IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS — I : REGULAR PAPERS , VOL . ? ? , NO . ? ? , ? ? ? ?

We have developed a grid network that broadcasts spikes (binary events) in a multi-chip neuromorphic system by relaying them from chip to chip. The grid is expandable because, unlike a bus, its capacity does not decrease as more chips are added. The multiple relays do not increase latency because the grid’s cycle-time is shorter than the bus. We describe an asynchronous relay implementation that automatically assigns chip-addresses to indicate the source of spikes, encoded as wordserial address-events. This design, which is integrated on each chip, connects neurons at corresponding locations on each of the chips (pointwise connectivity) and supports oblivious, targeted, and excluded delivery of spikes. Results from two chips fabricated in 0.25-μm technology are presented, showing word-rates up to 45.4 M events/s. I. ADDRESS-EVENT COMMUNICATION Neuromorphic engineers attempt to capture the computational power and efficiency of biological neural systems in their hybrid analog–digital VLSI systems. These neuromorphic systems employ a similar design strategy as biology: local computations are performed in analog, and the results are communicated using all-or-none binary events (spikes). A typical neuromorphic chip consists of a 2-D array of bio-inspired circuits (core), which may include models of synapses [1], spatiotemporal filtering [2], and spike generation [3]; the array is surrounded by a transceiver that handles spike communication to and from the core through an off-chip link. Techniques to morph neural wetware into VLSI hardware have been described previously for analog circuits [4], [5] and spikebased communication [6]. Neuromorphic chips routinely require tens of thousands of axonal connections to propagate spikes to and from the core — far too many to be implemented using dedicated wires. The address-event link, originally introduced by Mahowald and Sivilotti, implements these axonal connections using a timedivision-multiple-access (TDMA) link [7], [8]. Transceivers multiplex–demultiplex spikes over a few high-speed wires for efferent–afferent (outgoing–incoming) axons, respectively. For each event, the axon’s identity is encoded with a unique binary word, an address event. Manuscript received December; revised ????. This work was made possible by funding from the Whitaker Foundation, the NSF’s BITS and CAREER programs (EIA-0130822, ECS00-93851), and the Hong Kong Research Grants Council (HKUST6200/03E). P. Merolla and J. Arthur are with the University of Pennsylvania, Dept of Bioengineering, 3320 Smith Walk, Philadelphia, PA 19104-6392 USA (e-mail: {pmerolla,jarthur}@seas.upenn.edu). K. Boahen was also with the University of Pennsylvania at the time of this work, and is currently affiliated with Stanford University (e-mail: boahen@stanford.edu). B. Shi is with the Dept. of Electrical Engineering, Hong Kong University of Science and Technology, Clear Water Bay Kowloon, HONG KONG SAR (email: eebert@ee.ust.hk). Copyright (c) 2006 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending an email to pubs-permissions@ieee.org. ... ... T R A N S M I T T E R R E C E I V E R MERGE SPLIT T R A N S M I T T E R R E C E I V E R MERGE SPLIT T R A N S M I T T E R R E C E I V E R MERGE