Deploying Neural Networks on RISC-V with VPU

This work shows the results of an implementation of an AI application on a RISC-V processor with a Vector Processor Unit (VPU) using TensorFlow as a common tool. We used two commercial SoCs that uses RISC-V cores that implement the RISC-V vector extension (RVV) with a width of 128 and 256 bits and they are able to run a standard Linux distribution. The developed work application classifies clouds on satellite image and is designed to be deployed aboard satellites to improve their performance and yield. In this work we adapt TensorFlow to take advantage of the vector instructions present in the system and improve computing times. First, the performance of the vector unit with different data-types, word widths and length of the vectors has been outlined and characterized to be able to optimize TensorFlow with said parameters and later the entire framework is vectorized to take advantages of this processor units.