Jerome Toublanc, Synopsys: On AI-driven approaches addressing silicon-to-system engineering challenges

The effect of AI on EDA and across chip design is undeniable. Writing in Semiconductor Digest earlier this year, Srirangan S of ACL Digital described AI as ‘the driving force behind how chips are conceptualised, architected, verified, and optimised.’

It is the third of these which is of particular interest. Verification has traditionally been the toughest part of design – historically consuming up to 70% of design cycles, as Srirangan notes. As chip complexity has grown, the verification state space has grown with it. But how much of the bottleneck can be solved here?

Jerome Toublanc is a Principal Product Manager at Synopsys, leading the integration of multiphysics methodologies within the Product Management Group and focusing on AI-driven approaches that address silicon-to-system engineering challenges. Toublanc sees AI as a natural fit for verification, but also the first step.

“When we look deeper at verification, it’s a perfect fit for AI,” says Toublanc. “It’s about a lot of data, it’s about a lot of classification, a lot of report, a lot of things to consolidate, to filter, to classify. And this is definitely a natural fit for the existing AI technology.”

Once the value of AI for verification is proven, the next frontier will come into view – and Toublanc cites physics AI as an example. “The trouble is, it’s by far more complex; as soon as you move a cell, or move a chiplet, you have so many physics, so many couplings, so many interactions, you have to take into account that agentic AI won’t be enough,” says Toublanc. “It means we will have to have proper tools that will feed agentic AI properly, but make sure we can understand what’s happening, make sure we can understand the action.”

The goal for Synopsys, therefore, is to anticipate the next step. The company already has strong heritage in this space, claiming to be the first to deliver gen AI agents and ‘envision the future’ of agentic AI through its L1 to L5 framework – from basic, assistive automation to, ultimately, highly autonomous, self-directed agents.

Alongside this, the company’s acquisition of Ansys, completed last year, adds further heft across system simulation and analysis – particularly multiphysics, the integration of which within Synopsys forms part of Toublanc’s role. “The real opportunity for a company like us is to make sure we can combine AI on one side, analytics that will serve AI, metamodel, and, at the end, a multiphysics sign-off to make sure that everything is going to be under control,” says Toublanc.

Toublanc sees agentic AI ‘profoundly changing the way engineers will work.’ The goal is for the agents to do the simpler tasks while the engineers can think about engineering processes, rather than execution processes. “Engineers won’t be replaced, but definitely the skill they will need to have, to be efficient and to be a better engineer versus another… will be having a global engineering mindset, having a global physics understanding, and having AI literacy understanding as well,” says Toublanc.

It also changes how the product is working. Engineers, Toublanc notes, previously wanted simple, quick answers, and ‘did not care about tonnes of data.’ Now, the remit s to explore, judge, and think about trade-offs – and the product must evolve. “This is part of new EDA challenges,” says Toublanc. “In the past, we were always trying to make the tool as fast as possible, consuming less memory as possible. But now, if we need more data, we have to review the way we develop the tool. We have to review the way we export data.

“So that is a profound impact from a product point of view, to make sure we can serve engineering properly in the future.”

The impact of AI can be examined in other ways as well. What will it mean for the relationship between the EDA companies, the chipmakers, and the system vendors to maintain the value chain?

Toublanc believes it is a big impact as AI is ‘democratising’ semiconductor, making it a profound help for a system house. “We see more and more needs where people would like to optimise the system for specific semiconductor, but as long as they do not have the knowledge, know-how of the semiconductor, they really rely on the semiconductor company they are working with – and they only know at the end of the day if what they’ve done is the optimum way,” he says.

“This trend is also even stronger, I would say, with the generation of multi-die semiconductors. The multi-die by itself is becoming a system, and the system house, the PCB company, knows very well that the heat, for instance, will be very complex to handle,” adds Toublanc. “They want to anticipate as early as possible the optimisation of the system to get the temperature of the semiconductor under control, while they want to make sure the semiconductor is designed for their specific purpose.

“So, what we start to see already, thanks to AI, thanks also to technology like meta-modelling that helps people to anticipate the physics, we start to see more and more system houses exploring semiconductors and system optimisation much earlier in the design process.”

At Microelectronics UK, on September 29-30, Toublanc will be speaking on balancing performance and efficiency in next-generation chip design – and he will be exploring and outlining those next steps.

“One of my goals is to explain our thinking and where we are already anticipating the next generation of tools,” says Toublanc. “The idea is to make sure we can take the best of multiphysics and agentic AI, specifically for multi-die. The industry is moving on, especially in Europe.”