AI Agents Are Pushing Chip Design Earlier, but Verification Still Sets the Pace
- Ed Sperling
- 346 reads
Digital twins and AI agents expand early design exploration, while model accuracy, formal proof and changing workloads keep verification central.

AI agents and digital twins are widening the range of chip design choices engineers can examine early, from transistor structures and materials to architecture and implementation. The payoff depends on whether the models represent real hardware accurately and whether teams can verify the designs that agents produce. For now, earlier exploration shifts bottlenecks as much as it removes them.
That was a central theme in a discussion among 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; and Ashish Darbari, founder and CEO of Axiomise. Their views point to a shift in what “shift left” means: design work can begin earlier, but decisions may also need to stay open as engineers learn more.

From left: Jack Berg of Silvaco, Frank Schirrmeister of Synopsys, Ashish Darbari of Axiomise and Ronen Laviv of Hiveware.
More physics enters early design
Digital twins are moving analysis further upstream, but the models behind them need to capture more than digital behavior. Berg pointed to the growing role of multi-physics modeling as chip technologies introduce new materials and device structures. Designers may depend on materials science and physics solvers to create surrogate models they can use in design workflows. That approach is reaching areas such as complementary field-effect transistors (CFETs), where physical behavior and material choices shape the design space.
Compute capacity and software agents now make it possible to explore options much earlier than design teams could previously manage. Schirrmeister sees that expanded search as a major change, with a caveat: an agent cannot establish on its own that a surrogate model accurately represents the hardware. Someone must take responsibility for validating the model. If it fails to predict how the eventual design behaves, early decisions can send the project toward an expensive redesign or another silicon spin.
There’s a related gap between generating options and reasoning about their consequences. Agents can help explore alternatives, but engineers still need to judge whether a proposed architecture or implementation serves the product. The more choices teams can produce, the more important it becomes to understand what those choices mean for performance, power, functionality and verification.
Formal verification provides a trust check
Darbari said interest in AI agents has pushed more discussion toward moving from specifications to silicon. In principle, software could extract requirements from specification documents and help turn them into a design. In practice, teams may not have complete specifications, and extracting requirements does not prove that the resulting hardware behaves as intended.
Some companies are asking agents to rewrite existing designs to improve power, performance and area (PPA). That makes equivalence checking important: engineers need a way to establish that an optimized version preserves the required function. Darbari described formal proofs as one means of addressing that trust gap. He also sees functional verification discussed alongside security, functional safety and power concerns as hardware becomes part of AI systems.
Formal verification can help establish properties or equivalence, but adopting it early still depends on how well tools fit into the larger engineering environment. Darbari identified integration as a practical obstacle. The work also creates a review burden: agents can produce large volumes of output, and teams need useful ways to find the information that affects a decision. If every result requires extensive manual inspection, review itself can slow adoption.
Laviv framed the broader problem as a continuing search for bottlenecks. Moving a task earlier in the flow can expose a new constraint or make an old one more visible. Agents may accelerate parts of the process, but the gains depend on whether teams can keep analysis, verification and review moving together.
Design choices may need to remain adjustable
Traditional shift-left thinking assumes that teams can define the target early and then work steadily toward it. Schirrmeister questioned whether that assumption fits products whose workloads and implementation needs may change before deployment. Instead of fixing every detail far in advance, teams could start with a direction, build models and refine choices as new information arrives.
That could affect decisions such as whether a product needs an Arm processor or a RISC-V design. A team might discover through early modeling that its initial assumptions about the hardware were wrong, then revise the specification. Schirrmeister described this as a more iterative form of early design: engineers explore implementation choices sooner, while allowing later evidence to reshape them.
Such flexibility matters when the workload a chip will serve is uncertain. In a more sequential flow, product assumptions can take many months to meet implementation results. With ongoing data and model feedback, design teams may have more opportunities to adjust before the architecture is fixed. The challenge is reconciling those changes across the design and verification flow without losing control of which version has been checked.
Darbari agreed that many implementation details remain unknown at the outset, especially for machine learning hardware, where training and inference workloads bring different demands. But he cautioned that the architecture is not necessarily a blank slate: teams building a CPU, GPU or machine learning chip often have a prior design to inform the next generation. The uncertainty lies in choices around power and functionality, and in whether agent-generated alternatives can be evaluated and trusted quickly.
Agents need engineers in the loop
Berg described agents as junior engineers whose work needs review by people with the right expertise. They can help explore an inverse design problem: given constraints, find a transistor, chip or RTL implementation that meets them. Yet a set of constraints may be incompatible, leaving no feasible solution. In that case, the design team must decide which requirements can change and assess the trade-offs.
That is one reason the ability to revise a design matters. Digital twins and multi-physics solvers can help teams examine candidate solutions, but they can also reveal that the target is over-constrained. An agent can identify or generate options; engineering judgment is still needed to decide what to relax and how the resulting choice affects the product.
Laviv outlined a digital design flow that may include virtual platforms, formal methods, software simulation, emulation and FPGA prototypes, alongside engineers and AI engines. If teams resolve bottlenecks in formal analysis, simulation, virtual prototyping or the flow connecting them, they may be able to start later verification activities sooner. Earlier access to those results can support better-informed design decisions and make course corrections less disruptive.
Darbari sees validation extending all the way back to architecture. Some teams are revisiting architecture from the ground up to determine whether a design can support the range of intended uses. That assessment becomes harder when software behavior, training, inference and power requirements all feed into early decisions.
The practical measure of progress, then, is not how far left a tool can move a task. It is whether the models predict real behavior, whether teams can prove the design meets its requirements, and whether verification can keep pace with the alternatives agents generate. Early exploration is becoming more powerful; making its results dependable remains the work that will determine how much faster chip teams can move.