It may be used in devices that target machine learning applications to speed-up computation and improve efficiency. While the hardware specifications may not be superior to similar products in the market, we integrate it with a RISC-V core and publish it as open source, while most existing products are closed-source and work with closed ISAs.
SAURIA (Systolic Array tensor Unit for aRtificial Intelligence Acceleration) is a CNN (Convolutional Neural Network) accelerator that efficiently executes heavy computations needed for neural network algorithms, such as 2D convolutions and general matrix-matrix multiplications. The accelerator is meant to be loosely coupled to a host CPU, which must manage its local memories (fetch input data and extract results) and configure its execution through a memory-mapped interface. Once the host has prepared and started the execution, SAURIA will perform the convolution independently, and raise an interrupt upon full completion.
![](https://www.bsc.es/sites/default/files/public/styles/bscw2_-_simple_crop_style/public/bscw2/content/software-app/logo/captura_de_pantalla_35.png?itok=VAWouAvb&sc=ae35b44db124ad6286e5b54aaac5d0e8)
Contacto:
License:
Solderpad Hardware License (Version 2.1)