Question TSMC nodes over last 20 years. N2 is going to change everything.

[mikegg]

Node	Year	Density (MTr/mm²)	Wafer Cost ($)	Total Tr (M)	Transistors/$	Gain vs Prev
90 nm	2004	1.5	1,200	102,000	85.0M	–
65 nm	2006	2.1	1,800	142,800	79.3M	0.93×
40 nm	2008	4.5	2,500	306,000	122.4M	1.54×
28 nm	2011	9.8	3,000	666,400	222.1M	1.81×
20 nm	2014	18.0	4,000	1,224,000	306.0M	1.38×
16 nm FF	2015	28.9	6,000	1,965,200	327.5M	1.07×
10 nm	2017	52.5	7,500	3,570,000	476.0M	1.45×
7 nm	2018	91.2	9,300	6,201,600	666.8M	1.40×
5 nm	2020	171.3	16,000	11,648,400	728.0M	1.09×
3 nm (N3E)	2023	215.6	18,000	14,660,800	814.5M	1.12×
2 nm (N2)	2025	247.94	30,000	16,860,000	562.0M	0.69×
A16 (N2P)	2026	272.73	45,000	18,544,640	412.1M	0.73×
A14	2028	302.73	–	20,585,472	–	–

 

[mikegg]

My comments:

• The transistors/$1 for since N5 has stagnated
• For N2, it will reverse. It's going to get worse for the first time since 65nm. A lot worse.
• N2 has a ton of customers. TSMC says it has 2.5x more tape outs at the same timeframe as N5.
• Yet, for every $1, you're getting fewer transistors than N3.
• N2 is shaping out to be a node for AI accelerators and expensive enterprise products. Nvidia won't really care if the wafer cost is doubled when they're selling each chip for $50k and the biggest cost is actually the HBM.
• A16 is another regression from N2 based on rumored price per wafer.
• We don't have any rumored price for A14 but it should also be a regression from A16 in terms of transistors/$1.
• For consumers, I predict that gaming GPUs will stay on N3 for 2.5 generations. Maybe 3 generations. It doesn't make sense for consumers to buy an N2 GPU over N3 given that GPU gains mostly come from more cores (more transistors).

 

[Kryohi]

Where did you get these wafer cost estimates?
They are different from the most recent ones that were published, which showed much less of a difference between N3 and N2 wafer prices.

 

[mikegg]

They are different from the most recent ones that were published, which showed much less of a difference between N3 and N2 wafer prices.

Source then?

 

[dullard]

Where did you get these wafer cost estimates?

Since mikegg neglected to post his source, here is an article from June 4, 2025 with estimates that are close but do not exactly match his table:

TSMC could charge up to $45,000 for 1.6nm wafers — rumors allege a 50% increase in pricing over prior-gen wafers

That's a lot.

www.tomshardware.com

 

[johnsonwax]

For consumers, I predict that gaming GPUs will stay on N3 for 2.5 generations. Maybe 3 generations. It doesn't make sense for consumers to buy an N2 GPU over N3 given that GPU gains mostly come from more cores (more transistors).

Gaming PCs with top of the line GPUs are getting close to maximum household circuit wattage. They're going to have to start caring about power just like mobile does.

 

[gdansk]

They're going to have to start caring about power just like mobile does.

But Nvidia issued Order No. 227 after Maxwell in 2014. Not one step back. All power to the GPU.

 

[johnsonwax]

But Nvidia issued Order No. 227 after Maxwell in 2014. Not one step back. All power to the GPU.

They're going to have to put 220 on gaming PCs by that 3rd generation then.

 

[LightningZ71]

New PC power cords incoming: 4 spade 240VAC/30/40 AMP RV plugs... Plugs requiring 4 conductor 8 gauge lines with an Arc Flash safe pneumatic breaker...

 

[dullard]

delete

 

[jones377]

Good thread, do you think you could include tape-out cost per node in the table as well?

 

[NTMBK]

You've got to wonder if either Intel or Samsung might have an opportunity here. I can't see them catching up to TSMC in terms of performance, but if they can get better transistors/$ than 3nm they might have a chance to get some design wins.

(And we can only hope the AI bubble bursts.)

 

[LightningZ71]

Neither of them can reasonably hope to match TSMC in volume to get their facilities amortized low enough to support breaking even with lower per wafer costs than TSMC has the ability to match at will.