Generic Accuracy Configurable Matrix Multiplication-Addition Accelerator using HLS

Matrix Multiplication-Addition is one of the most common calculations when implementing Machine Learning (ML) algorithms for inference. Edge devices have limited processing power, due to resource and energy constraints, making the execution of these calculations a challenging task. This paper proposes the design of a generic configurable accelerator architecture for Generic Matrix Multiplication-Additions (GEMMA), implemented in untimed C++ for High-Level Synthesis and adaptable in matrix size, data bit-width, and data type for accuracy configuration, allowing tuning the impact on the overall design resource consumption. The overall analysis utilises existing Processing Elements (PE) from previous work as the execution units to perform the matrix operations. This work analyses the proposed architecture to spot design compromises regarding numerical accuracy. Also, this work points out that the proposed architecture inherits the behaviour of implementing each PE, presenting a trade-off between granularity and design efficiency.