Abstract
This paper presents ultra-compact, energy-efficient analog synapse and neuron circuits designed for a modern 28nm FD-SOI process. By utilizing sub-threshold operation and novel leakage-compensation techniques, the researchers demonstrate a neuromorphic architecture capable of both "slow" biologically realistic dynamics and "fast" high-speed data processing. The design achieves extreme density with synapse and neuron areas of 3 µm² and 20 µm², respectively, while maintaining stability against manufacturing noise and transistor leakage.
Introduction
Traditional computing follows the von Neumann architecture, where memory and processing are physically separated. This creates a "bottleneck" because moving data between the two consumes the vast majority of a system's power. Neuromorphic computing solves this by co-locating memory (synaptic weights) and computation (neural integration).
However, implementing these brain-inspired designs in advanced, highly scaled silicon processes (like 28nm) is difficult. As transistors get smaller, they become "leaky"—consuming power even when off—and their behavior becomes less predictable due to mismatch. This paper tackles the challenge of building stable, pA-to-nA (pico-to-nano Ampere) analog circuits that can maintain the long time-constants needed for real-world sensory processing while surviving the "noise" of modern manufacturing.
Circuit Design and Methodology
The authors employ several clever circuit-level strategies to make analog neuromorphic computing viable at 28nm:
- Sub-threshold Operation: Instead of using transistors as binary switches, they operate them in the sub-threshold regime, where the physics of the device naturally implements the exponential and logarithmic functions required for neural modeling.
- Leakage Canceling: A dedicated block mirrors the off-channel leakage current of the synapse array and subtracts it from the total signal. This allows the circuit to distinguish between actual neural signals and background leakage.
- Split-Transistor "Pseudo-Cascode": To achieve the tiny currents (pA) necessary for long time-constants (hundreds of milliseconds) without using massive capacitors, a "split-transistor" technique reduces the Early effect and improves linear performance.
- Differential Pair Integrator (DPI): This block implements the temporal dynamics of synapses, acting as a current-mode low-pass filter.
- Modular Neuron Design: The Integrate-and-Fire (I&F) neuron is built from modular "compartments" (LEAK, AHP, Na+, K+, etc.) that allow it to mimic complex biological behaviors like spike-frequency adaptation.
Key Insights and Results
The evaluation was performed using post-layout simulations of a chip taped out in 28nm FD-SOI technology, using 100 Hz input spike trains to validate temporal responses.
Computational Efficiency
By representing signals as currents, the summation of thousands of synaptic inputs happens inherently at the neuron's input node via Kirchhoff’s current law. In this architecture, a synapse acts as a gate where the weight determines the current flow into the neuron upon a pre-synaptic spike.
Versatility and Density
While neuromorphic circuits are often associated with "slow" biological time-scales, the authors demonstrate that by tuning the refractory period, these circuits can operate in the microsecond range. This versatility makes them suitable for both biological modeling and high-speed deep learning models (e.g., implementing ReLU functions). The design achieves extreme density, with a synapse layout area of just 3 µm² and a neuron area of 20 µm².
Biological Realism
The circuits successfully reproduced time-constants as long as hundreds of milliseconds, essential for processing natural signals like speech or gestures. The neuron's transfer function can be tuned to be highly linear, mimicking the activation functions popular in modern AI.
Conclusion
This work proves that analog neuromorphic primitives can scale alongside digital logic in advanced 28nm FD-SOI processes. By using asynchronous Address-Event Representation (AER) and sub-threshold currents, the system minimizes power consumption, making it ideal for "Always-On" IoT applications. Future work will focus on integrating these compact cores into larger multi-core architectures and exploring on-chip spike-based learning algorithms.
Summary updated