TSMC prepares for 2nm trial production using AI-assisted process

[Daniel Sims]

Rumor mill: As semiconductor mass production starts to enter the 3nm era, all of the major competitors are intensifying the race toward 2nm. While companies like Samsung and Intel look ahead to ever-smaller process nodes, industry leader TSMC is taking preliminary steps to stay ahead.

Sources have told Taiwan outlet Economic Daily that TSMC has started pre-production work on 2nm semiconductors, a crucial step toward trial production. The company's current timeline could keep it ahead of competitors like Intel and Samsung.

Reportedly, TSMC has sent engineers and support workers to the Baoshan, Taiwan Zhuke R&D plant to prepare for 2nm trial production, aiming to produce 1,000 wafers this year. The company declined to comment on the information. Previously confirmed roadmaps indicate plans for 2nm trial production in 2024, with mass production coming in 2025.

Interestingly, Economic Daily sources say TSMC's manufacturing process utilizes an AI-enhanced method called AutoDMP which is powered by Nvidia's DGX H100 chips. The AI optimizes chip designs 30 times faster than other techniques, improving energy efficiency and decreasing carbon emissions.

Nvidia's AutoDMP in action

With 2nm, TSMC will switch to gate-all-around (GAAFET) transistors, which can increase transistor density and shrink the leakage current. The new node could improve performance over 3nm by 10-15 percent at the same power level or use 20-25 percent less energy with the same performance.

Meanwhile, Intel wants to release its 20-Ångström (20A, essentially 2nm) node in 2024, with Samsung planning to start 2nm mass production in 2025. Both will use GAAFET and backside power delivery to increase logic density and decrease power leakage. Japanese company Rapidus aims to mass-produce 2nm semiconductors by 2027. Looking further ahead, Samsung intends to reach 1.4nm by 2027, while TSMC is making plans for 1nm.

In the near term, Samsung and TSMC are focused on improving their 3nm nodes. Samsung started on 3nm last June, while TSMC began mass production right before the new year.

However, Apple snapped up TSMC's entirely early supply of 3nm transistors, likely for upcoming products like the iPhone 15. This could enable Samsung to court other customers like Nvidia, Qualcomm, and Baidu for its forthcoming 3nm GAA node. Samsung, Intel, and TSMC plan to offer more efficient versions of 3nm in 2024.

Permalink to story.

https://www.techspot.com/news/98968-tsmc-prepares-2nm-trial-production-ai-assisted-process.html

 

[Kashim]

"Intel wants to release its 20-Ångström (20A, essentially 2nm) node in 2024"

Sure, I'll believe it when I see it. Intel was stuck on 14nm++++++++ for how man years? But suddenly they're going to leapfrog over all the competition, skip a bunch of steps and be the first to reach 2nm. Sounds totally legit.

(Hint: It's just a naming trick. They'll be behind TSMC and Samsung for a long time, but they'll name it something more advanced to pull the wool over the eyes of the gullible. The nano-meter naming convention is totally meaningless anyway. They might as well call it 20-Fermi, the next smallest size below Angstrom.)

 

[Axeia]

However, Apple snapped up TSMC's entirely early supply of 3nm transistors, likely for upcoming products like the iPhone 15. This could enable Samsung to court other customers like Nvidia, Qualcomm, and Baidu for its forthcoming 3nm GAA node. Samsung, Intel, and TSMC plan to offer more efficient versions of 3nm in 2024.

That only works if Intels/Samsungs 3nm are actually better (higher performance and/or lower power) than TSMCs older processes. The only time NVIDIA went with Samsung* instead of TSMC the generation after it showed massive power savings (RTX 3000 to 4000 series). So to I wouldn't be surprised if TSMCs 4N (or an even further optimized version of it) has Intels and Samsungs future nodes beat already.

*They have used Samsung before this as well but only for the low end products, so when performance matters they've always stuck to TSMC aside from one time.

 

[Porkous]

Wouldn't Intel be better by treading carefully first before investing heavily into fabs? Seems like a risky mistake that appeared with it's new CEO but not with the need coming from what it can do at the moment.

I dunno. For example, Intel cannot do chips for Samsung, AMD and apple but it can do for others.

Samsung would want to use it's process and not of competition. Amd would stick to tsmc for bleeding edge, while apple is already invested in tsmc.
So Intel would have to have the bleeding edge shrinking process for at least 5 years to attract clients