Engineers are still needed at key points throughout the design pipeline.
The proliferation of AI tools seems perfectly matched to fill a talent shortage, but a closer look shows the skills do not entirely overlap. Certain parts of the EDA pipeline require human engineers, and it seems likely to stay that way for the foreseeable future.
The dark art of analog design, the final word on safety-critical functional safety, high-level architectural decisions, product innovation, and creative problem-solving are where people shine.
And despite predictions of wholesale job replacement, the impact of AI on engineering jobs is more nuanced, depending on the nature of the tasks, the complexity of the work, and the current maturity of AI tools in each area.
“Creative, open-ended, and context-specific tasks — like analog design and conceptual work — are much harder for AI to fully replace,” said Anand Thiruvengadam, product management, senior director at Synopsys. “AI is more likely to augment, rather than replace, designers in complex areas, acting as a tool to boost productivity rather than a full substitute.”
The following are human-centric tasks, according to Thiruvengadam:
In the future, a Darwinian kind of AI system could become adept at solving new problems by randomly running through massive numbers of options, then choosing which ideas work best. “I don’t know if that is ever going to happen, but it’s certainly not going to be efficient,” said Matthew Graham, senior group director for verification software product management at Cadence. “This is what we do as human engineers. We say, ‘I need to solve this new problem. Let me think of all the different things that could potentially help me get creative.’ We self-select in our own minds. Go down a path, try that path, and so on. AI, at least right now, can’t quite do that.”
Fig. 1: Human brains and AI can solve problems better together. Source: Cadence
AI can offer a starting point for design. It can reduce the design cycle, help optimize faster, and improve time to market. But humans are needed to understand the concept of what the design is trying to do.
“That context is necessary,” said Nandan Nayampally, chief commercial officer at Baya Systems. “As much as you’d like an AI engine to understand the context, and there is certainly improvement there, it cannot fully translate all that. There are many things in context that we have that the machine doesn’t, so it needs guidance.”
To varying degrees, most companies are thinking about using AI to build hardware. “We’ve got investments looking at this area, so this is clearly the direction that the industry is going,” said Mohamed Awad, senior vice president and general manager of Arm’s Infrastructure Business [1]. “The question becomes, is it going to be all AI? Are you going to just push a button, and the chip is going to come up? This speaks to questions about the software. In the early days [of AI implementations], it’s going to be mostly higher-level functions. It’s going to be about streamlining and simplifying some of the rudimentary or less sophisticated tasks of the general engineer. But it would be naive to think that AI is not going to accelerate a lot of the core design principles and activity that happens, especially more mundane activity.”
Humans needed: verification, monitoring, and training
Human engineers need to understand how to get accurate and useful knowledge into the AI system and then check that what is coming out is correct.
“AI should not be a threat to your job, because in essence it is still going to require a human to verify what the AI is producing really is an optimal system,” said David Fritz, vice-president of hybrid-physical and virtual systems, automotive and mil-aero at Siemens EDA. “It’s not just read papers and consume a whole bunch of PDF files like large language models do. It is not that. It’s subtle knowledge. It’s the fact that there’s a likelihood that you know there’s a dependence between A and B that I’m unaware of. How do I go find that? You’re not going to discover that in a white paper somewhere, or a PhD dissertation. Engineers are going to be critical to getting the system to the point where it can do useful things for those who are less experienced. They’re still going to need to be there to monitor the system, to train the system as technologies progress. There is a bright future for engineering, but you’ve got to understand the role of AI, and it’s not just a natural language interface for our command line tool. That’s 90% of what’s been done, but it needs to be far more than that.”
Verification is crucial to avoid costly mistakes, and it’s the most time-consuming and expensive portion of the design process. But costs are going up everywhere. Mask sets and tape-outs are massively expensive in IC fabrication. There’s also a significant cost associated with PCB boards.
“Here, risk mitigation and time to market are the most important points,” said Alexander Petr, senior director at Keysight. “You’re trying to get everything right in your first iteration to save money and to be the first to market. Let’s assume you go with a group that says, ‘We retool the whole thing. We have an AI that creates a solution.’ Would you trust the AI to take something out that costs multiple millions of dollars? You must look at the workflow and decide where to inject humans. Where do you want to inject safety measures? Where do you want to inject verification steps? Trust is good, but verification is part of the game. If you follow that thought process, you will find that you need fewer people. But you’re trying to cover the skills gap. I don’t see that any jobs will be replaced. They will just be different. People won’t design. They will verify more.”
To this point, a number of new startups are focused on RTL verification. “Look at the signals, tell me what’s running the signals,” said Petr. “They have built AI tools, and they have been used for tape-outs, but if you ask people what they’ve done, it’s hardly production-ready. People don’t trust the system yet. It will take years, if not decades, to build that kind of trust, and there will be many iterations to go through. Also, if you look at the speed of some of the AI developments, they’re done in an agile fashion, meaning fail fast and try again. The first iterations of AI have issues. They hallucinate. They think they’re doing things right. Just recently there was an article about a whole database that was deleted by an AI. That’s not a good starting point. We need to go back to what skills we need to verify outputs. We need to figure out what tools we can build for those verification steps, then what we can automate. What can we retool? What can we replace? Overall, from a headcount perspective, AI is not going to change that.”
Analog, mixed-signal dark arts
Analog design is simply more difficult, not just a more difficult challenge to solve with AI/ML tooling. “Analog is a much harder domain, because everything is analog in the world,’ said Sathishkumar Balasubramanian, head of products at Siemens EDA. “You have abstraction, because analog is very close to physics. You have abstraction in digital design of 0s and 1s and Xs, as well as your system, software, and so forth.”
For analog/mixed signal, AI can be useful for analyzing, optimizing, and debugging, and it can serve as a natural language training buddy to help reduce learning barriers. However, something is lost in the move to AI tools.
“I used to read books, and they tell you how to solve this problem,” said Balasubramanian. “We work on it in the lab, we do breadboards, we do designs, and we solve a problem because we are focused on solving a problem. What has happened is that people are so far away from the problem because they’re trying to solve a tool that might help them solve a problem. They’re trying to learn it, to learn a different language, and learn how a tool operates that might help them solve a problem. They’re not really focusing on the crux of the problem itself, which is how to design the best op amp or the best PLL for this process node.”
Analog/mixed-signal became more difficult because engineers added more complexity through customized tools, customized skills, and customized databases. “We added a level of complexity on top of a tough subject that is becoming even more complex now,” Balasubramanian said.
Others agree that the AMS domain is tricky. “You still have people who consider some of that art,” said Petr. “They start drawing by hand. They go in, look at things, and say, ‘That doesn’t look right,’ which normally means, ‘It doesn’t look pretty. It doesn’t work.’ There’s a certain dimensionality to those problems, which is significantly higher because there are no standards, there are no rules established that you can follow, and that’s mostly because it’s analog. Analog is a much more dynamic signal than a discrete one, a binary one, so getting everything right in that domain takes significantly longer.”
AI companies in the AMS domain often promise to do synthesis, and contend it can be generated by an AI tool. “But if you dig deeper, they have established rules to do this, which restricts the degrees of freedom significantly,” Petr said. “To get to the point where you can explore, discover, and automate with an AI, with the degrees of freedom you have today in those domains, will take a very long time. You can simplify the problem and try to solve a smaller problem. You can make this look really impressive. But it’s a very limited scope.”
Safety-critical applications
For the aerospace, defense, and automotive sectors, functional safety engineers are especially critical. “Let’s say we need 8 or 10 full-time safety engineers, but we can only find 4 or 5 in the field. Let’s try to get this same overall net effect with 4 or 5 engineers with the support of machine learning,” said Andrew Johnson, engineering and technology leader, systems and functional safety engineering specialist at Imagination Technologies. “That could be possible, but I would suggest the 4 or 5 engineers you have left cannot be junior level. They need to be very experienced senior principal technical specialists in this field, who are hard to find. When you put AI in the mix, you need someone intelligent and experienced enough to know what the model is, what it can tell you, and whether to veto what the model is telling you. If you don’t know because you’ve got less or no experience, you’re just going to nod and say, ‘The computer says yes, so let’s move forward.’ That could be very dangerous.”
The aerospace/defense sector may see slower job loss to AI than other sectors. “Just by nature, certain industries will be slower to adopt because those industries move more slowly, for legitimate and less legitimate cultural reasons,” said Cadence’s Graham. “What is more likely is that the tools will adapt to those environments. We will necessarily create a set of AI-enabled tools that are secure, or that can operate in that secure environment effectively, rather than the industry completely ignoring it. I don’t think they have the choice.”
Those industries have to move cautiously, but they cannot ignore the inevitable. “It’s more likely that the tools will adapt,” said Graham. “Certain tools and specific versions of tools will adapt to that environment, rather than that environment not adopting them.”
Vibe coding supervisor
There is reason to be optimistic about digital-native Gen-Z and Gen Alpha’s abilities to take up AI and find new roles, especially those who have grown up coding. “I have a college-age kid right now who has a very specific view on AI,” said Michal Siwinski, chief marketing officer at Arteris. “Is he interested in doing some of the coding stuff that AI will probably end up doing? No, that’s boring level coding anyway. It’s low-level stuff that doesn’t even require a computing degree. Instead, figure out how you orchestrate AI and how you use it to do more interesting things. How do you do robotic systems? How do you figure out where electrical, mechanical, and all of that need to work together? That’s a whole different level of excitement that really is just starting. It’s going to be an evolution. I have two boys at home. They know how to code, and they continue to code, but their approach is that they’re not going to necessarily code something they can get on GitHub, because it’s already on GitHub. Coding still remains very relevant.”
Others warn of the dangers of unsupervised AI code. “I fall into the trap myself of relying totally on AI,” said Daniel Rose, founding AI engineer at ChipAgents. “If you rely on AI by itself, sometimes it will work, but there will be cases where it hallucinates and something goes wrong that you can’t figure out unless you actually know what’s going on in the background, what the language is working on. You have to understand the code that you’re generating. It’s just that AI will help you generate the code much more quickly than you could yourself.”
Vibe coding, which uses AI to generate functional code from natural language prompts, is rising in popularity, but it’s still not perfect. “AI does have delusions or hallucinations, and you need to have both domain expertise and understanding to make sure that what is given to you is viable and does what it’s supposed to do,” said Baya’s Nayampally. “This iteration moves faster, so you have to move faster, but at the same time you’re going higher level on what you need to get done.”
Conclusion
AI/ML is making inroads in semiconductor design. It is up to individual design teams and design infrastructure management to decide which tasks the AI can be trusted with.
The tasks AI will most easily take over include functional verification, regression testing, and coverage analysis, which are increasingly automated. “AI-powered tools can generate testbenches, predict coverage holes, and even suggest new test scenarios,” Synopsys’ Thiruvengadam noted. “This is because these tasks are rules-based, repetitive, and involve large datasets, making them well-suited for AI automation. In addition, automated place-and-route tools are already mature, and AI-enhanced tools are improving efficiency further. Routine digital layout tasks are becoming more automated.”
Finally, in a worst-case scenario, what if the power grid, internet, and AI suffer a long-term outage and we are left with a workforce that only knows how to use AI? “There’s always a possibility for blackouts,” said ChipAgents’ Rose. “We’ve seen that around the world, so you need people that aren’t totally reliant on AI, who can still get the job done without AI.”
Cadence’s Graham believes humans will always find a way. “I’m an optimist and I don’t see it happening. But necessity is the true mother of all invention, and if we are necessarily in that situation, we will necessarily invent a solution.”
Finally, because people have learned how to learn, they can go back and understand how processes existed. “If suddenly the concept of synthesis were vaporized from the Earth, and we say, ‘The only way that we can get to the actual mask is by doing this by hand,’ we would figure that out,” said Graham. “We have that capability. Not only that, but we have all the data that we have gotten to this point. We’re not discovering it through trial and error the way we did the first time. We got here once, we can absolutely get here again, and smart people are smart people, regardless of whether their starting point is here, here, or here.”
Reference
[1] DAC 62 Sunday SkyTalk: “The AI Imperative: What Will We Make of This Moment?”
Related Stories
The Limits Of AI’s Role In EDA Tools
AI is a set of algorithms capable of solving problems. But how relevant are they to the tasks that EDA performs?
How AI Will Impact Chip Design And Designers
How AI is reshaping EDA, and how it will help chipmakers to focus on domain-specific solutions.
Best Options For Using AI In Chip Design
Narrowly defined verticals offer the best opportunities for AI. Plus, what will the impact be on junior engineers?
AI’s Value In Chip Design Depends On Data Availability
AI can help engineers do their jobs better, but results can vary greatly by area of expertise and company size.
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply