EDITH: Efficient Design for Intelligent devices exploiting emerging TecHnologies

点击翻译按钮获取中文摘要

The current trend in Internet of Things is to design processing subsystems on the edge that are smarter, smaller and more autonomous. Therefore, it is expected that the Internet of Things of the future will not be composed of passive devices sending data to a server, it will be rather like a distributed computing fabric, where many of the IoT devices and subsystems themselves will process and analyze data in order to make autonomous decisions. Typical target applications include video-processing systems integrated in smart cars, video surveillance cameras, mobile phones, home personal assistants, drones, healthcare devices, industrial monitoring and control, etc. The key enablers for this new paradigm will be the innovative programmable devices and subsystems able to execute multiple complex algorithms including deep learning algorithms (ex. Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN)), under tight timing and power constraints. This new class of algorithms involve repeated multiplications of matrices, which are computationally intensive and require high data bandwidths. In addition, the power efficiency requirements are very high: up to 10 Tera Operations Per Second (TOPS) per Watt. Conventional computing architectures are not well suited to these new requirements. In this context, we aim to innovate at the architecture level of devices, by considering emerging technologies like silicon-photonics that have progressively attracted more and more attention for their use in tackling the high-power consumption and low bandwidth issues in conventional technologies. These technologies can be integrated into proposed application-specific architectures tuned to implement deep learning algorithms like CNN and RNN by exploiting the low power consumption and the unequalled speed and bandwidth specific to silicon-photonics. We will define an architecture exploration flow that supports the integration of photonics-based components, distributed parallel memories, and application-specific programmable processors that coordinate the data movement and computation. We will develop new performance models that integrate the components mentioned above. We will also propose efficient programming and deployment models for deep learning algorithms. In this models, security will be one of the cetric metrics.  The novelty of the proposed contributions resides mainly in: (1) the architecture-level and system-level view of the intelligent devices based on the silicon-photonics emerging technologies, (2) the combining of top-down and bottom up approach and (3) the perfect complementarity with the advanced physical-level research proposing innovative emergent technologies.