ABSTRACT
This chapter overviews current technologies for high-performance, lowpower neural networks. To cope with the high computational and storage resources, hardware optimisation techniques are proposed: Deep Learning (DL) compilers and frameworks, DL hardware coupled with hardwarespecific code generators. More specifically, we explore the quantization mechanism in deep learning, based on a deep-CNN classification model. We highlight the accuracy of quantized models and explore their efficiency on a variety of hardware platforms. Through experiments, we show the performance achieved using general-purpose hardware (CPU and GPU) and a custom ASIC (TPU), as well as the simulated performance for a reduced bitwidth representation of 4 bits, 2 bits (ternary) down to 1-bit heterogeneous quantization (FPGA).
