Samsung’s public roadmap was recently unveiled at the Foundry Forum 2021 in China. However, its absence on the slide was particularly striking – this is Samsung 3GAE (3nm, GAA-Early), the first iteration of their 3nm technology. 3GAE was originally introduced in 2019 along with its follow-up 3GAP (3nm, GAA-Plus), but only the latter was featured in the presentation.
Before this roadmap was published, there were ongoing concerns about node health; Originally slated for risky production of 3GAE in late 2020 and mass production in 2021, the company posted its first 3nm test chips on film last month.
In addition, at a recent Applied Materials event, SemiAnalysis estimated that Dr. Chidi Chidambaram, VP of Engineering at Qualcomm, Samsung Foundry’s largest third-party customer, estimates GAA technology will not go into production until 2023-24.
While the estimate is deliberately vague for NDA reasons, it still lags Samsung one year behind TSMC’s plans to mass-produce 3nm grade silicon next year. This, combined with the lack of 3GAE in the published roadmap, has given rise to speculation that the node has been abandoned entirely in favor of 3GAP.
So it looks like Samsung indeed has killed off 3GAE – 3GAP only for 2023 HVM:
4LPE at least seems reasonable in terms of an intermediary jump.
— Andrei F. (@andreif7) July 7, 2021
In response to a request from AnandTech, the company said that 3GAP should still be mass-produced in 2022, and that this node is probably not being published simply because it is reserved for internal use in the Samsung LSI.
However, although this has been done in the past with other forms of the “-early” process from the company, they were still mapped; 5LPE saw third-party use in Qualcomm’s Snapdragon 888, but the node itself was derived from 7LPE.
The company also manufactures 4LPP based on more traditional FinFET technology. It (and its predecessor 4LPE) are now listed as their own node family, rather than developing into 5/7 nm class technologies.
It may be that the process has significant enough improvements to justify marketing as a new “title” node, or simply because of more significant differences in design and production.
However, the lack of 3GAE in the public roadmap and the FinFET-based 4LPP node aligning with Samsung’s announced 2022 target for 3GAE makes comparisons to Intel’s early 10nm Canon Lake processor uncomfortable.
Technically, they were shipped in 2018 to fulfill obligations to investors, but were so bad that they abandoned the further development of the older 14nm node; The company now completely ignores this generation and insists that Ice Lake was its first true 10nm lineup.
But perhaps at least 14 +++ was easier to understand than the branching tree of the simulated processes and their subtrees.