Will AI Drive Scaling Forward?

The almost ubiquitous rollout of AI and its offshoots—machine learning, deep learning, neural nets of all types—will require significantly more processing power as the amount of data that needs to be processed continues to grow by orders of magnitude. What isn’t clear yet is how that will affect semiconductor manufacturing or how quickly that might happen.

AI is more than the latest buzz term. It’s a transformative horizontal technology, meaning it can be applied in a variety of different market segments, sometimes in multiple ways. These days everyone seems to have an AI/ML/DL plan, either to bundle it into end products such as assisted or autonomous driving or a security camera that can be weighted to ignore certain behaviors or intrusions and focus on others. It also can be used to improve the economics and/or reliability of those products at various stages of their development, including design, verification, test and manufacturing. And it can be applied to components or to entire systems and systems of systems.

Much of this already is apparent inside of data centers, where very large and increasingly custom chips are being developed to process large amounts of data very quickly, using new architectures that scatter processing elements and small memories around a die or a package. There seems to be no upper limit to the appetite for more processing power in these operations, which is helping to make investments in 7nm by both TSMC and Samsung look very timely. The big question going forward is whether AI will drive demand for 5nm and 3nm, and there is a mountain of money riding on the answer.

Unlike in the past, when moving to the next node pretty much guaranteed new business for the entire supply chain, the biggest drivers of transistor density at 7nm have been mobile phones, cryptocurrency, server chips, and AI. Two of those four applications are looking far less lucrative these days. Cryptocurrency has pretty much deflated. Alongside of that, the mobile phone market has flattened. While that is still good business, companies like Apple,Samsung and Huawei are now pushing heavily into advanced packaging. Apple has even started designing its own processors, and so far it’s not clear yet when or if the company will require 5nm, let alone 3nm. Apple is TSMC’s largest leading-edge customer, so any shift could have significant fallout.

AI already is a big driver of advanced-node server chips, including FPGAs and GPUs. But as ASICs begin taking over a larger chunk of the AI world—and particularly as AI processing moves closer to the edge, with more heterogeneous processors added into the mix— it’s possible those chips will be packaged as chiplets. That means not all of the components need to be developed at the latest node, which could significantly impact the volume for the next nodes.

So where do foundries and equipment makers place their bets and when? An advanced process at a foundry today costs somewhere in the neighborhood of $1 billion, and it could go up significantly as manufacturing of new nodes becomes more complicated. Add to that new tools and equipment, such as a high NA EUV scanner, changes in the manufacturing flow to allow more points for testing, and greater demand for rapid customization, and the investments in new materials and equipment across the manufacturing ecosystem become far less automatic.

Making a new process viable requires the involvement of an entire ecosystem, from IP and tool providers to new materials and various pieces of equipment that range from etchers and wafer cleaners to testers. And it requires all of them to come together at roughly the same time, which is why the delay of EUV caused so much chaos in the supply chain.

But 5nm, 3nm, and even 1.5nm, are very different challenges.
They require new types of transistors, new ways of building those transistors, including various types of etch, deposition and even growing structures on a substrate or other material. And they will need new inspection and metrology equipment for what could well be one-off types of designs with unpredictable demand and longevity.

That makes placing bets on future nodes a very high-stakes gamble. It will be interesting to see who has the stomach for placing those bets, and how the industry could change over the next five years if some of those bets don’t turn out quite the way everyone expects.