About this Event
Analog/mixed-signal (AMS) integrated circuits (ICs) play an essential role in electronic systems by bridging the analog physical world and the digital information world. Their ubiquitousness powers diverse applications ranging from smart devices and autonomous cars to crucial infrastructures. Despite such critical importance, conventional design strategies of AMS circuits still follow an expensive and time-consuming manual process and are unable to meet the exponentially-growing productivity demands from industry and satisfy the rapidly-changing design specifications from many emerging applications. Design automation of AMS IC is thus the key to tackle these challenges and has been a longstanding open research question.
Towards this goal, my dissertation explores various machine learning (ML) techniques to automate AMS design and infuse learning abilities in these circuits to meet diverse application requirements. First, I will present our design framework to use neural approximation--a type of supervised learning (SL) method to learn analog-to-digital converters (ADCs), one of the most essential AMS circuits, by training rather than design by hand. The method not only achieves an automated synthesis flow for various ADCs but also enables the designed ADCs of intelligence to flexibly learn different quantization schemes to accommodate miscellaneous applications. Our results demonstrate superior performances of learned ADCs that exceed the limits of conventional ADCs. Next, I will propose a strategy to generalize the neural approximation method for arbitrary AMS circuits and apply them to the system-level designs. As a case study, we leverage the method to transform the design of AMS peripheral circuits in in-memory computing (IMC) architectures that lead to much-improved computing performance and energy efficiency. Finally, I will introduce how reinforcement learning (RL) techniques are exploited to automate the design of AMS and radio-frequency (RF) circuits. Unlike prior arts, we train our RL agent with sufficient domain knowledge of AMS circuit design, thereby beating all existing methods and experienced human designers. Particularly, we apply our RL method to solve the two most challenging problems of AMS circuits at the schematic level: 1) finding device parameters to meet given circuit specifications, and 2) performing design space exploration to achieve the best figure-of-merit (FoM) of a given circuit.
In conclusion, our exploration to the automated design of AMS and RF ICs demonstrates ML techniques' great potential to bring us closer to a future where circuit designers can be assisted by various learning methods to achieve faster and more efficient design of high-performance IC products.
Event Details
See Who Is Interested
0 people are interested in this event
User Activity
No recent activity
About this Event
Analog/mixed-signal (AMS) integrated circuits (ICs) play an essential role in electronic systems by bridging the analog physical world and the digital information world. Their ubiquitousness powers diverse applications ranging from smart devices and autonomous cars to crucial infrastructures. Despite such critical importance, conventional design strategies of AMS circuits still follow an expensive and time-consuming manual process and are unable to meet the exponentially-growing productivity demands from industry and satisfy the rapidly-changing design specifications from many emerging applications. Design automation of AMS IC is thus the key to tackle these challenges and has been a longstanding open research question.
Towards this goal, my dissertation explores various machine learning (ML) techniques to automate AMS design and infuse learning abilities in these circuits to meet diverse application requirements. First, I will present our design framework to use neural approximation--a type of supervised learning (SL) method to learn analog-to-digital converters (ADCs), one of the most essential AMS circuits, by training rather than design by hand. The method not only achieves an automated synthesis flow for various ADCs but also enables the designed ADCs of intelligence to flexibly learn different quantization schemes to accommodate miscellaneous applications. Our results demonstrate superior performances of learned ADCs that exceed the limits of conventional ADCs. Next, I will propose a strategy to generalize the neural approximation method for arbitrary AMS circuits and apply them to the system-level designs. As a case study, we leverage the method to transform the design of AMS peripheral circuits in in-memory computing (IMC) architectures that lead to much-improved computing performance and energy efficiency. Finally, I will introduce how reinforcement learning (RL) techniques are exploited to automate the design of AMS and radio-frequency (RF) circuits. Unlike prior arts, we train our RL agent with sufficient domain knowledge of AMS circuit design, thereby beating all existing methods and experienced human designers. Particularly, we apply our RL method to solve the two most challenging problems of AMS circuits at the schematic level: 1) finding device parameters to meet given circuit specifications, and 2) performing design space exploration to achieve the best figure-of-merit (FoM) of a given circuit.
In conclusion, our exploration to the automated design of AMS and RF ICs demonstrates ML techniques' great potential to bring us closer to a future where circuit designers can be assisted by various learning methods to achieve faster and more efficient design of high-performance IC products.