# Low-latency hierarchical routing of reconfigurable neuromorphic systems

## TL;DR

This paper introduces a new routing architecture for FPGA-based spiking neural network simulations that reduces latency and improves performance.

## Contribution

A novel hierarchical tree-based routing architecture with stochastic arbitration is proposed to reduce latency in multi-FPGA systems.

## Key findings

- The hierarchical tree-based architecture maintains constant local bandwidth while being scalable.
- Stochastic arbitration reduces FIFO congestion and improves worst-case latency compared to round-robin methods.
- Performance measurements confirm lower latency and better overall performance with the proposed routing scheme.

## Abstract

A reconfigurable hardware accelerator implementation for spiking neural network (SNN) simulation using field-programmable gate arrays (FPGAs) is promising and attractive research because massive parallelism results in better execution speed. For large-scale SNN simulations, a large number of FPGAs are needed. However, inter-FPGA communication bottlenecks cause congestion, data losses, and latency inefficiencies. In this work, we employed a hierarchical tree-based interconnection architecture for multi-FPGAs. This architecture is scalable as new branches can be added to a tree, maintaining a constant local bandwidth. The tree-based approach contrasts with linear Network on Chip (NoC), where congestion can arise from numerous connections. We propose a routing architecture that introduces an arbiter mechanism by employing stochastic arbitration considering data level queues of First In, First Out (FIFO) buffers. This mechanism effectively reduces the bottleneck caused by FIFO congestion, resulting in improved overall latency. Results present measurement data collected for performance analysis of latency. We compared the performance of the design using our proposed stochastic routing scheme to a traditional round-robin architecture. The results demonstrate that the stochastic arbiters achieve lower worst-case latency and improved overall performance compared to the round-robin arbiters.