Abstract:As the demand for efficient data processing escalates, reconfigurable analog hardware which implements novel analog compute paradigms, is promising for energy-efficient computing at the sensing and actuation boundaries. These analog computing platforms embed information in physical properties and then use the physics of materials, devices, and circuits to perform computation. These hardware platforms are more sensitive to nonidealities, such as noise and fabrication variations, than their digital counterparts and accrue high resource costs when programmable elements are introduced. Identifying resource-efficient analog system designs that mitigate these nonidealities is done manually today.
While design optimization frameworks have been enormously successful in other fields, such as photonics, they typically either target linear dynamical systems that have closed-form solutions or target a specific differential equation system and then derive the solution through hand analysis. In both cases, time-domain simulation is no longer needed to predict hardware behavior. In contrast, described analog hardware platforms have nonlinear time-evolving dynamics that vary substantially from design to design, lack closed-form solutions, and require the optimizer to consider time explicitly. We present Shem, an optimization framework for analog systems. Shem leverages differentiation methods recently popularized to train neural ODEs to enable the optimization of analog systems that exhibit nonlinear dynamics, noise and mismatch, and discrete behavior. We evaluate Shem on oscillator-based pattern recognizer, CNN edge detector, and transmission-line security primitive design case studies and demonstrate it can improve designs. To our knowledge, the latter two design problems have not been optimized with automated methods before.
What problem does this paper attempt to address?
The problems that this paper attempts to solve are as follows: With the increasing demand for efficient data processing, reconfigurable analog hardware realizes new analog computing paradigms at the sensing and execution boundaries, which is expected to achieve efficient energy - based computing. However, these analog computing platforms are more vulnerable to non - idealities (such as noise and manufacturing variations) than their digital counterparts, and will bring high resource costs when introducing programmable elements. Currently, identifying resource - efficient analog system designs and alleviating these non - idealities mainly rely on manual operations.
Specifically, this paper aims to solve the following problems:
1. **Sensitivity of analog hardware**: Analog hardware is very sensitive to non - idealities such as noise, parameter variations caused by manufacturing, and environmental sensitivity, which will affect the fidelity of computing.
2. **High resource cost**: Digital interface circuits for programming hardware and performing digital - to - analog conversion increase the amount of resource usage, especially when the accuracy is improved.
3. **Lack of automated optimization methods**: Existing design optimization frameworks are usually for linear dynamic systems or specific differential equation systems, and solutions are obtained through manual analysis. These methods do not require time - domain simulation to predict hardware behavior. However, the described analog hardware platforms have non - linear time - evolving dynamics, lack closed - form solutions, and require the optimizer to consider time factors.
To solve these problems, the author proposes an optimization framework named Shem, which can optimize analog systems with non - linear dynamics, noise and mismatches, and discrete behaviors. Shem utilizes the recently popular back - propagation method for training neural ODEs, making it possible to optimize analog systems. In addition, Shem can also improve the design in case studies such as oscillator pattern recognizers, CNN edge detectors, and transmission line security primitive designs.
### Core contributions of the paper include:
- Introducing techniques for design optimization of analog systems in the time - domain, and achieving differentiation for noise, manufacturing variations, and digital logic.
- Proposing a framework named Shem, which is built based on JAX and supports automatic differentiation and optimization of non - linear time - domain simulation systems.
- Verifying the effectiveness of Shem through case studies, showing that it can improve the design in oscillator pattern recognizers, cellular non - linear network edge detectors, and transmission line security primitive designs.
### Formula summary:
- **MSE loss function**:
\[
\text{MSE}(Y, A, B, z)=\frac{1}{N}\sum_{i,j}(y_{ij}-f(x_{ij}[A, B, z](t_3)))^2
\]
- **Expected loss function**:
\[
\text{Loss}(Y, A, B, z)=E_\delta[\text{MSE}(Y, A, B, z)]
\]
- **Ordinary differential equation model**:
\[
\frac{dx_i(t)}{dt}=f(x_i(t),\theta,t)
\]
- **Gradient calculation**:
\[
\frac{dL}{d\theta}=\frac{\partial L}{\partial x_{\tau_N}}^\top\left(\frac{\partial x_{\tau_N}}{\partial x_{\tau_{N - 1}}}\frac{dx_{\tau_{N - 1}}}{d\theta}+\frac{\partial x_{\tau_N}}{\partial \theta}\right)
\]
Through these techniques, Shem can effectively optimize analog systems, making them more robust when facing non - idealities such as device mismatches and transient noise.