Optimization of energy and power driven architectural exploration for multi-core and heterogeneous System on Chip

Abstract

The contribution of this work builds on top of the established virtual prototype platforms to improve both SoC design quality and productivity. Initially, an automatic system-level power estimation framework was developed to address the critical issue of early power estimation in SoC design. The estimation framework models the static and dynamic power consumption of the hardware components. These models are created from the normalized values of the basic design components of SoC, obtained through one-time power simulation of RTL hardware models. The framework allows dynamic technology node reconfiguration for power estimation models. Its instantaneous power reporting aids the detection of possible hotspot early into the design process. Adding this additional data in conjunction with a steadily growing design space of complex heterogeneous SoC, finding the right parameter configuration is a challenging and laborious task for a system-level designer. This work addresses this bottleneck by optimizing the design space exploration (DSE) process for MPSoC design. An automatic DSE framework for virtual platforms (VPs) was developed which is flexible and allows the selection optimal parameter configuration without pre-existing knowledge. To reduce exploration time, the framework is equipped with several multi-objective optimization techniques based on simulated annealing and a genetic algorithm. Lastly, to aid HW/SW partitioning at system-level, a flexible and automated workflow (SW2TLM) is presented. It allows the designer to explore various possible partitioning scenarios without going into depth of the hardware architecture complexity and software integration. The framework generates system-level hardware accelerators from corresponding functionality encoded in the software code and integrates them into the VP. Power consumption and time speedups of acceleration is reported to the designer, which further increases the quality and productivity of the development process towards the final architecture. The presented tools are evaluated using a state-of-the-art VP for a range of single and multi-core applications. Viewing the energy delay product, a reduction in exploration time was recorded at approximately 62% (worst case), maintaining optimal parameter accuracy of 90% compared to previous techniques. While the SW2TLM further increases the exploration versatility by combining modern high-level synthesis with system-level architectural exploration.