What is it about?

Integrated Circuits (Chips) are getting increasingly complex. The circuits thereon can be categorized as either Digital or Analog. Where Analog circuits are the ones dealing with continuous signals and as such require special attention when designing, since they are more susceptible to interference. Part of this design process is sizing the devices (transistors) in the circuit. Usually a designer will manipulate the geometric dimensions of all devices until the overall circuit performance meets a given specification. This is done manually and predominantly based on the intuition and experience of the designer. Previous Reinforcement Learning (RL) based automation approaches have agents interact with these device geometries to reach specifications guided by a weighted and meticulously engineered reward signal. Since the performance specification of circuits are usually comprised of electrical characteristics, we make use of a transformed action space in this work. Here, the agent gets to interact with electrical characteristics of individual devices, which are then translated back into geometries in a separate step. Additionally, we make use of an alternative experience replay scheme that copes with a sparse and binary reward signal. Doing so treats target specifications more like coordinates in space and he agents learn how to navigate the space efficiently, to reach any given coordinate as quickly as possible.

Featured Image

Why is it important?

Technology nodes are defined by the smallest possible geometric structure that can be fabricated. Translating a working design from a 350nm technology to a 180nm technology, for example, is a non-trivial task. There is no easy or formal way to translate the geometries from one to the other. Taking an approach, that is decoupled from the technology, where agents only know the electrical domain, makes it portable without any additional effort. Using a sparse and binary reward signal during training ensures, that the agent learns it's task from scratch without any biased expert knowledge being encoded. Instead of merely imitating an analog design expert, these agents are able to find entirely new (and potentially better) solutions than their human counterparts.

Perspectives

Decoupling the geometry from electrical behavior is the foundation for many exciting opportunities in the field of analog circuit design automation. The translation from electrical behavior to geometries of a device can be done through look-up tables, or approximated with AI or ML approaches. Doing so, transforms the design task from the geometric domain into the electrical domain, where technology dependent differences can be disregarded. In this space design becomes much easier and intuitive for humans and AI alike.

Yannick Uhlmann

Read the Original

This page is a summary of: Deep Reinforcement Learning for Analog Circuit Sizing with an Electrical Design Space and Sparse Rewards, September 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3551901.3556474.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page