Abstract-Given the limited current understanding of the neural model of computation, hardware neural network architectures that impose a specific relationship between physical connectivity and model topology are likely to be overly restrictive. Here we introduce, in the SpiNNaker chip, an alternative approach: a mappable virtual topology using an asynchronous network-on-chip (NoC) that decouples the "logical" connectivity map from the physical wiring. Borrowing the established digital RAM model for synapses, we develop a concurrent memory access channel optimised for neural processing that allows each processing node to perform its own synaptic updates as if the synapses were local to the node. The highly concurrent nature of interconnect access, however, requires careful design of intermediate buffering and arbitration. We show here how a locally buffered, one-transaction-per-node model with multiple synapse updates per transaction enables the local node to offload continuous burst traffic from the NoC, allowing for a hardware-efficient design that supports biologically realistic speeds. The design not only presents a flexible model for neural connectivity but also suggests an ideal form for general-purpose high-performance on-chip interconnect.
I. HOW TO MAP NEURAL NETWORKS TO HARDWARE?W HILE the dynamics of spiking neural networks at the neuron level are fairly well-understood, questions of network organisation at the functional unit level, data representation, and inter-system communication remain largely unanswered [1]. Important biological discoveries [2] have fuelled ongoing debates on the nature of neural structure and dynamics [3], which seem unlikely to lead to any resolution without powerful and sophisticated modelling probably incorporating dedicated hardware [4]. How to implement this hardware, however, is an almost equally contentious question.One popular method is the "neuromorphic" chip, exemplified by [5], using analogue circuits to model neural properties directly. While these devices remain an important research area since analogue circuits are in principle more biologically accurate as well as more space-efficient, in practice analogue circuitry suffers from a significant device and process technology lag behind digital. The second method is the "neural accelerator", relying on carefully optimised digital technologies with shared synaptic memory. MASPINN [6] and SP 2 INN [7] are examples of this approach. In principle digital devices are programmable and thus more flexible than analogue, however, to date, the need to use bit-mapped inputs over standard bus interfaces for synaptic memory access limits model choice. Historically, both models, relying on a direct mapping from the hardware to the physical structure More recently, FPGA's [8] offer the possibility for configurable topologies and processing functions, but use a circuit-switched connectivity model that severely constrains utilisation of routing resources. If the mapping of the neural model to the FPGA remains one-to-one, then, just...