Extension of an Embedded AI Framework for RISC-V Systems

Abstract

Embedded AI deals with the execution of neural networks on microcontrollers and other small-scale devices. These systems aim to perform AI tasks on limited hardware resources, without relying on cloud-based services. This thesis presents the extension of an embedded AI framework for floating-point and quantized neural networks. The framework is extended with 1D convolutions, batch normalization, batch normalization folding, concatenation and global pooling while also lowering the memory requirements for quantized neural networks. At the beginning of this thesis, the challenges of embedded AI are explained and the RISC-V IP core used is presented. The embedded AI framework Emmi is introduced, the fundamentals for understanding convolutions and batch normalization are presented, and a brief overview of its features and the quantization technique used is given. Before implementing the extensions, Emmi is analyzed and the requirements are summarized. During the test-driven implementation, unit and integration tests ensure the functionality of the framework. After testing and benchmarking it on a non-vector RISC-V system, it is compiled with automatic vectorization and tested in a RISC-V simulator that supports vector instructions. Finally, it is shown that the framework can be used in a predictive maintenance application.