Retinomorphic Vision Systems Ii: Communication Channel Design

Kwabena Boahen Physics of Computation Laboratory California Institute of Technology MS 136-93, Pasadena, CA 91125, USA [email protected] ABSTRACT I discuss the tradeo s faced when asynchronous pulse trains are transmitted among large, two-dimensional, arrays of neurons on di erent chips, using time-division multiplexing, and present an implementation of an arbitered, random-access, channel. The long cycle time that plagues arbitered channels is addressed in the implementation described here by pipelining. Cycle times ranging from 420ns to 730ns were achieved, for 64 64 arrays, in a 2 m CMOS process, yielding a peak throughput of 2.38M spikes/second. 1. CHANNEL-DESIGN TRADEOFFS The two-chip neuromorphic system shown in Figure 1 uses an interchip communication channel to transmit spiketrains between neurons at corresponding locations on each chip. The channel performance may be rated according to the following criteria: Capacity: The maximum rate at which spikes can be transmitted. It is equal to the reciprocal of the minimum communication cycle period. Latency: The mean time a spike spends in transit between a neuron in the sending population and a neuron in the receiving population. Temporal Dispersion: The standard deviation of the channel latency. Integrity: The fraction of spikes that are delivered to the correct destination. All four criteria together determine the throughput, which is de ned as the usable fraction of the channel capacity, because the load o ered to the channel must be reduced to achieve more stringent speci cations for latency, temporal dispersion, and integrity. I discuss tradeo s between adaptive quantization versus xed quantization, and between arbitration versus free-for-all in Sections 2 and 3, respectively. Then, I describe an implementation for a pipelined, arbitered, random-access channel in Section 4, and present test results from a working multichip neuromorphic system that uses this communication channel in Section 5. My conclusions are in Section 6. 2. ADAPTIVE VERSUS FIXED We are given a desired sampling rate fNyq and an array of N signals to be quantized. We use adaptive, 1-bit, quantizers that sample at fNyq when the signal is changing, and sample at fNyq=Z when the signal is static. Let the probability that a given quantizer samples at fNyq be a: That is, a is D em ul tip le xi ng