Our paper "NoC-Based Support of Heterogeneous Cache-Coherence Models for Accelerators" has been presented at the NOCS'18 Symposium, during the Embedded Systems Week. The paper and the slides are now available.

The paper shows that choosing the appropriate coherence model for accelerators can improve performance and decrease energy consumption. Moreover, the best coherence choice varies at runtime for each accelerator's invocation. For these reasons, we propose a cache-coherence protocol and a NoC-based architecture that supports: three models of coherence for accelerators, the runtime selection of the coherence model, the coexistence of heterogeneous coherence models for accelerators.