On-chip and system-area multi-processor interconnection networks in advanced processes for maximizing performance minimizing cost and energy

Abstract

A chip design environment is disclosed which accepts application specific processing, memory and IO elements and declarative specification of function, cost and performance of peripheral, low-level and infrastructural elements and of overall design and generates synthesizable module RTLs and relevant place-and-route constraints. The generated elements include the network interconnecting all the elements, a programming memory consistency model and its coherence protocol, allocation and scheduling processes realizing run-time inference of optimal parallel execution and processes for control of coherence action and prefetch intensity, task-data migration, voltage-frequency scaling and power-clock gating. The environment employs knowledge bases, models to predict performance and to assign confidence scores to predictions and, in turn, the predictions to explore space of topology, architecture, composition, etc options. The environment generates synthesizable module RTLs to complete the design and relevant place-and-route constraints. User may simulate the synthesized design. If a user shares simulation results, the environment may evaluate the predicted performance against performance determined by simulation and use the results to update its knowledge and models.