This project will address the synergy of machine learning based computing models in Big-data analytics problem, from the perspective of the subjacent computer architecture. The target is to combine efficiently into a single architecture the dissimilar requirements, minimizing the impact on the complex software stack.
The project will analyze how current system behaves in this context. From the knowledge derived, we will propose techniques to improve system efficiency and performance with minimal impact in programmer’s productivity. In particular we will explore how to define a dynamically reconfigurable processor-architecture and how to design the memory hierarchy in order to provide a transparent use of a heterogeneous set of computing elements (CPU, GPU and specialized ML accelerators). Additionally the project will explore the benefits of using specialized hardware for performing machine learning, with emphasis in unsupervised learning.