Handel Jones, CEO of International Business Strategies, sat down with Semiconductor Engineering to talk about the growth and impact of AI. What follows are excerpts of that conversation.

SE: What do you see as the impact of AI on semiconductors?

Jones: The fact that you have a 5G smart phone is because of AI. Steve Jobs changed the smart phone to a portable device, but the image processing you have is because of AI. The analogy is the impact of machine power on human muscle power, which allowed us to build tall buildings and transport people on jets. These advances changed society over the past 200 years. AI will have a similar impact on improving brain power, which is the ability to analyze and process data. You will have a digital twin, and that will have a lot more compute power and memory power. It will be something that you use to analyze different things. For an individual, initially it will have some limited capabilities. But in the industrial and military worlds, the impact will be huge very quickly.

SE: Will this all be digital? Or will it be a combination of digital and analog?

Jones: The way humans communicate with digital twins has to change, and that’s going to be analog. There will be different ways of communicating information to and from the digital twin. There won’t be huge limitations in the digital capabilities over the next few years. If you look at what some companies are doing, it’s more of a transformer concept, where you can use data from unsupervised learning. Leveraging information from data is going to explode. Today, for the data being generated, there are estimates that maybe 1% is being utilized.

SE: How do you determine how much accuracy is required?

Jones: Well, 3D face recognition in China right now is pretty accurate. And they’re doing a lot of it, using it in shops, and for medical and transportation. This is going to change dramatically. What you don’t have yet is a swarm concept, but that’s coming.

SE: AI architectures today are not terribly efficient. Do you see that changing?

Jones: The transformer algorithm is a big breakthrough in AI, because now you can basically change the architecture of your chips. We’ve been looking at NVIDIA’s Hopper GPU. We’ve also been looking at Google’s intelligent placement and routing. And Cadence’s Cerebrus technology is incredible, because it checks based on what you’ve done in the past, which allows you to narrow down the elimination of bugs and optimize PPA.

SE: But even though we’re seeing some interesting pieces, the evolution of real AI seems to be happening more slowly than you might expect, right? So if you look at semiconductor design and manufacturing, AI is just beginning to get used.

Jones: We have to do this in pieces, and there are several reasons for that. AI, and design in the cloud, can dramatically change the revenue streams of EDA vendors. They’re the ones responsible for developing all of this, so there is a kind of movement to the cloud. But it’s not happening as fast as we thought it should because you need to balance that against income, and because the big companies that spend money on EDA have their own cloud. So there is some resistance. There also is a big gap between what you have in manufacturing and what you have in design. The reticle and mask databases are bridges. We’re seeing some real innovation in that area. The real key area is functional verification. Emulation is big right now, and it’s expensive, but functional verification is where you really need to eliminate bugs. AI is a natural fit there. We think in three to five years AI will become a key part of functional verification. We see that within the EDA vendors, and you’ll need big money to develop these capabilities. You can’t do it with a small company. That’s going to make the designers more productive, which is good because we have a global shortage of design engineers.

SE: Where else do you see AI being applied in the semiconductor world?

Jones: In the pre-qualification, which is the architecture stage. That’s the weak link right now from an industry point of view. If you change the architecture, the amount of data you have to consider is incredible. In a new design, 80% of the IP was used in an older design. So if you do a design at 3nm, and there was a previous chip done at 4nm or 5nm, then you re-use 80% of the IP. Otherwise, the amount of time needed to re-qualify that IP is too long.

SE: So you see the chiplet approach working well here?

Jones: Yes, and if you look at what AMD has done with four blocks or eight blocks for the CPU, that’s brilliant. And it has been very successful for AMD. Apple is now adopting it, and we’re going to see others doing the same. For the high-volume companies, particularly if the chip size goes over 200mm², chiplets work well. But for DARPA, where they have low volume and many different types of technologies, this is very difficult.

SE: What’s your outlook for that approach?

Jones: The value of chiplets is really at the advanced nodes, where debugging of the chip is very extensive. Second, where you have long interconnects, the RC time constant and the transmission line effects are killing you. When we look at going to 2nm, the transistor is scaling nicely but the middle-of-line and back-end-of-line are not. The chiplet solves part of the back-end-of-line issues, although it doesn’t do much for the middle-of-line. But at 28nm, there’s no real value in chiplets. You probably have an RF block, a mixed-signal block, and maybe a bit of non-volatile memory. Making those into multiple chips becomes very cumbersome.

SE: There’s a lot of data to process here. Will AI help with that?

Jones: The amount of data is going to explode, and the way to solve that is with AI because humans can’t do it. The brain power of AI is going to be significantly greater than the brain power of humans. And China is moving ahead a lot faster with this than the U.S. The use of the phone in China is much more extensive than here. They have almost 2 million 5G base stations. It’s behind schedule, but coming up fast. They’re using digital currency in restaurants, and they can check your COVID status. They have your picture. Everyone in China is in the central database now, and China is much more focused on this than we are. They have mandated schedules for autonomous vehicles by 2035.

SE: So when AI does rule the world, are we better off? And how does it change?

Jones: If you look at machine power, some things are good and some are bad. But overall, society has progressed. But in AI, China is moving so fast that it’s going to give them incredible power. There is a military aspect, as well as automation of factories, and so on. And if you look at electric vehicles, China is going to be really focused on exports. A key part of electric vehicles is the battery. They can get the silicon carbide, and they are coming up fast on software for robo-taxis. It’s still limited, but they’re pushing it as a government initiative. One company is coming out with an electric vehicle that costs about $7,000, and it has reasonable capabilities. And Neo is pushing a replaceable battery, which is starting to get some traction.