At the recent Design Automation Conference, experts discussed the development of multi-agent systems in chip design. Matt Graham, Senior Director of the Software Verification Product Management Group at Cadence, explained that the architecture of multi-agent systems is evolving towards a horizontal organization: “It’s becoming much flatter than an army. If you asked me a year ago, I would have said, yes, it’s a large hierarchy of agents. But we are moving towards a flat set of skills, capabilities, functions. All of this is organized into dynamic agents working depending on the task.” Alexander Petr, Senior Director and Portfolio Manager at Keysight EDA, offered a comparison with an orchestra: “It’s better to compare it to an orchestra. You have one conductor and groups of musicians playing certain instruments. They can play alone or together, synchronously or asynchronously to achieve certain melodies, and then play together again.” Satyash Balasubramanian, Head of Product for EDA AI at Siemens EDA, added: “Even if the system is flat, the key – is that everyone knows their role and how they are connected to everyone else. Everyone knows the ultimate goal. Nothing is isolated.” Balasubramanian noted the growing popularity of so-called “headless agents” without a single main agent controlling all actions. “It’s popular in software and is coming to hardware. We are building agent capabilities from physical, deterministic engines, and it will be an agent on its own. It can talk to other agents, understand different formats and designs.”
Anand Thiruvengadam, Executive Director and Head of AI Product Management at Synopsys, emphasized the importance of flexibility in such systems: “It will not be a rigid structure. Spill and flow will occur, it can change depending on the complexity of the commands. Fluidity is very important. You will have a layer that organizes and calls resources as needed to execute.” Harrison Balisteri, Head of Business Development and Strategic Partnerships at ChipAgents, stressed the importance of context management: “If all agents develop the same hypotheses for a problem, it doesn’t matter how many agents you deploy or how good your judging agent is. It’s a lot in terms of context management and local versus global context.” The speakers discussed the issue of accuracy in chip design. Balasubramanian clarified: “In chip design, 99.99% accuracy – is a failure. All it takes is one picosecond, and you have a whole violation.” Petr added that deterministic, mathematically optimized approaches remain critical: “We have spent decades optimizing modeling engines for deterministic results. And my mathematically based optimizer is still faster and more accurate.” Thiruvengadam pointed out the problem of the “march of nines”: to achieve 100% accuracy when replacing deterministic tools with probabilistic models, it is necessary to overcome the threshold of 99.99999% for hardware.
The main challenge – is controlling token consumption and intellectual property protection. Graham noted the need for protective barriers in the infrastructure: “It’s part of the preparation before deployment. In IC design, there are infinitely many problems, so there must be time, token, and tool coverage barriers.” According to Balasubramanian, IP rights – are a serious issue: “IP rights – are one of the biggest issues I hear about. All it takes is one protected piece of code getting into your codebase, and if you don’t check it, you will incur losses.” He also pointed out that every prompt – is intellectual property: “Any prompt you give – is your intellect. You are telling people the algorithm for solving certain problems.” Petr added a legal aspect: “U.S. copyright law does not allow us to copyright AI-generated code. This means that every chip designed by AI is not protected by copyright.” Balisteri offered a solution: using your own model deployed locally. “If the design never leaves your environment, it is de facto still yours. It is not copyrighted as such, but no one else can access it.” He also noted that the ChipAgents model uses 40% fewer tokens when performing the same tasks, which solves the cost problem of deploying agent systems in chip design.
Related articles
