AI restructures Siemens EDA tools
Cette publication existe aussi en Français
Siemens EDA is restructuring the structure of its EDA tools to adopt a new architecture for artificial intelligence (AI) and machine learning (ML).
The Questa, Calibre and Solido EDA tools have relied on individual complex databases to hold information about custom chips, chiplets and 3D-IC substrates and even chips on a printed circuit board designs. Now those databases are being opened up into a data lake so that AI engines can access the information across the different tools.
This is a fundamental shift in the tool architecture that allows the creation of digital twins of the chip and package.
“What we are all seeing is software driving functionality. That pervasive growth is enabled by silicon and the amount of compute is driving an explosion in the need for compute. Layered on top of that is AI,” said Michael White, senior engineering technologist at Siemens EDA
“We are seeing this in hyperscalers first, then autonomous cars and consumer electronics. This need is driving investments across Siemens EDA with architectural exploration, functional verification, advanced node design and physical verification,” he said.
“AI is integral to all of this, not only in the use of the system but as part of the products to enrich the designer and make them more productive, using AI in many of the engines. It is pervasive in Siemens EDA to make the IC community far more productive,” he added.
The AI is a combination of several techniques such as Reinforcement learning (RL) and rule-based machine learning. Retrieval Augmented Generation (RAG) uses transformer models to enhance the user interface using existing training material
“AI can boost productive by 50%, especially at 2nm but this is industrial grade AI vs consumer AI,” said Sathish Balasubramian, head of product for verification and AI at Siemens EDA.
“This means Verifiability, usability, robustness and generality and not just certain corner cases, but the most important thing is never to compromise on accuracy. We have been doing AI for the last 20 years at Siemens EDA across all the product lines, with pattern recognition in 2005, reinforcement learning to tweak models, and now there is genAI moving to AI agents to automate parts of the workflow and then to agentic AI and using all of this for digital twins,” he said,
“What we are focussed on is ML and RL at the backend with GenAI and AgenticAI at the front end in a hybrid AI system and we are launching an agentic AI system across silicon and PCB tools. This enables generative and agentic AI across all the Siemens Eda tools with a multimodal data lake and APIs for customers to tap into the system.”
Whether the database remains separate depends on the customer choice. They can integrate their own EDA data and create custom workflows to deploy AI where it adds the most value – enhancing adoption and competitiveness without disrupting workflows. This can be on-premises or in the cloud, and a ‘data flywheel’ using the centralized multimodal data lake.
“This is a combination of the databases and data lake,” he said. “We use RAG a lot even though the databases are separate. The separation depends on the different customers to open up RTL database into the data lake.”
As well as in-house infrastructure and third-party models, the AI system also supports Nvidia’s NIMS microservice and Llama Nemotron models, NIM enables scalable deployment of inference-ready models across cloud and on-premises environments for real-time tool orchestration and multi-agent systems. Llama Nemotron adds high context reasoning and robust tool-calling for more intelligent automation across the EDA workflow.
“AI agents can dramatically boost productivity for complex electronic design automation to support engineers across layout optimization, simulation and verification, freeing engineers to focus on creative problem-solving and advanced design challenges,” said Tim Costa, senior director of CAE and CUDA-X at Nvidia.
The latest version of the tools is being launched at the Design Automation Conference in the US this week, with early access to key customers.
These include Aprisa AI for RTL-to-GDS design exploration that adaptively optimizes for power / performance / area (PPA) for a given design, as well as integrated generative AI-assist, delivering ready-to-run examples and solutions. A natural language interface provides 10x productivity, 3x faster time to tapeout and 10 percent better power, performance and area optimisation for digital designs across all process technologies, says the company.
Calibre Vision AI identifies and fixes critical design violations in half the time of existing methods by loading and organizing them into intelligent clusters. Designers can then prioritize their activity based on this clustering and achieve a higher level of productivity.
Generative and agentic AI is used throughout the Solido Custom IC platform, including schematic capture, simulation, variation-aware design and verification, library characterization, layout and IP validation.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
