TSMC Introduces Brand New 3nm “FinFlex” Node with Customized Variants Suited for Every Task

TSMC, or Taiwanese Semiconductor Manufacturing Company, is the world’s most valuable semiconductor manufacturer and on track to become the biggest overall in its industry. Since the company’s inception, it has only seen sustained growth, a trend that has exponentially sky-rocketed in the past few years.

In many ways, TSMC is the only manufacturer in the world capable of producing bleeding-edge chips employed in today’s CPUs, GPUs, MacBooks, iPhones, and so much more. The company supplies its semiconductors to the biggest names in tech, including AMD, NVIDIA, and Apple. 

TSMC’s current flagship product is its 5nm FinFET node. Right now, the aforementioned Apple is TSMC’s only customer that uses 5nm nodes, in its M1 and M2 chips. However, reports suggest the company’s M2 Pro and M2 Max SoC might actually be manufactured on a 3nm node. And today TSMC has made that official.

TSMC 3nm “FinFlex”

Just two years after launching 5nm FinFET, TSMC has just announced its next-gen 3nm process node and it comes with an interesting nuance. The company is planning to offer different variations of its 3nm node geared for specific use cases such as higher efficiency or maximum performance. 

See, not every device using TSMC’s 3nm node would use it for the same output. In some cases, squeezing every last bit of pure, raw performance from the chip is the priority, whereas, some clients would look to maximize efficiency. And, of course, some will opt for a best-of-both-worlds approach to achieve a balance of both.

Whatever the case may be, one thing is consistent through all of those scenarios and that’s compromise. A single design cannot provide solutions for every problem thrown at it and therefore multiple variations of that same design, each tailored for a particular use case. 

TSMC is calling this “FinFlex” so the full product name is technically 3nm FinFlex. It’s essentially a small menu of varying 3nm flavors which the client can pick and choose to best fit its desire. Under the hood, the company is achieving this through offering different numbers of fins-per-transistors in different variants of the process.

In total, there are four variants TSMC has come up with. Firstly, there’s the original, base variant of the node, which is followed up by three variations. One of those is made for higher efficiency, one for higher performance, and one for the biggest die size. TSMC is calling these 3-2 FIN, 2-2 FIN, and 2-1 FIN, respectively, however their actual names are a bit different.

As you can see in the graph above, the 3-2 FIN is optimized to offer the most amount of performance possible but it’s also the least efficient out of all. On the contrary, the 2-1 FIN puts paramount importance on delivering as much efficiency as possible. Lastly, the 2-2 FIN is almost like a balance between the two with minor improvements in all aspects. 

So, all in all, the standard 3nm node will be titled “N3“, the efficiency-focused one will be called “N3E“, the performance-oriented variant is referred to as “N3P“, and the final balanced variant which will be ideal for producing the largest size dies, that one is named “N3X“.

What this brings to the table

Interestingly, TSMC says that clients are not bound to just one of these variants either, instead they can mix and match the fins to customize the node according to their preference for use on the same die. A good example of this would be how Intel is using the new big.LITTLE hybrid core architecture for its new CPUs. In such a design, the N3P (performance) node could be used for the bigger cores while N3E (efficiency) could be used for small cores.

This is a very intriguing move by TSMC. By no means is this the company’s first foray into customized process nodes, but it has never been done on this level with such bleeding-edge hardware. In recent memory, the company has offered a 6nm node that was actually just an enhanced version of 7nm, and there were plans to do similar things with 5nm upgrading into 4nm. 

Moreover. the company’s enhanced 16nm node is labeled 12nm, which means actual transistor gate length is not representative of the product name anymore. The closest TSMC has ever been to its new 3nm offerings is with its old 28nm process node, which the company made multiple flavors of, each tuned specifically for certain use cases. 

TSMC is said to begin production on its 3nm process in just a few short weeks, come the second half of 2022. That being said, we’ll see it being utilized in actual products in 2023, at best. And you can bet Apple will be the first one to adopt it. In fact, TSMC will allocate the entire initial run of its 3nm process nodes exclusively to Apple.

Now, that’s not because Apple and TSMC have struck an unfair and anti-competitive deal together (hopefully), it’s actually because none of the other potential customers would be interested in 3nm right now. AMD, NVIDIA, and Intel are likely the only clients to think of upgrading to 3nm and all of them are booked right now.

AMD and NVIDIA’s next-generation of productions including GeForce RTX 40-series GPUs, and Ryzen 7000 CPUs and RDNA 3 GPUs, all will use 5nm processes instead. Intel will rely on its own, in-house fabrication for its processors. However, it is rumored that the company might be interested in using the new 3nm nodes to manufacture the GPU tiles in its upcoming Meteor Lake CPUs, but they’re due in 2023-24 as well.

3nm is still far away

So, no one is really in a hurry to jump on the 3nm train, maybe except Apple. Regardless, TSMC has a real winner on their hands with the new FinFlex strategy. By offering different variations of the same process node, the company is essentially appealing to as many clients as possible. The prospect of customizing the hardware for any specific use case is ultimately going to make 3nm stand out from the rest of the nodes on the market.

Speaking of other nodes, TSMC’s only competition ahead looks to be Intel, who’s also gearing up to obtain “unquestioned leadership” in the silicone fabrication sector by 2025. Just a few years from now, in ~2025, the Blue Team will ditch FinFET for RibbonFET transistors on its 20A process, which will apparently revolutionize chips for Intel. 

That’s around the time N3X, the most powerful iteration of the FinFlex technology, will also become prevalent around that time and that’s when the real fight will begin. While it’s roses and daises right now, a silent war is brewing between Intel and TSMC and with semiconductors proving to be as imperative as they are, it’s almost as if there won’t be any losers in this battle.