Heterogeneous Architectures For Parallel Acceleration

To enable a new generation of digital computing applications, the greatest challenge is to provide a better level of energy efficiency (intended as the performance that a system can provide within a certain power budget) without giving up a systems's flexibility. This constraint applies to digital system across all scales, starting from ultra-low power implanted devices up to datacenters for high-performance computing and for the "cloud". In this thesis, we show that architectural heterogeneity is the key to provide this efficiency and to respond to many of the challenges of tomorrow's computer architecture - and at the same time we show methodologies to introduce it with little or no loss in terms of flexibility. In particular, we show that heterogeneity can be employed to tackle the "walls" that impede further development of new computing applications: the utilization wall, i.e. the impossibility to keep all transistors on in deeply integrated chips, and the "data deluge", i.e. the amount of data to be processed that is scaling up much faster than the computing performance and efficiency. We introduce a methodology to improve heterogeneous design exploration of tightly coupled clusters; moreover we propose a fractal heterogeneity architecture that is a parallel accelerator for low-power sensor nodes, and is itself internally heterogeneous thanks to an heterogeneous coprocessor for brain-inspired computing. This platform, which is silicon-proven, can lead to more than 100x improvement in terms of energy efficiency with respect to typical computing nodes used within the same domain, enabling the application of complex algorithms, vastly more performance-hungry than the current state-of-the-art in the ULP computing domain.