NeuralMatrix: Compute the Entire Neural Networks with Linear Matrix Operations for Efficient Inference

The inherent diversity of computation types within individual Deep Neural Network (DNN) models imposes a corresponding need for a varied set of computation units within hardware processors. This diversity poses a significant constraint on computation efficiency during the execution of different neural networks. In this study, we present NeuralMatrix, a framework that transforms the computation of entire DNNs into linear matrix operations. This transformation seamlessly enables the execution of various DNN models using a single General-Purpose Matrix Multiplication (GEMM) accelerator. Extensive experimental results spanning different DNN models demonstrate that our approach preserves network accuracy while providing both generality and application-specific levels of computation efficiency. This allows a broad spectrum of DNN models to be executed using a single GEMM accelerator, eliminating the need for additional special function units.
View on arXiv