Analog Accelerator for Simulation and Diagnostics

Alexander Feldman, Ion Matei, Emil Totev, Johan de KleerAAAI 2020 · DOI

Abstract

Traditional digital solvers for ordinary differential equations (ODEs) often struggle with the inherent trade-off between simulation speed and numerical accuracy, particularly in time-critical applications like Model-Based Diagnosis (MBD). This paper proposes a hybrid analog-digital accelerator based on Field Programmable Analog Arrays (FPAAs). By mapping mathematical models directly onto a reconfigurable grid of analog components, the system eliminates discrete time-stepping, achieving simulation speeds up to 1,000 times faster than real-time. We detail a framework incorporating automated scaling, local feedback calculus, and a calibration-in-the-loop mechanism to mitigate analog hardware noise.

Introduction

In engineering domains such as power grid management or biological modeling, physical systems are governed by complex ODEs. While digital methods like Runge-Kutta are standard, they are computationally intensive for "stiff" equations or high-speed diagnostic needs. This research explores returning to analog computing—which is inherently continuous—to simulate the physical world more efficiently.

System Architecture and Methodology

The proposed accelerator utilizes a hybrid architecture where a digital ARM core manages the configuration and calibration of an analog computational core.

Automated Workflow and Calibration

The simulation process follows a structured pipeline from high-level mathematical description to hardware execution. An automated workflow handles the mapping of an Abstract Syntax Tree to the physical op-amp grid and performs necessary voltage scaling to prevent signal saturation.

The hybrid system workflow, illustrating the transition from host-side compilation and mapping to the controller-managed simulation and calibration loop.

To ensure precision despite the inherent noise of analog hardware, the system employs a "Calibration-in-the-Loop" strategy. The controller executes the simulation $N$ times, resetting and re-calibrating between iterations, and then averages the results to produce a stable, high-fidelity output.

Local Feedback Calculus

Mathematical operations are performed using a library of computational blocks. These blocks utilize operational amplifiers with local feedback to execute summation, integration, and non-linear operations. Complex math, such as multiplication, is handled in the log-domain ($\ln(x \cdot y) = \ln(x) + \ln(y)$), allowing for sophisticated dynamics like predator-prey models to be represented on-chip.

Experimental Results

The architecture was validated using SPICE simulations targeting a 180 nm CMOS process. Evaluation focused on classic models, including the Lotka-Volterra predator-prey equations.

Speed and Linearity

The system demonstrated a massive performance gain, achieving a time scale factor of $10^3$. For the Lotka-Volterra model, 40 seconds of real-world dynamics were simulated in just 4 milliseconds. Furthermore, the CMOS implementation exhibited strong linearity within the primary operational range of 0.25V to 1.25V.

Comparison of output voltage linearity between the ideal LTspice Universal model and the proposed 180 nm CMOS implementation across varying resistance ratios.

Discussion and Future Work

While the results show significant promise for high-speed simulation, challenges remain regarding DC offsets and hardware footprint. Future research will focus on "chopper stabilization" to reduce offset errors and migrating the design to a 40 nm feature size to improve density and bandwidth.

Conclusion

By moving away from clocked digital logic toward continuous analog propagation, this accelerator provides a path for ultra-fast, real-time simulation of physical systems. The integration of digital control for calibration and scaling addresses the traditional weaknesses of analog computing, creating a robust platform for modern diagnostic applications.

Summary updated

BibTeX
@inproceedings{feldman2020analog,
  author = {Alexander Feldman and Ion Matei and Emil Totev and Johan de Kleer},
  title = {{Analog Accelerator for Simulation and Diagnostics}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i08.7034},
}
Figure