2025-09-12 Group Meeting Report

This report summarizes several recent pieces of work: improvements to RoRD-Layout-Recognation, the Geo-Layout-Transformer idea and code skeleton, notes from Siemens EDA Forum, and several automation scripts for improving server stability and usability.

Improvements To RoRD-Layout-Recognation

The main improvements in this phase focus on robustness and usability.

Improved Training Strategy

Training no longer requires pre-annotated datasets. The Trainer now automatically generates positive and negative samples during training and dynamically keeps the two groups balanced. This means raw data can be used directly for training.

This makes the training process closer to real usage and reduces the cost of early data preparation.

Improved Documentation

The README in the Git repository has been updated with more detailed setup and usage instructions.

A source link can be added later for the RoRD-Layout-Recognation repository.

Geo-Layout-Transformer

Geo-Layout-Transformer is a proposed large model for IC layout analysis based on a GNN-Transformer architecture. The goal is to borrow ideas from ViT, or Vision Transformer, while modeling IC layouts in a more structural way.

Technical Report

The technical concept and design have been written up as a report. A source link can be added later for the Geo-Layout-Transformer technical report.

Code Skeleton

The code skeleton for Geo-Layout-Transformer has been completed. The main to-dos are clarified, while concrete implementation details are currently represented by placeholder functions.

A source link can be added later for the Geo-Layout-Transformer repository.

Siemens EDA Forum

On August 28, I attended the Siemens EDA Forum in Shanghai. The forum was more commercial than technical, but it still provided useful insight into recent industry trends in AI-EDA, especially around yield improvement.

A detailed technical report link can be added later for Siemens EDA Forum.

Also, the refreshments were very good.

Server Scripts

To improve server stability and usability, I wrote several automation scripts.

NVIDIA Library Check And Reconnection Script

This script addresses the issue where nvidia-smi frequently fails because its version does not match the kernel driver version, specifically 575 versus 570.

The script checks the status of nvidia-smi every 10 minutes. If a problem is detected, it automatically reconnects to the correct 570 library version.

ZJUnet Auto Check And Reconnect Script

This script addresses intermittent ZJUnet dial-up disconnections on the server. It periodically tests network connectivity and automatically runs the reconnect command when the test fails.

Reconnect after network test failure	Network test success

MDView Script

MDView is a small tool that uses two mature terminal tools, glow and chafa, to browse Markdown files in the terminal. It can render headings, tables, lists, code blocks, images, and links.

Heading rendering	Table rendering

List rendering 1	List rendering 2

Image rendering	Link rendering

Summary

Recent work falls into two threads. One continues the AI-EDA research direction through RoRD and Geo-Layout-Transformer. The other improves the engineering environment through server automation and daily tooling. The former supports long-term research goals, while the latter improves the stability of the experimental environment and the efficiency of day-to-day development.