Design of a Coarse Grain Reconfigurable Array for Neural Networks

In recent years, the development of Systems on Chip (SoCs) is facing increasing challenges, which are starting to expose the limitations of current design philosophies. In spite of many conservative measures, often at the expense of computational efficiency, realising a functional SoC has come to require tens of Millions USD. Such a tendency threatens to hinder innovation and keeps away applications that require the efficiency of ASIC. Among them, a prime example are Artificial Neural Networks (ANNs), which have become one of the most popular research topics and already see many commercial applications in countless fields. The increasing complexity of ANNs makes them quite demanding in terms of computational power, so the inefficient approach of modern SoCs is quickly becoming unsuited for them. This thesis proposes a novel VLSI design framework called SiLago, which has the potential to overcome the main architectural limitations of modern SoCs while also cutting their engineering costs. One of the main innovations is the adoption of a complete hardware approach, where all computation relies on a Coarse Grain Reconfigurable Array (CGRA) of custom blocks that are able to accelerate different applications reusing the same hardware. Implementing popular algorithms such as ANNs on a SiLago platform is a good opportunity to prove its advantages and this is precisely the rationale behind this thesis, which is about the design of two CGRAs compatible with SiLago and customised to support three classes of ANNs. One CGRA targets Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM), while the other is suited for Self-Organizing Maps (SOM). In particular, DataPath Unit and Compression Engine for both types of SiLago blocks have been designed from scratch. Special care has been given to obtaining versatile and efficient implementation for compression algorithms and for the nonlinear functions sigmoid, hyperbolic tangent, exponential and softmax required by ANNs. The design flow has been carried out end-to-end, from algorithm specifications to post place & route verification. The final results are two fully functional CGRAs, detailed down to the physical level and accompanied by extensive reports on their area occupation.