How Agentic AI Chip Design Built a Full RISC-V Core

In 2020, researchers fine-tuned a GPT-2 model to design fragments of logic circuits; in 2023, researchers used GPT-4 to help design an 8-bit processor with a novel instruction set; by 2024, a variety of LLMs could design and test chips with basic functionality, like dice rolls (though often these were flawed).

Now Verkor.io, an AI chip design startup, claims a bigger milestone: a RISC-V CPU core designed entirely by an agentic AI system. The CPU, dubbed VerCore, has a clock speed of 1.5 gigahertz and performance similar to a 2011-era laptop CPU.

Suresh Krishna, cofounder at Verkor.io, says the team’s key claim is that this approach is more effective than using only specialized AI systems for specialized tasks within the overall design process. “ What we learned is that the better approach is to let the AI agent solve the whole problem,” he says.

Bringing Human Workflows to Agentic AI

Verkor.io’s agentic system is called Design Conductor, and it’s not itself an AI model. It’s a harness for large language models (LLMs). A harness is software that forces an AI agent to proceed through structured steps. In this case, the steps are like those a team of human chip architects would follow: design, implementation, testing, and so on. The harness also manages subagents and a database of related files.

That means it can work autonomously with only an initial prompt—in this case a 219-word design specification—from the user. (The prompt is published in the Design Conductor paper.) It outputs a Graphic Design System II (GDSII) file, which can be used in existing electronic design automation (EDA) software.

Synopsys and Cadence, two major players in EDA software, also have agentic AI tools. These allow chip architects to automate some tasks with AI agents. Design Conductor is different because it’s built to handle chip design from spec to completion with full autonomy, something major EDA companies have not yet touted.

Ravi Krishna, founding engineer at Verkor.io, says Design Conductor’s workflow is “mirrored after the traditional process a human engineer might use.” It analyzes the specification, then writes and debugs a register-transfer level, or RTL, file (an abstraction of the CPU’s data flow) before iterating through subtasks like power delivery, signal timings, and layout, which are again checked against the specification. Some tasks, like layout, call tools to assist the agent. “It’s an iterative system.”

The system took 12 hours to create the VerCore design. That’s not long, but, because it uses AI agents, you might imagine it taking more or less time based on the number of agents thrown at it. However, Ravi Krishna says it’s not that simple, because some design tasks aren’t easily parallelized.

However, the general improvement of AI models over time has proven essential. “I remember that around the middle of last year, we tried to build a floating-point multiplier with the models of that time. It was slightly…