Samsung uses AI to write chip-check scripts

Samsung, working with researchers from Seoul National University, has unveiled Rule2DRC, an AI agent designed to handle one of chip design’s most repetitive bottlenecks: turning written design rules into working design rule check (DRC) scripts.

That matters because chip teams must validate thousands of constraints before tape-out, and a move to a new process node often means rewriting large parts of those scripts almost from scratch. The problem gets worse when rules are written in natural language but need to be enforced in tools such as KLayout or SVRF.

Rule2DRC is aimed squarely at that gap. The system was built around 1,000 pairs of design rules and verification code, along with 13,921 test chip layouts. According to the developers, that lets them evaluate more than whether generated code merely resembles a reference answer. They can test whether an AI-generated script actually runs in a real verification engine and produces valid results.

The team also introduced SplitTester, a method for selecting the best option from several scripts generated by the model. Instead of rewarding polished-looking output, it ranks code based on how it behaves when executed. In EDA environments, that is the metric engineers actually care about: not elegant text, but a script that holds up on real layouts.

Recommended reading

6G signal breakthrough filters interference at the source

The project also includes a layout-native AI agent GUI. In that interface, an engineer can view the chip layout and see generated verification code in the same window. Samsung says the system has already been adapted to its internal environment, while integration with the company’s own language model is still underway. In demos, the agent parsed text instructions, selected relevant layout regions, and modified geometry on individual elements, including corners.

Samsung’s pitch is not that AI replaces engineers. The system translates textual requirements into executable scripts and quickly explores alternatives, while engineers and verification tools remain responsible for correctness. That is why the company is testing the setup on real engines rather than synthetic benchmarks.

The broader context is clear: Synopsys, Cadence, and Siemens EDA already dominate the EDA market and have spent years pushing automation in rules, verification, and supporting code. Samsung’s effort looks less like a lab demo than an attempt to cut expensive manual work in a part of semiconductor development where late mistakes can cost weeks and a lot of money.