NEUROWARE: Efficient and Robust Hardware for Brain-Inspired Computing (A project of the Knowledge Generation National Program)

The NEUROWARE project deals with one of the topics with most intensive activity for these and coming years: big data processing, and in particular the area of efficient brain-inspired computing. Many applications require deep learning algorithms to be computed near to the source of information, what imposes strict restrictions in terms of memory, computational costs and energy consumption along with adaptability to the source environment. However, there is a growing gap between the needs of algorithms designed by data scientist and the capabilities of the hardware available to accommodate them. These efficiency and energy gaps are not bridged by the technology reduction imposed by Moore’s Law, there is a need for novel circuital and architectural paradigms.

In NEUROWARE we propose hardware architectures for brain-inspired computing that contribute to increase the efficiency of current solutions to provide better performance and reduced energy consumption. The challenge is to implement reliable intelligence in edge devices and platforms in a way that is efficient in terms of performance and power. The general objectives of this project address some of these issues from the device, circuit and architectural levels. NEUROWARE plans to explore the architectural limits of circuits implementing the basic processing element in deep neural networks (DNN), the artificial neuron.

• First it proposes digital implementations at circuit and architectural level targeting novel designs for DNN. Furthermore, since reliability is a key issue in current nanometer technologies, the adaptability of DNN hardware to process, temperature, VDD, radiation and aging variations will be also studied.
• A second objective of NEUROWARE is the exploration of circuits and architectures for the robust design of memristive-based brain-inspired computing. There is a clear potential of ReRAM for DNN implementations, because it provides efficient implementation of interconnections and computation in memory for brain-inspired circuits. This results in a reduction in data management and a significant improvement in performance and energy savings. However, the immaturity of ReRAM technology makes designers face serious issues like device non-uniformity, conductance level instability, sneak path currents, and wire resistance. This directly affects the reliability of the implementation and has serious impact on the array size and system performance.
• A last goal of NEUROWARE is the design of configurable hardware for the implementation of DNNs. This configurability will be studied both for the design of digital artificial neurons and memristive architectures. Finally, the project proposes the combination of both digital and analog memristor-based circuitries to conform efficient mixed-signal architectures.