Half-precision Floating Point on Spiking Neural Networks Simulations in FPGA

The use of half-precision floating-point numbers (hFP) in simulations of spiking neural networks (SNN) was investigated. The hFP format is used successfully in computer graphics and video games for storage and data transfer. The IEEE 754-2008 standard settles that arithmetic operations must occur at least on single-precision floating-point format (sFP). This means that it is necessary to convert hFP to sFP for arithmetical operations and reconvert the results to hFP before storing it. The influence of successive conversions when simulating SNN is the main concern of this article. Three methods were used to evaluate the impact of hFP on SNNs: (i) F-I curve, (ii) subthreshold regime, and (iii) the time for the next spike. We have tested the leaky integrate-and-fire and the Izhikevich's neuron model; both presented similar results. The data show that SNNs simulated with sFP present equivalent results when compared to the ones simulated with hFP with identical topology. Such results are important because hFP requires half of the memory space, simpler buses, and lower bandwidth for transferring data. We may infer they require lower clock frequency consequently lower power consumption. These are essential factors for real-time simulation of SNN on embedded electronics. The sFP to hFP conversion circuits, and vice versa, may be implemented using few logical blocks in a field-programmable gate arrays (FPGA) with no relevant Iatency. We conclude that data in the hFP format are suitable for SNNs synthesized in FPGAs, even though such implementations require conversion circuits.