Reconfigurable field-programmable gate arrays (FPGAs) provide an effective programmable resource for implementing hardware-based artificial neural networks (ANNs). They are low cost, readily available and reconfigurable-all important advantages for ANN applications. However, FPGAs lack the circuit density necessary to implement large parallel ANNs with many thousands of synapses. This paper presents an architecture that makes it feasible to implement large ANNs with FPGAs. The architecture combines stochastic computation techniques with a novel lookup-table-based architecture that fully exploits the lookup-table structure of many FPGAs. This lookup-table-based architecture is extremely efficient: it is capable of supporting up to two synapses per configurable logic block (CLB). In addition, the architecture is simple to implement, self-contained (weights are stored directly in the synapse), and scales easily across multiple chips.
展开▼
