(Photo: Samsung)Last week, rumors surfaced that Samsung would soon claim it had beaten TSMC to the 3nm punch. Yesterday, Samsung confirmed the news with a celebratory press release. It is the first global foundry to begin 3nm silicon production. It’s also the first foundry to move beyond FinFET to gate-all-around (GAA) transistors, which is a major achievement. We should note that nowhere in Samsung’s press release does it say it’s begun high volume manufacturing (HVM). Therefore, it’s unclear how many wafer’s it’s capable of producing at this time.
The news marks the culmination of a multi-year effort by Samsung to hit this crucial milestone without serious delays. Samsung first announced its GAA design in 2019, calling it MBCFET. That stands for Multi-Bridge Channel Field Effect Transistor, which is a nanosheet design. This is different than Intel’s RibbonFET process, which uses thin nanowires. TSMC will also be moving to nanosheets, but not until 2nm. For 3nm it’s sticking with FinFET via a customizable design named FinFlex. TSMC is expected to begin 3nm production sometime in late 2022.
The announcement from Samsung offered some numbers to quantify the benefits of the move to GAA. Compared to 5nm, its 3nm process allows for a 45 percent reduction in power consumption, 23 percent more performance, and a 16 percent reduction in area. The company’s second generation 3nm process will take things even further. It’s promising a 50 percent reduction in power, 30 percent more performance, and a reduction in area of 35 percent. Samsung has said previously its second generation design will arrive approximately one year after its predecessor. Its first generation 3nm nanosheet design is focused on “high performance, low power computing” the company stated. It will then move to focusing on chips for mobile applications.
Samsung’s nanosheet design will allow it to precisely fine-tune the characteristics of the transistors. By tweaking the width of the sheets, it can customize the power and performance curve with more precision than what FinFET allows. For more power/performance it can use wider sheets, or narrower sheets to improve efficiency.
So far Samsung hasn’t announced any official customers or products for its 3nm process. It’s also not clear when it will begin high volume manufacturing, and what kind of yields it’s achieved. Still, it’s a shot across the bow of its main rival TSMC. As noted by Anandtech, Samsung has been losing customers to TSMC lately. In May Qualcomm announced it was moving to TSMC for its Snapdragon 8+ Gen 1 SoC. The company also saw Nvidia use its 8nm node for Ampere, then switch to TSMC for its upcoming RTX 40-series GPUs. Those will be built on TSMC’s 5nm node.
Obviously, Samsung would like to reverse that trend, and reaching 3nm is a feather in its cap along those efforts, if it can get yields up. When TSMC begins 3nm production later this year, it’ll be interesting to see how its flexible FinFET design stacks up against Samsung’s GAAFET. Intel isn’t expected to leave the FinFET era until 2024 with its 20A “Angstrom” process. Like TSMC’s, its Intel 3 node will still use FinFET.