Exploration of Dataflows for Neuromorphic Accelerators

Description

Spiking Neural Networks (SNNs) are emerging as an energy-efficient alternative for traditional edge-AI applications and dedicated neuromorphic hardware, due to their inherent event-driven and sparse nature. However, realizing this algorithmic efficiency in hardware is often bottlenecked by the memory subsystem and the execution strategy. Specifically, irregular data access patterns, frequent memory fetches for neuron states and unbalanced workloads hinder efficiency. This thesis aims to explore dataflows prioritizing the reduction of data movement. Different focus can be set.

Requirements

• Good Knowledge of Python or C++
• Good Knowledge of digital logic
• Knowledge of VHDL/Verilog
• Interest in performance modeling

References

• Rathi et al._ „Exploring Neuromorphic Computing Based on Spiking Neural Networks: Algorithms to Hardware“
• Sharma et al. "Identifying Efficient Dataflows for Spiking Neural Networks"