Generative AI Will Be the Next Evolution in Embedded Design

Disruptive technologies, like artificial intelligence (AI), force us to change our behavior. As engineers, AI is reshaping the way we approach research, design, test, and verification. Generative AI, based on large language models (LLMs) and plain text prompts, has a lot to offer here.

Anyone not familiar with design methodologies may be surprised to learn that electronic design flows rely on the intermediate exchange of data. This exchange invariably uses a text format intended to be readable by humans.

This could be seen as serendipitous. Researchers working on integrated design would not have recognized the term large language model when hardware description languages (HDLs) were developed in the early 80s. Yet we can draw a straight line between that moment and today, where electronic design automation (EDA) companies now use LLMs in their HDL synthesis and design verification workflows.

LLMs are also being used to write embedded code. The syntax of structured programming languages makes AI extremely good at turning natural language prompts into functioning software.

Recent developments in AI include retrieval augmented generation (RAG) and agentic AI. With RAG, agents can refer to existing documentation to refine and improve their understanding of a prompt. These techniques have led to AI agents that can interface directly with integrated development environments (IDEs) to not only suggest code snippets but deliver complete builds.

AI Will Positively Impact the Design Cycle

A recent survey conducted by Avnet asked over 1,200 engineers across Asia, EMEA, and the Americas how they are using AI. The results show over 75 percent of respondents feel strongly that AI will positively impact the design cycle of new products and the automation of design tasks.

Figure 1. How confident are engineers about AI positively influencing the design process? (Image: Embracing Artificial Intelligence, Avnet Insights)

Iteration in design involves refining all or part of a solution to arrive at an optimum result. The measure of ‘optimum’ is dependent on the application and requirements set by the end market. Generative AI can assist engineers by automating design iteration; those engineers can then evaluate the results in the context of the application and requirements of the end market. Ultimately, using RAG, AI may be able to apply the context to refine the process and arrive at an optimum solution with less input required from an engineering team.

The current frenzy reflects that the AI industry is quickly maturing past adolescence. There are now numerous examples of how AI is being used in the design process. Breaking down the design cycle indicates the opportunities.

Using AI in PCB design

Table 1. Avnet’s impression of where AI is currently being developed to support the design process.

Looking at Table 1, component layout and routing show high potential because we know generative AI is already being used in these phases by some EDA software providers to augment design engineers. This will expand to include front-end design phases including detailed design and schematic capture.

Several smaller companies and startups are targeting frontend design using AI. The main challenge here is capturing the design intent accurately enough for a generative AI agent to interpret the requirements and synthesize a working design.

The ambitious goal would be to use natural language prompts to initiate the design process, resulting in a working design delivered as a full suite of design drawings ready for PCB manufacture.

While ultimately achievable, there are intermediate steps to reaching this goal. The process would involve training a generative AI model on existing designs for a specific application and asking the model to provide rapid design iteration for evaluation by an engineering team.

To prove the concept, Softweb Solutions created a pilot program with a key customer. Following a rigorous requirement gathering stage, the Softweb team understood the customer’s challenges and the limitations it faced in reaching its objectives. Using this understanding, the team examined the customer’s existing design processes, to identify how and where generative AI may be used. The solution was to use generative AI to rapidly deliver design variations, taking into account design constraints, performance requirements, and manufacturing restrictions ( Figure 2).

Figure 2. Using a generative AI model trained on the customer’s own designs, Softweb Solutions created a framework that enabled rapid design iteration in a format compatible with the EDA tool being used. (Image: Generative AI Solutions for Improved Hardware Schematics, Softweb Solutions)

The detailed process followed is shown below.

Data Flow and Communication with Eagle (Autodesk):

Input from Eagle:
- Eagle exports schematic, layout, and BoM in JSON/XML format (Figure 3).
- Data is ingested into a Component Repository (Vector Store) for structured indexing (Figure 2).
Processing by Gen AI Design Generator:
- The LLM (self-hosted, fine-tuned on electronic designs) processes input files.
- Natural language queries are converted into design elements.
- Corrected or optimized designs are generated with AI-driven insights.
Feedback to Eagle:
- The processed design is converted into a compatible Eagle format.
- JSON/XML-based output ensures easy import into Eagle for validation and further modifications.

Figure 3. Leveraging JSON to remove the barriers between EDA software and trained generative AI models. (Image: Softweb Solutions.)

The process involved training the generative AI model on existing PCB designs provided by the customer. For the pilot program, approximately 200 historical designs were ingested. Other training sources included open-source schematics and circuit layouts, along with industry-standard formatted bill of materials (BoM).

The fine-tuning stage used data processing to extract key patterns from the EDA information. Using this information, the PCB studio injected synthetic errors to train the AI model to recognize common design mistakes. By implementing machine learning operations (MLOps), the solution exercised continual learning based on a cycle of model training, validation, and deployment.

The AI-driven PCB design workflow can be summarized as:

Engineer inputs design requirements via Eagle or NLP interface.
Gen AI design generator processes inputs and suggests:
- component placements
- routing recommendations
- error corrections (AI-based validation)
PCB Studio and AI error correction refine the design.
Validated design is exported to Eagle for final adjustments.

The results from this pilot program are encouraging, indicating that the customer would get to market faster (42 percent) with improvements in design accuracy (81 percent) and design quality (95 percent).

AI and the Future of Design

The use of AI in the design process is seeing strong adoption from engineers of all disciplines and management levels. It can be proven to deliver a faster time to market and an overall improvement in design quality.

Building solutions that integrate into the existing tool flow is part of the challenge. However, professional engineering is incredibly structured and disciplined. This level of discipline already translates well into the EDA flow, making it potentially simpler to integrate AI.

The above case study provides a glimpse of how the data-driven engineer will work in the future, leveraging advancements in technology at each stage of the design process, improving efficiencies and outcomes.

This article was written by Alex Iuorio, Senior Vice President, Supplier Management and Business Development, Avnet (Phoenix, AZ). For more information please visit here .