An Army of Agentic EEs Unleashed by Synopsys
Last month, I had the pleasure of attending Synopsys’ first Executive Forum in Santa Clara. For those who don’t follow Synopsys, the firm has a multi-decade history of providing foundational technology to the semiconductor industry, and they have not missed a beat under CEO Sassine Ghazi’s leadership. And while I knew that the leadership team would be sharing the latest innovations that would help fuel the next generation of processors, what Synopsys put together was stunning in terms of its long-term implications to semiconductor design.
Taking a step back, a processor has historically maintained a five-to-seven-year development cycle from vision to production, and over time market pressures have shrunk this development cycle significantly. According to Jensen Huang, who was running his own conference across town as the Synopsys event unfolded, we have now entered a realm of hyper-Moore’s law advancements with silicon design cycles within a one-year window. Unheard of, aggressive, and seamlessly impossible with yesterday’s tools.
The first advancement to address this challenge has been the introduction of chiplet architectures and the movement of different IP blocks, sometimes manufactured with different process technologies, connected together to form complex and scalable systems. This allows for re-design of chiplets across a portfolio of products or even across a portfolio of vendors for design elements that are not critical to differentiated solution delivery. Synopsys itself operates in this space, delivering IP and multi-die integration solutions to the semiconductor ecosystem. We’ve seen the industry move in this direction with all major players favoring chiplet architectures, and while we’re still waiting for a vibrant chiplet ecosystem to emerge for open, interoperable designs, we are seeing an uptick in adoption of chiplet designs across the industry delivered to bespoke customers.
But what Sassine and his team envisioned next rips the doors off of semiconductor design to the point of considering that this is a next era of advancement: agentic AI controlled design. Imagine deploying millions of AI agents that a semiconductor designer can utilize to craft semiconductors. This starts with something Synopsys has already delivered – an AI Copilot that assists a designer in the process of his or her design, trained specifically on an organization’s technology and sitting aside other Synopsys tools and IP. This is a great advancement and a good use of LLM training for a pressured workforce. But…it’s only an amuse bouche of what is to come.
We are on the precipice of what Sassine called level 2 implementation of a journey taking us from copilot to auto-pilot design. Synopsys will be delivering agentic capability in their portfolio later this year to offload design of specific functions to agents, with human controlled integration of this design into larger workflows. Synopsys sees a reality where a portion of design is delivered through chiplet and IP blocks, a portion is delivered by agents, and actual engineers focus on the most challenging elements of execution, speeding design significantly.
This will be followed by level 3 delivery, where multi-agent models will include specialists orchestrated together for more complex work, followed by the integration of adaptive learning improving designs as they progress, and finally level 5 full autonomy delivery with agents empowered with decision making. Yes…in level 5, agents in theory can design and validate microprocessors autonomously.
This is stunning. It’s stunning for the acceleration of compute, given that we can also advance process technology and continue delivery of performance improvements, but it’s stunning for another reason. Microprocessors are the most complex inventions on the planet, and Synopsys’ vision includes processors of a trillion transistors, reaching new levels of design complexity. If AI agents can deliver this complex of problem solving, we need to ask what they will be incapable of doing.
While we do not have an acute time horizon for this advancement, one thing is clear. Just like the accelerated microprocessor design windows, the speed in which generative AI and agentic computing has been advancing makes this reality much closer than we’d historically predict. And while this story feels a lot like the backstory of Skynet, in the near term there is no escaping that AI agent assist will help Jensen and his team at NVIDIA, and the entire silicon ecosystem, accelerate design to meet the insatiable customer demand represented by AI. AI…creating AI.
Last month, I had the pleasure of attending Synopsys’ first Executive Forum in Santa Clara. For those who don’t follow Synopsys, the firm has a multi-decade history of providing foundational technology to the semiconductor industry, and they have not missed a beat under CEO Sassine Ghazi’s leadership. And while I knew that the leadership team would be sharing the latest innovations that would help fuel the next generation of processors, what Synopsys put together was stunning in terms of its long-term implications to semiconductor design.
Taking a step back, a processor has historically maintained a five-to-seven-year development cycle from vision to production, and over time market pressures have shrunk this development cycle significantly. According to Jensen Huang, who was running his own conference across town as the Synopsys event unfolded, we have now entered a realm of hyper-Moore’s law advancements with silicon design cycles within a one-year window. Unheard of, aggressive, and seamlessly impossible with yesterday’s tools.
The first advancement to address this challenge has been the introduction of chiplet architectures and the movement of different IP blocks, sometimes manufactured with different process technologies, connected together to form complex and scalable systems. This allows for re-design of chiplets across a portfolio of products or even across a portfolio of vendors for design elements that are not critical to differentiated solution delivery. Synopsys itself operates in this space, delivering IP and multi-die integration solutions to the semiconductor ecosystem. We’ve seen the industry move in this direction with all major players favoring chiplet architectures, and while we’re still waiting for a vibrant chiplet ecosystem to emerge for open, interoperable designs, we are seeing an uptick in adoption of chiplet designs across the industry delivered to bespoke customers.
But what Sassine and his team envisioned next rips the doors off of semiconductor design to the point of considering that this is a next era of advancement: agentic AI controlled design. Imagine deploying millions of AI agents that a semiconductor designer can utilize to craft semiconductors. This starts with something Synopsys has already delivered – an AI Copilot that assists a designer in the process of his or her design, trained specifically on an organization’s technology and sitting aside other Synopsys tools and IP. This is a great advancement and a good use of LLM training for a pressured workforce. But…it’s only an amuse bouche of what is to come.
We are on the precipice of what Sassine called level 2 implementation of a journey taking us from copilot to auto-pilot design. Synopsys will be delivering agentic capability in their portfolio later this year to offload design of specific functions to agents, with human controlled integration of this design into larger workflows. Synopsys sees a reality where a portion of design is delivered through chiplet and IP blocks, a portion is delivered by agents, and actual engineers focus on the most challenging elements of execution, speeding design significantly.
This will be followed by level 3 delivery, where multi-agent models will include specialists orchestrated together for more complex work, followed by the integration of adaptive learning improving designs as they progress, and finally level 5 full autonomy delivery with agents empowered with decision making. Yes…in level 5, agents in theory can design and validate microprocessors autonomously.
This is stunning. It’s stunning for the acceleration of compute, given that we can also advance process technology and continue delivery of performance improvements, but it’s stunning for another reason. Microprocessors are the most complex inventions on the planet, and Synopsys’ vision includes processors of a trillion transistors, reaching new levels of design complexity. If AI agents can deliver this complex of problem solving, we need to ask what they will be incapable of doing.
While we do not have an acute time horizon for this advancement, one thing is clear. Just like the accelerated microprocessor design windows, the speed in which generative AI and agentic computing has been advancing makes this reality much closer than we’d historically predict. And while this story feels a lot like the backstory of Skynet, in the near term there is no escaping that AI agent assist will help Jensen and his team at NVIDIA, and the entire silicon ecosystem, accelerate design to meet the insatiable customer demand represented by AI. AI…creating AI.
.png)
