

[Geddagod]

Idk if this has been discusses yet, but anyone check out the difference between the X4 on the mediatek 9300 vs the X4 on the mediatek 9400?
20-30% lower core area (both without the L2 cache arrays and with the L2), ~ same Fmax, and ~10-20% lower power on the low end of the curve.
The specFP curve actually shows even worse power numbers than the specint numbers...