The objective of this SBIR proposal is to develop a design space exploration tool to rapidly evaluate different neuromorphic processing options for a given application. The neuromorphic options to be examined include deep learning accelerators, spiking neuron accelerators, memristive systems, and photonic accelerators. The tool will work with both individual algorithms and with chains of algorithms, where it will help with mapping individual algorithms within a chain to multiple neuromorphic processing options in a heterogeneous compute system. The tool will be able to take inputs from traditional neural frameworks and applications.
Benefit: Our model will be designed to take a variety of algorithms as inputs, including traditional non-neural algorithms, second generation neural algorithms (primarily deep learning), and spiking network algorithms. This flexibility will allow designers to examine the impact of going to different types of neural hardware for specific algorithms in their overall system and understand the impact of doing so.
Keywords: design space exploration, design space exploration, Performance Modeling, neuromorphic systems
