A Paradigm Shift: Leveraging Artificial Intelligence And Machine Learning In Design Engineering

Matthew Yap, Sr. Manager Design Engineering, Microchip Technology

Matthew Yap, Sr. Manager Design Engineering, Microchip Technology

Reflecting on the past two decades, most IC design engineers spent most of their time in schematic capture and custom layout tweaking. Starting off as a young engineer, I was fascinated by the idea of computer-aided design, which eventually evolved into design automation, giving engineering scaling leverage. That led me to switch from schematic design and venture into the RTL2GDS domain. Today, I'm leading a central engineering team at Microchip Technology and seeing a similar paradigm shift in the form of AI/ML technology in design engineering.

As in typical central engineering functions, the demand for engineering resources has been a roller coaster ride. While it is easier to manage the peak and valley of EDA licenses and compute with peak-license EDA cards and cloud services, it is trickier when it comes to manpower. With the exponential scaling of Moore's Law, we are challenged to enable higher blocks-per-engineer productivity to catch up with the work demand. This is where I observed a similar paradigm shift two decades ago, whereby one engineer working on a small portion of a custom layout was enabled to take on a whole complex block with automatic place-and-route EDA technology. Within 20 years, the number of transistors on microchips has increased 1000-fold from 50M to 50B! And I foresee that it will continue to scale upwards, not just by process node scaling, but coupled with the rise of chiplets and heterogeneous integration.

Product development and innovation are a core foundation at Microchip, as demonstrated by one of our Guiding Values, “Continuous Improvement is Essential,” which drives our team to invest in emerging technology. When taking productivity gain as one of the objectives, we can break it down into three major components:

While It Is Easier To Manage The Peak And Valley Of Eda Licenses And Compute With Peak-License Eda Cards And Cloud Services, It Is Trickier When It Comes To Manpower

Productivity = schedule (day/experiment) + scalability (block/engineer) + efficiency (cost/engineer).

From pure runtime, each experiment does not differ much between the traditional versus AI/ML approach. However, we started to notice the effectiveness of AI/ML for babysitting tasks with zero downtime, which can terminate whenever the optimization does not make sense and before testing new experiments automatically leveraging reinforcement learning (RL) technology. That paved the way for scalability, in which an engineer can now work on multiple blocks while the AI/ML babysits most of the exploration runs. Efficiency at this infancy stage presents challenges, as the cost per engineer with the additional loading of AI/ML EDA license and more compute power could be higher. As in any emerging technology adoption, it is expected to have a window of paying the premium, and it is an investment I believe is worth spending to stay in the game.

Apart from putting a dollar value on AI/ML, one misconception that many may struggle with is the fear of diminishing the human role in IC design. History shows a similar sentiment, in the 1980s during the “computer revolution” people feared for their jobs and the same could be said with the “Internet revolution.” Today, most of us know that we leverage computers and the Internet to the fullest to get things done efficiently and effectively. At Microchip, another guiding value is “Employees Are Our Greatest Strength,” and the company will continue to invest in human capital and equip its engineers with the latest emerging technology. There is very little sign of diminishing human role, and in my opinion, we will leverage AI/ML down the road as we did with other emerging technologies that came before us.