Deep Learning (DL) offers the advantages of high accuracy performance at tasks such as image recognition, learning of complex intelligent behaviors, and large-scale information retrieval problems such as intelligent web search. To attain the benefits of DL, the high computational and energy-consumption demands imposed by the underlying processing, interconnect, and memory devices on which software-based DL executes can benefit substantially from innovative hardware implementations. Logic-in-Memory (LIM) architectures offer potential approaches to attaining such throughput goals within area and energy constraints starting with the lowest layers of the hardware stack. In this paper, we develop a Spintronic Logic-in-Memory (S-LIM) XNOR neural network (S-LIM XNN) which can perform binary convolution with reconfigurable in-memory logic without supplementing distinct logic circuits for computation within the memory module itself. Results indicate that the proposed S-LIM XNN designs achieve 1.2-fold energy reduction, 1.26-fold throughput increase, and 1.4-fold accuracy improvement compared to the state-of-the-art binarized convolutional neural network hardware. Design considerations, architectural approaches, and the impact of process variation on the proposed hybrid spin-CMOS design are identified and assessed, including comparisons and recommendations for future directions with respect to LIM approaches for neuromorphic computing.
Domino circuits are used to achieve higher system performance than static CMOS techniques. This work briefly surveys domino keeper designs for high fan-in domino circuits. A new domino circuit structure is shown in this paper that reduces the power-delay-product over 16% as compared to previous domino techniques with keepers
Due to various breakthroughs and advancements in machine learning and computer architectures, machine learning models are beginning to proliferate through embedded platforms. Some of these machine learning models cover a range of applications including computer vision, speech recognition, healthcare efficiency, industrial IoT, robotics and many more. However, there is a critical limitation in implementing ML algorithms efficiently on embedded platforms: the computational and memory expense of many machine learning models can make them unsuitable in resource-constrained environments. Therefore, to efficiently implement these memory-intensive and computationally expensive algorithms in an embedded computing environment, innovative resource management techniques are required at the hardware, software and system levels. To this end, we present a novel quality-of-service based resource allocation scheme that uses feedback control to adjust compute resources dynamically to cope with the varying and unpredictable workloads of ML applications while still maintaining an acceptable level of service to the user. To evaluate the feasibility of our approach we implemented a feedback control scheduling simulator that was used to analyze our framework under various simulated workloads. We also implemented our framework as a Linux kernel module running on a virtual machine as well as a Raspberry Pi 4 single board computer. Results illustrate that our approach was able to maintain a sufficient level of service without overloading the processor as well as providing an energy savings of almost 20% as compared to the native resource management in Linux.
30 June 2009 in Journal of Signal Processing Systems
Data centers worldwide have consumed massive amount of energy. The server processor energy efficiency has the biggest impact on energy usage in data center. In this paper, factors determining the processor circuit power are introduced and different power saving techniques are briefly reviewed. A novel methodology of using double edge triggering in server processor is proposed. By using the proposed double edge triggering implicitly pulsed CBS_ip flip-flop, the estimated energy saving of data center could be 4.3%; the energy saving on worldwide data centers will be estimated over 12 B kWh.