AI agents in EDA will augment, not substitute, human expertise. For example, AI agents will assist with repetitive tool setups using natural language and provide a productivity boost, but will rely on chip designers\u2019 strategic expertise in following a specific chip design approach. The roadmap envisions a future with a collaborative human-agent ecosystem for addressing complex design challenges. Dan Yu, an AI & ML leader at Siemens EDA, explained: \u201cAI optimizes and accelerates EDA workflows. Think of the human EDA engineer as the pilot, making strategic decisions, while AI acts as the co-pilot, boosting productivity and accelerating time to market.\u201d<\/p>\n\n\n\n
EDA data is both the foundation and a major challenge for AI implementation<\/strong><\/p>\n\n\n\n Data is fundamental for AI in EDA, especially for fine-tuning models or building Retrieval-Augmented Generation (RAG) solutions. However, obtaining well-organized, high-quality EDA data (e.g., RTL\/Verilog, sample PDKs, etc.) is a challenging task. Furthermore, EDA data typically remains siloed and underutilized and needs to be aggregated in centralized, multimodal EDA data lakes to unlock a strong data flywheel effect for uncovering critical insights across the entire EDA workflow. At the same time, data security and intellectual property (IP) protection pose significant challenges due to the sensitive nature of design information, necessitating robust security measures from the outset.<\/p>\n\n\n\n Compute infrastructure is a key enabler for the EDA AI transformation<\/strong><\/p>\n\n\n\n Robust cloud and data center infrastructure is of paramount importance not just for model training or fine-tuning, which may occur occasionally, but it is critical for smooth and continuous inference, which happens daily. Specifically, inference is getting more importance, as reasoning models and agentic systems typically require 10x-20x more compute. Thus, semiconductor companies must make a strategic choice between on-premises capital expenditures and cloud computing, or embrace a hybrid solution. To avoid huge costs of constant model retraining, and the ensure you are using the best-in-class models given the rapid AI advancements, Jeff Dyck, Senior Director of R&D at Siemens EDA, suggested: \u201cRAG is a viable methodology. RAG offers flexibility in utilizing the best available models without the need to retrain for every new model released almost monthly. Furthermore, RAG allows enabling an access control layer to restrict RAG operations to authorized documents, thus providing a structured pathway for controlled access to proprietary data.\u201d<\/p>\n\n\n\n Embrace AI by moving fast while not breaking things<\/strong><\/p>\n\n\n\n Panelists emphasized the importance of proactive AI experimentation, encouraging the implementation of solutions to identify strengths, weaknesses, and areas for iterative improvement. One of the main challenges is reconciling the AI development mindset of \u2018moving fast and breaking things\u2019 with the EDA industry\u2019s principle of \u2018do not break things.\u2019 The goal is to \u2018move fast but not break things\u2019, and striking this balance means fostering a culture of rapid progress without compromising reliability.<\/p>\n\n\n\n
![\"EDA](\"https:\/\/blogs.sw.siemens.com\/wp-content\/uploads\/sites\/58\/2025\/06\/Panel-photo-for-blog-1-scaled.jpg\")