DAC 2026: How Undo Makes AI for Chip Design Work Better

DAC was abuzz this year with new AI-fueled approaches to chip design and debug. New tools powered by deeply trained models opened substantial possibilities to deliver higher quality designs much faster, potentially with less resources. Every new technology has its risks. For AI-driven chip design some of those risks center around a class of issues associated with AI losing its focus and context. The designs in front of us are incredibly complex so this isn’t hard to imagine.

Manifestations of these issues include model hallucinations or simply incorrect behavior. All problems have solutions, and Undo has focused on this class of problem. The company showcased technology that up-levels the performance of AI by adding detailed context about the design. This kind of information can be the margin of victory for a design using new AI tools. Let’s look a bit closer at how Undo makes AI for chip design work better at DAC 2026.

What Undo Delivers

I had the opportunity to meet with Dr. Greg Law, CEO of Undo at DAC.  Greg is a systems engineer at heart. He has spent over 25 years building and leading software teams working on complex, high-performance codebases where failures are costly and debugging is a bottleneck. His career spans pioneering computer company Acorn, fast-growing startups including NexWave and Solarflare, and ultimately the creation of Undo. Greg holds a Ph.D. from City University, London, and lives in Cambridge, UK.

Greg explained that Undo enables AI to understand complex runtime behavior to deliver accurate agentic root cause analysis. The main focus is on C/C++ and SystemC and the impact is an up-level in the performance of AI. By giving coding agents the precise execution history of a program (that is, high-quality runtime context), Undo enables evidence-based automated root-cause analysis up to 100x faster.

He went on to explain that Undo’s deterministic technology records a program’s exact execution once: every variable, every event, every I/O, across every module. Agents then query the recording to reason about dynamic behavior the way they already reason about code. As a

result, rare, intermittent and complex problems become easy to isolate and explain.

Looking closer, for AI this context is quite valuable. It transforms an agent from a “clever guesser” into a reliable problem-solver. With Undo, agents can explain not just what went wrong but why, backed by a recording of what actually happened, thanks to Undo’s technology.

Stepping back a bit, we can look at coding agents such as Cursor, Claude and Copilot that are being adopted across engineering teams. These teams often question whether they can scale to large and complex codebases. There is a concern about how much hallucination (returning results not grounded in fact) limits the benefits.

The problem is you can’t truly see what happened at runtime. This makes it hard to reason about cause and effect. Logs help, and AI is good at reading them. The problem is logs only show what was predicted to be of interest. If you already knew which data would expose the problem, you’d have prevented it in the first place.

Coding agents get stuck in the same human loop of re-instrumenting and rerunning. This is done faster, but it is just as wasteful. This isn’t a gap in the model’s intelligence; it’s a gap in its context. The hardest problems in complex software live in runtime behavior: execution paths, state changes, and values flowing across module boundaries that source code and logs never reveal. Models keep getting better, but context is now the bottleneck.

This is the problem Undo solves. After my discussion with Greg, I started to see the potential impact across many AI applications.

The Impact Undo Delivers

We discussed two use cases to illustrate the impact of Undo’s approach.

1) System Design Architects. The context here is GPU model testing (test workload performance validation). Architects are running test workloads on GPU models. The problem is the activity is constrained by limited simulation coverage. Functional bugs prevent enough applications from running on top of the model (which would enable them to define the best parameters / architecture). So, coverage may only reach 50 – 60% of meaningful use cases. The inability to perform more tests impacts the final quality of the chip.

The solution is to provide runtime context. This is a major enabler to agentic exploration of architectures. Architects can quickly diagnose and resolve functional defects using design-specific runtime context (the complete picture of what actually happened when the program was executing). Architects can then evaluate more scenarios and optimize the chip architecture across more workloads, and deliver higher quality chips

2) System Level Verification. According to the Wilson Research Group, up to 70% of design investment goes into verification. And roughly 50% of that time goes into debug. So, about 35% of total silicon design effort goes into debug. Root cause analysis is challenging since bugs can be anywhere (the test bench, fast models, the synthetic software load, or RTL hardware design). Bugs can be in any of these domains, including in the interaction between these components. And large SystemC and C++ simulation models are often opaque and multi-threaded, making them especially hard and time-consuming to debug.

The opportunity is to significantly increase design productivity. Agentic workflows struggle with complex codebases – AI hallucination is commonplace. Without deep runtime context from the live design, AI cannot reliably perform root-cause analysis for complex verification issues.

System verification teams can reduce debug time by giving AI agents the runtime context they need to understand complex system behavior and autonomously identify root causes. Runtime context from the live design is far more valuable than static training data or generic LLM intelligence.

In both use cases above, Undo provides the needed runtime context – and you can’t get that data any other way.

To Learn More

If you want to up-level the performance of your AI tools, Undo is a company you should know. You can learn more about them on SemiWiki here. And you can probe deeper, including a demo request on Undo’s website here.  And that’s how Undo makes AI for chip design work better at DAC 2026.

Also Read:

The Other Side of Bug Localization

Enabling AI to Understand Complex Runtime Behavior for Accurate, Automated Root Cause Analysis — DAC 2026

Revolutionizing Hardware Design Debugging with Time Travel Technology