Coming from AI/CS? A Starter Guide

You do not need a hardware background to start. Here is how I would suggest getting into chip-design research from AI or CS.

You do not need to already know EDA (and please do not be scared by these terms!). Some of my favorite students are people who arrive from AI, algorithms, systems, or architecture, then discover that chips are an absurdly good playground for their taste (including me: my undergraduate major was computer science, and I knew almost nothing about EDA or testing before I started research). TBH, I even believe chips are one of the most promising directions from a very practical point of view (great industry jobs, faculty opportunities, and, if we are lucky and stubborn enough, a real chance to leave your name somewhere in the history of computing). What matters most is curiosity, discipline, and the courage to work on problems that are larger than one person.

I Come from AI/CS and Have No Hardware Background. Can I Really Do This?

Yes. Your concern is reasonable, but I do not think the lack of a traditional hardware background is a major obstacle for the kind of research I want to do. At the beginning, your background in AI, algorithms, or computer science may actually be more important. Many chip-design problems can be abstracted as optimization, graph, generative-modeling, reinforcement-learning, or agentic decision-making problems.

You will need to learn enough domain knowledge to understand the problem formulation, constraints, and evaluation. However, I would not expect you to become an expert in circuits or hardware design before you can contribute. The more effective approach is to learn the relevant hardware concepts alongside a concrete research problem.

A simple roadmap I would suggest:

  1. Start with one specific chip-design problem, not all of VLSI. Placement is a good example: learn what cells, macros, and nets are; what wirelength, congestion, and timing mean; and what makes a placement “good.”
  2. Study how different AI methods enter the problem. Compare reinforcement-learning-based placement, learning-enabled differentiable placement, diffusion-based placement, and more recent LLM- or agent-based approaches. Try to understand not only what model was used, but how the original engineering problem was reformulated.
  3. Think critically from the AI perspective. Why does a method work here? What assumptions does it make? Where does it fail? Most importantly, does this chip-design problem require genuinely new AI methodology, rather than simply applying an existing model to a new dataset?

A short reading path:

  1. Chip Placement with Deep Reinforcement Learning: an early representative example of applying RL to chip placement.
  2. DREAMPlace: useful for seeing how a classical placement problem can become differentiable optimization accelerated by modern deep-learning toolkits.
  3. Chip Placement with Diffusion: an example of using a modern generative model to place macros simultaneously.
  4. Order Matters: Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution: a recent example of bringing LLM-based search and reasoning into placement.
  5. ChipNeMo: a representative industrial example of domain-adapted LLMs for chip-design tasks.

You do not need to understand every technical detail at this stage. Read these works with the help of AI tools and ask yourself a simpler question: Do I find these problems interesting enough to understand why existing AI methods are insufficient, and do I want to develop better ones? If the answer is yes, your current background is sufficient to get started. The domain-specific knowledge can be learned gradually during the Ph.D.