TSMC will lead rivals Samsung and Intel for years in advanced semiconductor nodes with the recent launch of its 2-nm process, according to analysts surveyed by EE Times.

The Taiwan foundry’s initial 2-nm customers will include top chip designers like Nvidia, Apple, AMD, and Qualcomm, as well as hyperscalers like Microsoft, Amazon, and Google, according to the analysts.

Late in December, TSMC announced the production start of N2, the company’s term for 2 nm. N2 marks the first time for TSMC to adopt nanosheet transistor technology, also known as gate all around (GAA), years after foundry rival Samsung. Claiming full-node advances in performance and power consumption, TSMC developed low-resistance redistribution layer and super high-performance metal-insulator-metal capacitors to boost performance of the 2-nm node.

TSMC said its 2-nm tech will lead the chip industry in density and energy efficiency.

“With our strategy for continuous enhancements, N2 and its derivatives will further extend TSMC technology leadership well into the future,” the company said in a prepared statement.

The champion

Analysts agreed that TSMC will maintain its title as champion of the advanced nodes. The company will keep its most advanced tech on Taiwan, and that is a concern for some. The U.S. government has called the company’s advanced tech leadership a national security risk.

TSMC will have some 2-nm capacity in the U.S. by 2028, but leadership capacity and most advanced technology nodes will remain in Taiwan, International Business Strategies CEO Handel Jones told EE Times.

“TSMC’s N2 is looking like it will be one of the company’s greatest hits, with over 15 customers working on it,” TechInsights vice chair Dan Hutcheson told EE Times. “Publicly known customers with interest include Apple, Nvidia, Google, Broadcom, Marvell, AMD, Intel, and MediaTek.”

The analysts noted that design costs of up to $800 million for a new chip are a tiny part of wafer production costs reaching $20 to $30 billion per year for some TSMC customers. Chip designers in advanced nodes must make huge bets on foundries supplying wafers with high yield and on schedule.

TSMC has continued to execute its tech roadmap, providing customers sufficient capacity while its foundry rivals Samsung and Intel have struggled to deliver on promises for advanced-node production.

In October last year, Intel announced the ramp of its 18A GAA process, the company’s entrant against TSMC’s N2. Intel is using 18A to make its Panther Lake and Clearwater Forest processors at the company’s new Fab 52 in Arizona. Intel originally scheduled 18A for volume production by early 2025, with initial products like Panther Lake launching in 2025.

The gap between TSMC and the competition is widening rather than closing, Jones noted.

“It’s a big problem for the industry for there to be only one high-volume source for wafers and packaging,” Jones said.

Samsung, which in 2022 became the world’s first chipmaker to adopt GAA in its fabrication process, stumbled with adoption of the new tech that is used to make 3D chips for AI processors.

“Samsung is still having issues, but now that Intel announced Panther Lake is in production, as the process matures, I expect Intel to gain some new foundry customers,” independent analyst Mike Demler told EE Times.

“For TSMC, 2 nm is its first GAA node, but it won’t have the backside power technology of Intel’s 18A,” Demler said. “TSMC will incrementally introduce improvements later in 2NP and 1.6 nm, then 1.4 nm. Collectively, those technologies will comprise the leading edge for the next three to five years.”

N2 will validate GAA even as it lacks backside power, Hutchison noted.

“The result is that while N2 makes great speed and power gains, its density gain is mediocre compared to [TSMC’s previous] N3E,” he said.

Competition

Intel and Samsung have failed to grab top customers from TSMC, which makes more than 90% of the world’s most advanced chips. TSMC’s position may slip for a variety of reasons, including geopolitics, Paul Triolo, China and Technology Policy Lead at DGA Group, told EE Times.

“The competition at the most advanced nodes here has become more complex with both Samsung and Intel, and now Rapidus in the mix,” Triolo said. “Intel and Samsung could have a shot in terms of capacity and supply assurance, multi-site redundancy, and geopolitical risk hedging—non-Taiwan based production—and commercial terms such as wafer pricing, engineering support and guaranteeing schedules.”

TSMC itself has recently built more fabs outside Taiwan, an island which could be in the crosshairs of a future conflict involving the U.S. and China.

Triolo notes that TSMC’s overall market share lead is “very large,” possibly exceeding 70%. From that base, competitor “inroads” are more likely to look like selective wins in high-value segments than a rapid displacement of TSMC’s overall dominance, he added.

“Intel can win meaningful share in U.S.-sovereign/hyperscaler custom silicon and some networking/ASIC niches if 18A ramps credibly and the ecosystem hardens—AWS-type programs are the template here,” Triolo said. “Samsung can win share via its U.S. fabs and anchor ramps. Here, Tesla/Apple supply relationships are the template, especially where customers explicitly want a second advanced-node source.”

Japanese startup Rapidus is a bit of a wildcard, he said.

Rapidus, which aims to start 2-nm production in 2027, could become a credible option for Japanese/Asia-based customers prioritizing supply chain sovereignty and specific advanced node programs they are willing to co-develop with the foundry, Triolo said.

TSMC appears to have reduced the probability of a Samsung-style GAA stumble by waiting longer to introduce GAA (first at N2 rather than N3), running an extended baseline/yield-enhancement program with early customer tape-outs, and now declaring on-schedule volume production, Triolo added.

“The only conclusive demonstration that the ‘hard parts’ are fully sorted is when multiple high-volume customer products, [mobile and HPC], ship on N2 with stable yields and predictable ramp rates through 2026,” Triolo said. “One area to watch is how SRAM and high-current blocks are behaving as this can be where early GAA pain can reveal itself, and also look at time from tape-out to high-volume manufacturing and whether there is significant slippage in schedules.”

Given TSMC’s track record, it seems likely that the company will cross these hurdles in 2026, he added.

“Expecting TSMC to stumble and provide openings for Intel and Samsung may not be the best bet,” Triolo concluded.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved