Impact of Input Encoding and ADC Resolution on Matrix-Vector Multiplication Accuracy

Analog In-Memory Computing based on Non-Volatile Memories is a promising candidate to reduce latency and power consumption of neural network inference in edge-computing applications. This kind of computational accelerators allows both storing weights and performing in-situ analog computation inside the array. This analysis evaluates the impact that input encoding schemes could have on the accuracy of Matrix-Vector Multiplication (MVM), depending on ADC resolution. Two commonly used encoding technique, Pulse Width Modulation and Bit-Serial, are assessed in term of computation time, system resolution and average MVM error.