How AI Will Impact Chip Design And Designers

Experts at the Table: Semiconductor Engineering sat down to discuss the role and impact of AI in chip design with Chuck Alpert, Cadence Fellow; Sathish Balasubramanian, head of product marketing and senior director for custom IC at Siemens EDA; Anand Thiruvengadam, senior director and head of AI product management at Synopsys; Sailesh Kumar, CEO of Baya Systems; Mehir Arora, head of engineering at ChipAgents; Daren McLearnon, product manager at Keysight. What follows are excerpts of that conversation.

L-R: Keysight’s McLearnon; Synopsys’ Thiruvengadam; ChipAgents’ Arora; Cadence’s Alpert; Siemens’ Balasubramanian; Baya Systems’ Kumar.

SE: Where are we with AI in terms of, how do you see it playing into chip design?

McLearnon: AI is transforming the microwave and III-V design processes, as well as the systems and applications surrounding them. If you look at 6G, the entire physical layer is adaptive. Those designs are reconfigured on the fly, and they require different architectures to provide that experience. So the design tools underneath them are changing, but also the what’s delivered and how it’s used is going to be needs to be more dynamically related.

Thiruvengadam: If you look at what is really driving AI, the blocks we see today in computers are getting pushed to the limits. Chip design in general is getting pushed to the limit. Compute, memory, interconnect, and then that is compounded by shorter development cycles. Right? You have many of our tier-one SoC customers going to yearly rhythms for development cycles. And then there is one more aspect to it, which is the talent shortage. If you take the design getting pushed to the limits, compounded by shorter development cycles and a talent shortage, that’s a perfect recipe for disruption from AI. AI is going to be pervasive. It is going to impact every aspect of chip design. We see it as a horizontal that’s going to impact every vertical, from a design or verification point of view, for the entire flow from spec to tape-out and beyond.

Alpert: In the short term, we’re trying to improve designer productivity. One of the ways to do that is help-assistance and co-generation. A lot of what we do is Tcl scripting, Python scripting, and in Cadence, SKILL scripts. We generate a lot of collateral, including SDC constraints. It’s helping with those kinds of fairly mundane tasks, and accelerating them. As we hire junior engineers, they might not have Tcl experts anymore. You can use natural language to do your debug. Those are short-term things that will really help aid productivity. But that’s not changing the way we do design. What will change the way we do design is when we start building agentic workflows, where we have agents discover different tools, and then use those tools to optimize around that. That’s what is going to be the game changer. So there are two areas of focus. One is just productivity. The other is an agentic workforce.

Arora: AI already has impacted everybody horizontally across the entire stretch. People are trying to use it for code generation at every level, up and down the stack, the front end, the back end, all the way to GDS and beyond. What we’re really interested in is that as AI accelerates dramatically, under a sort of crucible of constraints, the human starts to become the bottleneck for many different agentic workflows. A lot of serious progress is going to be made by expanding the vertical. Many serious verticals within hardware design are going to be carried out completely, end-to-end, because this is the only way you can further accelerate progress. We’re dealing with monumental tasks here — the tier-one SoC. If you’ve got a one-year tape-out timeline for every single generation, how do you even do that? How do you continue to compress and continue to accelerate, unless you can take some of these entire verticals end-to-end? That’s what we’re very interested in. The latest crop of foundation models is making this possible with the onset of reasoning models, fully agentic end-to-end training. Especially with reinforcement learning, we’re going to see this happen, and that’s where the industry needs to head.

Kumar: AI definitely is changing the whole compute paradigm. Ten years ago the entire compute was based on CPU compute, with x86 and Arm. Now the lion’s share has moved to a GPU-based computer architecture. So the computer architecture paradigm is completely changing, and we are still in the early phase. We are doing GPU-based compute and accelerator-based compute, but it’s just a matter of time before we get into true neural network compute and those types of things. New software paradigms will come as a result of that, because software always follows hardware. Once we have a mature software/hardware platform architecture, then software will start to develop around that. That’s already happening with CUDA and other things. As this transition happens, there will be lots of opportunities for everybody. We’re finally seeing massive amounts of investment coming into the hardware world. The chip industry, from a VC standpoint, was mostly dormant for the last two decades. NVIDIA is the most valued company now, and it’s a chip company. This will impact every aspect of hardware, then software, IP, CAD tools. Engineers want to change things and build new things, and this is a great opportunity for that.

Balasubramanian: We’ve been doing AI for a long time. AI is a pretty loaded word. It includes machine learning, reinforcement learning, GenAI, agents, and agentic AI. Within the EDA industry we have several products that already are doing machine learning and reinforcement learning. We view it as hybrid AI, where you have tools within tools, with ML and reinforcement learning already built in on top of that with GenAI and LLMs coming into play. But the biggest change is that customers are telling us, ‘I don’t want to hire a tool expert. I want to hire a domain expert.’ So with GenAI and agentic AI, the key thing is that you’re not going to solve every problem. It’s a vertical flow. It’s purely based on use cases. It’s not like I can go and type something in a tool and say, ‘Design me a PLL at 2nm at this frequency.’ That’s not going to happen today. It’s still a long way off. But you can easily automate a lot of manual tasks that had humans in the loop. So you don’t need to have humans in the loop as often, but you still need them just to make sure it’s correct. We are at that stage. But it’s definitely a productivity gain.

SE: So whose jobs are at risk? And are they really at risk, or do they just need to retool?

Alpert: We have an engineering talent shortage. We’ve had one forever. EDA tools have made immense progress over the years. If you think about where EDA is now versus 34 years ago, the productivity improvements are incredible, and yet we still have a talent shortage. AI is going to improve our productivity. Let’s say — pick a number — it’s been 10%. Now it will be X plus 10%. Or instead of it being 20%, it will be 30%. That’s great, but we’ll just have more work for them to do, and they’ll have to change their skill set. But they’re engineers. They always change their skill set because they’re smart people. So I don’t see anybody losing their job — at least not in the next 10 years.

Thiruvengadam: A talent shortage is one of the fundamental problems we have in this industry. The other thing we should look at, especially when it comes to agentic AI, is the opportunity to offload highly repetitive tasks to agents so that the humans in the loop can focus on higher-value tasks. That’s value creation. So humans will be focused on the higher aspects in the value chain, and the lower aspects are being offloaded to agents. That’s one way to think about it, and it takes away the negative connotation of, ‘AI is going to take away engineering jobs.’ That’s an important message to drive out in the industry, because when you talk to customers and show improvements in productivity and turnaround time, the first question they ask is, ‘Can I reduce my workforce? How much cost savings can I extract?’ But that’s not the right message to share with the industry. It’s more about how to get engineers to do more with the existing resources, or to accelerate innovation engineering velocity.

Balasubramanian: From the customer side, I do see that kind of thinking. They’re a little bit scared. But if you go back 20 years, we used to do layouts differently, and at the time people thought they were going to lose their jobs because EDA was going to take over. But we just started doing more. We are at a very good stage where I can go hire people for their domain expertise. That’s key. Customers are really focusing on the problem they want to solve, rather than figuring out how to use what I call passive software, where we provide them with EDA software and the customer builds stuff around it. We are transitioning to more interactive software where it’s easier for customer to start automating,

Alpert: Those were polygon pushers. But if you can’t adapt to new recipes, then you’re at risk. We’re engineers here. Everybody should be able to do that.

Kumar: We have gone through centuries of disruption, starting from industrial age — the motor, telecom, the internet. And every time we thought half of the population is going to lose their jobs. But it never happened. We’ll always find things to do. And the people who are going to excel are the ones who are going to use the latest cutting-edge technologies and tools. In my view, AI is an amazing tool, and whoever is going to be able to leverage it effectively to provide some sort of value or service is going to succeed.

Alpert: For the general population, I do think there’s risk. But for the engineering population I don’t see it.

Kumar: Even non-engineers will benefit. There always are ways to provide some value to others that they are willing to pay for.

Alpert: But certain jobs could be automated out of existence. Most innovations happen over time, so there’s time to make those changes. But AI is accelerating at such a pace that there could be huge swaths of jobs that will change.

Arora: We’re talking about AI in kind of an additive way — 10% on top. But think about it as a force multiplier, especially for small teams today. This is the best time for someone with a little bit of domain expertise, a little bit of senior experience, and maybe you’ve done a tape-out, to start your own company. You can start your own effort because now you can use agentic AI to carry out huge swaths of the verification and design effort and make those designs higher quality. You can have smaller, more focused teams with more domain expertise. There’s an infinite demand for more specialized hardware, and for AI accelerators. And the algorithm du jour changes every six months when a new set of papers is published out of frontier labs. We need better algorithms for attention. Maybe we’ll need better algorithms for state space models and so forth. In the future, there’s going to be a huge amount of demand for all this specialized material to come out of here. So that’s why people who have a little bit of senior expertise are going to be doing remarkably well. They’re going to have an enormous force multiplier on what they can do. The big question is what happens to junior positions, and this has always been the case with disruption.

McLearnon: I agree with the point on differentiated design — automating things that are poorly differentiated and just sort of monetized. There will always be these frontier designs on materials or processes or applications that push the edge of what’s ever been done before, and you’re always going to have human hands architecting the automation of the mundane to liberate the noble engineers. But when you automate entry level jobs out of the pipeline, where do these experienced people come from? That’s a challenge for our university educators and the pipelines of matriculating new talent.

Thiruvengadam: For any of the EDA vendors, it cannot be focused on one persona. It’s focused on all the personas. For any chip design you have a junior engineer persona with a certain level of tasks and responsibilities, a mid-level, and a senior level. When we develop solutions, we have to map those personas. There is a set of problems that can be addressed with agentic AI. So I’m not too worried about the junior engineer being pushed out.

Kumar: But if you look at the venture capital-backed startups, there are $1 billion companies that have less than 10 people. That’s a single-digit workforce with a billion-dollar-plus valuation. And VCs are looking for their next billion-dollar company with one employee, and they claim it will happen next year. From that standpoint, if you have this very concentrated wealth, it is going to create very interesting trends in terms of employment. But then the money always trickles down. That person who is the billionaire is going to create more employment.

Alpert: And if one person created a billion dollar company, then it should be copyable, and other people will copy them.

Balasubramanian: Exactly. You have to differentiate between what’s commercial and what is solving complex problems. So there are certain standards from us or our customers. I can’t have Calibre give a wrong whistle. That’s the sign-off. If I’m rolling out an AI technology in Calibre, it needs to be correct to the physics. That’s important.