General-purpose Dataflow Model with Neuromorphic Primitives

TL;DR

This work addresses the challenge of executing general-purpose programs with control flow on neuromorphic hardware by introducing neuromorphic dataflow (NDF), which decouples control into routing ('where') and token-generation ('when') primitives. The where primitive reconfigures data routing with $m_{where}$ inputs and $n_{where}$ outputs, potentially via an adjacency matrix $A$ of size $n_{where}\times m_{where}$, while the when primitive leverages a modified spiking neuron to generate switch tokens based on a multi-dimensional membrane potential $v^k$ and a weight matrix $W$, with the update $v^k_{t+1} = f\left(v^k_t + I^T W\right)$. This combination reduces graph complexity, improves neuromorphic compatibility, and enables plasticity through mechanisms like STDP or surrogate-gradient BP, with hardware mappings to 2D-mesh routers and soma-like modules supporting adaptation. The authors demonstrate that NDF preserves programmability and leverages hardware parallelism, showing potential for broad applicability across neuromorphic platforms. Overall, NDF offers a practical pathway to deploy general-purpose, Turing-complete computation on neuromorphic hardware by embedding neuromorphic primitives directly into the dataflow representation. It combines compact control-logic representation with learnable components to deliver both efficiency and adaptability on neuromorphic systems.

Abstract

Neuromorphic computing exhibits great potential to provide high-performance benefits in various applications beyond neural networks. However, a general-purpose program execution model that aligns with the features of neuromorphic computing is required to bridge the gap between program versatility and neuromorphic hardware efficiency. The dataflow model offers a potential solution, but it faces high graph complexity and incompatibility with neuromorphic hardware when dealing with control flow programs, which decreases the programmability and performance. Here, we present a dataflow model tailored for neuromorphic hardware, called neuromorphic dataflow, which provides a compact, concise, and neuromorphic-compatible program representation for control logic. The neuromorphic dataflow introduces "when" and "where" primitives, which restructure the view of control. The neuromorphic dataflow embeds these primitives in the dataflow schema with the plasticity inherited from the spiking algorithms. Our method enables the deployment of general-purpose programs on neuromorphic hardware with both programmability and plasticity, while fully utilizing the hardware's potential.

Figures (6)

• Figure 1: A basic demonstration of conventional dataflow. (A) A program segment written in Algol-like syntax designed for von-Neumann architecture. (B) The equivalent dataflow representation of program A. This example is taken from dennis1974data
• Figure 2: A basic demonstration of neuromorphic dataflow. (A) Same program in Fig. \ref{['fig:conventional_dataflow']} (A). (B) The equivalent neuromorphic dataflow model of program A.
• Figure 3: A dynamic where primitive takes switch token as input to change its connections of data in/out arcs.
• Figure 4: The firing rule of when primitive with two-dimensional membrane potential and four separate regions.
• Figure 5: Approximation error of a simple program.