A System for Validating Resistive Neural Network Prototypes

Building prototypes of heterogeneous hardware systems based on emerging electronic, magnetic, and photonic devices is an important area of research. The novel implementation of these systems for artificial intelligence poses new and unforeseen challenges in mixed signal data acquisition, hyperparameter optimization, and hardware co-processing. Many emerging devices exhibit unpredictable and stochastic behavior as well as poorly repeatable hysteretic effects or performance degradation. Dealing with these device challenges on top of more traditional hardware problems, like quantization errors, timing constraints, and even hardware and software bugs is an enterprise fraught with pitfalls. Equally important to the construction of the physical prototype is the co-development and integration of a design verification framework that can extensibly allow for predictable behavior of not only the entire system but also all of its parts in a modular way, allowing for seamless integration in both simulation and implementation. This work discusses Daffodil-lib, a Python based prototyping framework which, from hardware to software, enables everything from a script-based simulation to a compiled hardware-timed experiment, to everything in between with no syntactical changes for the end user.