PartsCube Global

Back to Home
Semiconductor Engineering

Can AI Agents Really Shift Chip Design Left?

Semiconductor Engineering
  • Agent orchestration at various stages of design, verification, and validation will require new methodologies and coordination across tools.
  • Engineering expertise remains essential, and proprietary data is still siloed.
  • Cost and complexity could limit adoption, giving larger companies an advantage.

Experts at the Table: AI agents are moving into more parts of the chip design flow, but connecting them across different tools and processes remains a challenge. Questions also remain about how to capture decades of engineering expertise, protect proprietary design data, and keep costs manageable. Semiconductor Engineering sat down to discuss these issues with Frank Schirrmeister, executive director of strategy programs and system solutions at Synopsys; Ronen Laviv, vice president of business development and strategic partnerships at Hiveware; Jack Berg, vice president of business development at Silvaco; Ashish Darbari, founder and CEO of Axiomise. What follows are excerpts of that conversation. To view part one of this discussion, click here. Part two is here.


L-R: Silvaco’s Berg; Synopsys’ Schirrmeister; Axiomise’s Darbari; Hiveware’s Laviv.

SE: If AI agents are modifying code, and you keep updating AI models and agents, how much can you really shift left?

Darbari: This is going to be a new methodology direction the industry will be taking, much like UVM, where you combine these different flows together. How do you combine all the agents together? I don’t get a sense that anybody really knows today how this combination is going to work. From what I have seen in different presentations, they’re still looking at how these things unfold. A lot of agentic AI companies lack the field experience of having seen these problems firsthand, so what they’re building is not what the consumer wants. In many cases, the big consumers are building their own stuff because they know what they want. They have the field experience. They have data from many years. So where will the big EDA vendors go to streamline this, and how that will shape the end use? I don’t know.

Laviv: One of the challenges with emulation is that when running in co-simulation mode, the simulation is slowing down. Maybe, instead of having UVM or other languages that slow it down, AI will get its own language and get SystemVerilog RTL synthesizable code that you can put in the machine. Then you’ll need less of the software, and you can run faster in emulation. I’m not sure yet if the methodology would change quickly, but UVM might not stay forever.

Darbari: That’s what Nvidia’s head of GPU said at DVCon U.S., isn’t it? The formal is replacing simulation, or is going to replace simulation. I’m less interested in how each of these individual technologies work with each other, but I’m very interested in whether all of this adoption will lead to a coherent story.

Schirrmeister: You have lifecycle, manufacturing, digital implementation, functional verification, and you have specification engineering. Making sense of it is the light bulb — making sure it works, getting it to layout, getting it into the actual physical silicon and 3D, and then observing it during the lifecycle. Each of these flows has many sub-components. Functional verification to emulation is like one of the five. There is virtual prototyping, formal, simulation, emulation, prototyping. Everybody gets their own agent. So you have an agent doing something localized. Now you start getting an agent that orchestrates these things together as a functional thing, and then you get an agent that connects to the front end. This is an onion of agents, with more and more layers on top. EDA vendors are all saying we have Level 5 stuff, and there’s truth to that within the scope of the operating domain.

Berg: This is very much about hierarchy.

Darbari: But who’s going to pay for all this? Do the small companies have the budget to tokenize their banks and run all of this?

Schirrmeister: Do you mean the guys building the product, or the guys building the tools?

Darbari: The ones running the product. These are the consumers. They would need to pay for the tokens. I have a feeling that only the big companies will be able to afford this. They’ve been running it full-on, saying they haven’t written any code. All of this is agents. But it would get very interesting if the cost of doing all of this could be significantly reduced.

Schirrmeister: What’s exciting about this is that all the smaller designs that used to be difficult because you had to go through all of this will benefit. If you want to be an Nvidia competitor, you still cannot do this out of your garage. That still has to be a big company.

Berg: You’ve got levels of agents at this point, and a Level 5 orchestrating the whole thing. The industry is going to work on building each of these, and then iterating and improving them. That’s the only way in which you’re going to get agents to learn and replace large groups of engineers doing the grunt work that’s necessary here.

SE: But that hierarchy needs to start with whatever shifts left to include to each one of these different blocks, right? You need to at least define, in general, what you’re going to be doing.

Berg: You’re going to have information being fed back to ensure that when they go to specification of functional that there are constraints place here. That way you can improve the probability of manufacturability.

Schirrmeister: Can I have an agent model and predict that path?

Berg: Yes, you can.

Schirrmeister: But you want to check, right? You need to make sure early on what the multi-physics effects of what you’re doing will be.

Berg: Yes, and without that it will be dead on arrival in manufacturing.

Schirrmeister: So now the question becomes, ‘How much can you really abstract?’

Laviv: AI is doing a great job on, say, 80% of the challenges. But then you have experience from individuals who have been in the industry for 30 years with unique skills that AI doesn’t have. We had a discussion where we talked about 250 different skills for power and emulation and format and verification. AI doesn’t have those skills today. The industry needs to figure out how to gather those skills in order to really get to a flow where an agent can talk to an agent and has all decades of experience that is hidden in Apple and Qualcomm and Intel.

Berg: It took 50 years, but in 2024 the Novel Prize in chemistry was given in part to a software called AlphaFold, which predicted with uncanny certainty how proteins fold. Proteins are much more complicated than the crystalline structures that we deal with on a daily basis. But it took 50 years to get there. That said, it’s a proof of concept. Whether or not you believe it can be done in semiconductors, it’s going to happen.

Laviv: The difference is that with proteins, everything is open.

Schirrmeister: Yes. Where does the data come from? Simon Davidmann (currently a professor at the University of Southampton) framed it as, ‘We have the big three. They have the base tools, and they build agent stacks on it. That’s why you have a chip stack, or super stack, or whatever.’ We are doing the same thing at Synopsys, and we’ll see who has the best traditional EDA. People who know AI well found EDA. But they need to learn EDA first, rather than the other way around, and they don’t have access to the data they need. The users have access to the data. Samsung talked about optimizing quality of service for a network-on-chip. They had a fast classic engine, an SoC simulator, an emulator, and they built their own LLM to feed all the data in there. That’s their company data, and it’s not going out to anybody. That’s why they’re doing it themselves. Cognichip said you need LLMs that are trained with real chip designs and people, which is why they hired chip designers and build chips.

SE: They’re also using that intelligence for the LLMs rather than the agents, right?

Schirrmeister: Exactly. And that’s the balance you need.

The post Can AI Agents Really Shift Chip Design Left? appeared first on Semiconductor Engineering.