Power-Aware FPGA Mapping of Convolutional Neural Networks

With an unprecedented accuracy in numerous AI tasks, convolutional neural networks (CNNs) are rapidly deployed on power-limited mobile and embedded applications. Existing mapping approaches focus on achieving high performance without explicit consideration of power consumption, leading to suboptimal solutions when power is considered in a subsequent stage. In this context, there is an emerging need for power-aware methodologies for the design of custom CNN engines. In this work, a methodology is presented for modelling the power consumption of FPGA-based CNN accelerators using a high-level description of modules, together with a power-centric search strategy for exploring power-performance trade-offs within the CNN-to-FPGA design space. By integrating into an existing CNN-to-FPGA toolflow, the proposed power estimation method can yield a prediction accuracy of 93.4% for total system power consumption. Furthermore, it is demonstrated that the associated power-oriented exploration approach can generate CNN accelerators with a 20.1% power reduction over a purely throughput-driven design for AlexNet, maintaining the design's throughput.