A scalable multi-core architecture with heterogeneous memory structures for Dynamic Neuromorphic Asynchronous Processors (DYNAPs)

Saber Moradi, Qiao Ning, Fabio Stefanini, Giacomo Indiveri

Abstract

This paper presents a novel routing architecture for Dynamic Neuromorphic Asynchronous Processors (DYNAPs) designed to overcome the memory scaling bottleneck in large-scale neuromorphic systems. By integrating hierarchical and mesh routing strategies with distributed, heterogeneous memory structures, the proposed architecture significantly reduces memory requirements while maintaining high programmability. A mixed-signal prototype chip with 1,000 neurons demonstrates ultra-low power consumption (17 pJ per routing hop) and low latency, achieving 100% accuracy on a real-time visual classification task using event-based sensor data.

Introduction

Neuromorphic computing leverages asynchronous spikes, typically using the Address-Event Representation (AER) protocol, to achieve high power efficiency. However, scaling these systems to match biological complexity is challenging due to the massive "fan-out" of neurons. Traditional hardware implementations require large routing tables, leading to bottlenecks in bandwidth, latency, and silicon area. Existing solutions often compromise on network flexibility or rely on external DRAM, which reintroduces the von Neumann bottleneck. This work addresses these challenges by proposing a scalable routing fabric that minimizes memory usage through a two-stage "tag-based" shared-addressing scheme.

Methods

The core innovation is a two-stage routing mechanism that exploits the clustered connectivity patterns observed in biological brains.

Two-stage tag-based shared-addressing routing scheme.
  1. Stage 1 (Point-to-Point): Source neurons transmit events to intermediate routers targeting specific clusters (cores).
  2. Stage 2 (Broadcast): The intermediate router broadcasts the event to all neurons within the target cluster. Destination neurons utilize local Content Addressable Memory (CAM) "tags" to selectively accept or ignore events.

This approach allows memory requirements to scale logarithmically ($O(\log N)$) rather than linearly with the number of neurons. The physical implementation combines a 2D-mesh for inter-chip traffic with a hierarchical tree structure for intra-chip communication, as shown in the multi-level router hierarchy.

Hierarchical routing architecture combining mesh and tree structures.

The digital routing fabric is built using ultra-low power Quasi Delay-Insensitive (QDI) asynchronous circuits. The prototype chip employs a mixed-signal design: sub-threshold analog circuits emulate biophysical neuron dynamics, while asynchronous digital circuits manage event routing. Memory is distributed heterogeneously, with SRAM in routers for source addresses and CAM in synapses for target tags, effectively bypassing the von Neumann bottleneck.

Results and Evaluation

The architecture was validated using a prototype chip fabricated in a 0.18 µm CMOS process, featuring 4 cores, 1,000 neurons, and 64,000 synapses. Performance was evaluated using the Poker-DVS dataset, which contains high-speed event-based data from a Dynamic Vision Sensor.

Key performance metrics include:

  • Energy Efficiency: 17 pJ per routing hop at 1.3V, competitive with state-of-the-art digital solutions on much smaller process nodes.
  • Latency: Local broadcast time of ~27 ns and cross-chip latency of 15.4 ns.
  • Application Performance: A 9-chip system implementing a 3-layer CNN achieved 100% classification accuracy on the Poker-DVS dataset with a latency of less than 30 ms.

Unlike architectures like IBM's TrueNorth, which scale quadratically for large fan-out models due to the need for extra routing cores, this architecture scales linearly for Convolutional Neural Networks (CNNs).

Conclusion and Future Work

The DYNAP architecture demonstrates that combining hierarchical routing with distributed heterogeneous memory can provide the scalability and efficiency required for real-time neuromorphic applications. While current performance is limited by I/O circuit speeds and a local throughput cap of 38 Mevents/sec, future iterations on advanced technology nodes (e.g., 28nm FDSOI) are expected to reach throughputs of 1 Gevents/sec. Integration with emerging resistive memory (ReRAM) technologies offers a path toward even higher density and lower power consumption.

Summary updated

BibTeX
@article{moradi2018scalable,
  author = {Saber Moradi and Qiao Ning and Fabio Stefanini and Giacomo Indiveri},
  title = {{A scalable multi-core architecture with heterogeneous memory structures for Dynamic Neuromorphic Asynchronous Processors (DYNAPs)}},
  journal = {IEEE Trans. Biomed. Circuits Syst.},
  year = {2018},
  url = {https://openreview.net/forum?id=IOgyp45Tjo},
}
Figure