AlphaDesign AI Experts Wade into Design and Verification

I mentioned in an earlier blog that multiple presentations at DVCon 2025 went all-in on AI-assisted design and verification. The presentation was one such example, looking very much at top-down AI-expert application of agentic flows to design and verification. AlphaDesign is a new startup out of UC Santa Barbara headed by William Wang (Professor of AI and a track record of research engagements with Amazon, Intel, Nvidia and others.)

The role of AI experts in advancing AI for design and verification

The promise of AI in this domain is both exciting and concerning. Exciting because there is potential to revolutionize productivity. Concerning not for loss of jobs (that will never happen), but because AI is still viewed as about approximate, probabilistic answers while engineering is about precision; approximate may be helpful for beginners and quick starts but not for production quality. We will still lean heavily on production tools (synthesis, simulation, etc) to validate and optimize, initially all the way through the flow, likely moving later in the flow as we build confidence in the quality of AI-based design generation.

Yet if we want to seize this opportunity and not talk ourselves out of big advances before we start, we need native AI experts to be involved in this journey as much as native EDA experts. EDA teams with their own AI experts will continue to push from the bottom up, very much with a focus on near-term profitability in optimizing proven flows, because that’s how they can run a healthy business. Top-down AI experts meanwhile can push what could be possible in generation and analysis from natural language prompts/specs, and beyond. That’s where I see AlphaDesign fitting in.

Certainly trust will need to be built along that journey, initially in helping refine verification suites for improved coverage. And perhaps in generating snippets of RTL as designers start to become comfortable with that idea. Later becoming a more accomplished aid in the design and verification process. We’ve been down similar paths before in EDA. I see what AlphaDesign is proposing an yet another improvement in productivity, initially helping tune current flows, gradually switching us over to new ways of thinking about the design task.

Agentic flows

The company calls their solution ChipAgents™, reflecting that the approach uses LLM agents to accomplish a goal. An agent in this world incorporates planning (decomposing a task into subtasks and refining past action), memory (managing context over a long period of time), and tool use (for assessment on a proposed solution and to elaborate designs/tests).

Agentic flows have been making big strides in the software engineering world. For automatically locating and fixing bugs in a software repository, LLM-only success rate is pretty sad (<3%). Adding basic agent support improved that rate by 8X. Amazon Q developer agent doubles that rate to 55%. A further refinement gets to 62%+. Not hands-free yet but an impressive advance.

Of course this is for software which can draw on a massive training corpus. Hardware is much more difficult, not because we have to be more clever but because there is so little of it to use in training (by one estimate, SystemVerilog code amounts to 0.28% of the lines of code accessible in Python+Java+Go+Javascript.) Also toolchains in hardware design compound complexity over software engineering.

Progress

This is an early-stage company, receiving their seed funding round in August 2024. Initial staffing has drawn from UCSB graduates with a heavy emphasis on AI and data science training.

The first step has been to build a serious reference design they call ChipAgentsBench, curated from OpenSource projects like OpenTitan. Good move, since many GenAI demonstrations for design that I have seen so far have been based on toy examples. This reference has 2.8k SystemVerilog files, amounting to over 600k lines of code. AlphaDesign say they plan to open-source a subset of this design at some point.

Details on demonstrated agent capabilities are thin so far. There is a CoverAgent aiming to help improve coverage in testing. As a general concept, using AI to help improve bottom-up coverage is not new. What looks intriguing here in talking to a couple of the R&D folks is looking at coverage based on reading natural language specs. As an example, finding ways to boost coverage in error-handling logic is always challenging in bottom-up testing but may be easier to spot/exercise based on reading a spec.

Unfortunately I missed their talk, thanks to conflicts, so take my limited understanding with a grain of salt. The company cites common use models in early engagements include generating DV documentation and code snippet generation for utility scripts. They also mention code summarization, RTL/testbench generation predicated on existing files and design verification IPs – all high value targets if/when proven.

Definitely a company to watch. You can check out the website HERE.

Also Read:

An Imaginative Approach to AI-based Design

Powering the Future: How Engineered Substrates and Material Innovation Drive the Semiconductor Revolution

The Double-Edged Sword of AI Processors: Batch Sizes, Token Rates, and the Hardware Hurdles in Large Language Model Processing