Taiwan chip-making behemoth TSMC is about to embark on a long journey with its all-new 3nm process. It’s been reported the company would begin production on its most advanced node in the second half of 2022, which is just weeks away. Now, TSMC has officially unveiled the technology behind it. In a departure from its recent history, TSMC will be offering multiple versions of its 3nm node with customized transistor configurations.
TSMC has offered different versions of the same node before, but not always in the same way. Most recently, the company has offered a “6nm” product as an enhancement of 7nm and plans to offer a 4nm variant of 5nm. At the 16nm node, TSMC opted to label its enhanced 16nm as 12nm. Back at 28nm, the company ran multiple product lines and offered multiple flavors of the process tuned for different characteristics. While not all of these offerings were equivalent with each other, TSMC has a history of offering multiple variants on the same node.
In this case, the new approach is called FinFlex, and it’s a novel “menu” that allows chip designers such as AMD and Nvidia to customize their designs to enhance performance in different areas for 3nm products. TSMC says the reason for giving clients options is because in this line of work, everything is a compromise. For example, in one scenario a client might want absolute maximum clock speeds and performance. However, in a different scenario, efficiency might be the paramount concern. As we all know, achieving both of those goals with a single design is not easy. To achieve this TSMC will be offering different variations of the number of fins-per-transistor, which is a first. Overall, TSMC will be offering four versions of its N3 technology. There will be the original 3nm, then three variations that boost either performance, efficiency, or size (cost). The options include: 3-2 FIN, 2-2 FIN and 2-1 FIN, which will allow companies to exert precise control over the design of its 3nm products.
As the chart above shows, each of the three variations provide different benefits. The 3-2 arrangement provides the highest level of performance. The 2-2 configuration is a compromise, providing improvements in all three areas: efficiency, die size, and performance. The 2-1 design is the most efficient. As Tom’s Hardware points out, these different nodes will be named N3, N3E, N3P, and N3X. TSMC says a client can even mix and match these different fin types, all on the same die. As an example it could use several 2-2 blocks, and one 3-2 block, similar to what Intel calls a “hybrid” design. Although N3 will begin production this year, the variants will come online in 2023 and 2-24.
The “standard” N3 architecture will begin production soon and will be for the early adopters. So far this is reported to be Apple for its M2 Pro and Max chips, and Intel for its Meteor Lake GPU tiles. Companies such as AMD and Nvidia are expected to tap TSMC’s 3nm node, eventually. That will likely not occur until 2024 or thereabouts. Both companies have already confirmed they’re using the company’s 5nm node for its next-generation products. This includes the RTX 40-series GPUs, and AMD’s Zen 4 CPUs and RDNA3 GPUs. Presumably, other clients will adopt the FinFlex variants in the coming years, as its roadmap for this extends into 2024. TSMC has stated previously it expects 3nm to be a “long node,” which a lifespan of over three years. Also, 3nm is the company’s final FinFet node, after which it’ll move to a gate-all-around design with nanosheet transistors.
Overall this is an extremely interesting move by TSMC. No foundry has ever offered anything quite like it, and it arrives at a time when Intel is also ramping up its foundry efforts. We’re not sure if there’s a correlation there, because a technology like this probably began planning several years ago. Still, it’s a bold move on TSMC’s part. It should theoretically allow for a very wide range of products to be built on its most advanced node. That in turn helps it attract as many clients as it can. Over the last few years, foundries have been fighting for a dwindling pool of leading edge customers. That’s due to prohibitive costs and the fact that a lot of chips just don’t need to be made on the most advanced node. It is not clear if the recent surge of VC funding for AI made a difference in this state of affairs or not.
TSMC seems to be gearing up for a fight, come 2024 and 2025. That’s when Intel is expected to be transitioning away from FinFet as well, as it moves to RibbonFET transistors on its 20A process. As you may recall, Intel’s IDM 2.0 strategy calls for it to obtain “unquestioned leadership” in silicon fabrication by 2025. That’s precisely when TSMC’s N3X will be rolling out, which is its most powerful iteration of its FinFlex technology.