As the feature size of the semiconductor process is scaling down to 10nm and below, it is possible to assemble systems with high performance processors that can theoretically provide computational power of up to tens of PLOPS. However, the power consumption of these systems is also rocketing up to tens of millions watts, and the actual performance is only around 60% of the theoretical performance. Today, power efficiency and sustained performance have become the main foci of processor designers. Traditional computing architecture such as superscalar and GPGPU are proven to be power inefficient, and there is a big gap between the actual and peak performance. In this paper, we present the MaPU architecture, a novel architecture which is suitable for data-intensive computing with great power efficiency and sustained computation throughput. To achieve this goal, MaPU attempts to optimize the application from a system perspective, including the hardware, algorithm and corresponding program model. It uses an innovative multi-granularity parallel memory system with intrinsic shuffle ability, cascading pipelines with wide SIMD data paths and a state-machine-based program model. When executing typical signal processing algorithms, a single MaPU core implemented with a 40nm process exhibits a sustained performance of 134 GLOPS while consuming only 2.8 W in power, which increases the actual power efficiency by an order of magnitude comparable with the traditional CPU and GPGPU.

Detecting targets in hyperspectral image under compressive sensing (CS) is crucial for real-time applications in front-ends. However, applying existing deep learning methods directly to compressed hyperspectral image has been challenging, including the uncertainty introduced by the random sensing matrix in the CS acquisition model, and the ineffectuality of traditional similarity metrics. In this paper, we present a solution to these challenges by introducing a Compressive Sensing-based Triplet Transformer Detector (CS-TTD). Our approach involves data augmentation using the Restricted Distribution Property (RDP) to generate samples for a balanced number of positive samples. followed by the Siamese Network with a triplet transformer to transition spectral vectors from the compressed domain to an embedding space. In addition, we introduce a Combine Convolution Net (CCN) classification module to replace traditional metrics for obtaining classification results. To improve the training and make full use of dataset labels, we also present a two-stage training approach, known as intercategory separation and intracategory aggregation (ISIA), combined with Hard-negative mining, Semi Hard-negative mining. Besides, to estimate the minimal number of compressed bands for sufficient detection performance, we introduce a kernel distribution estimation-based sequential iteration (KDE-SI). Our method achieves hyperspectral target detection without reconstruction, and experimental results demonstrate its performance at 0.16 compression ratio is comparative to existing methods in the original domain.

During recent years, with the rapid development of wireless sensor technologies, Wireless Sensor Network (WSN) systems are widely used in various applications, such as environmental monitoring, smart home and so on. A WSN system is a self-configuring network of sensor nodes connected by wireless links. The union of sensor nodes forms an arbitrary topology. Routing in a WSN system is a critical task. Therefore, the routing protocol has much effect on the throughout performance and energy consumption of the whole network. Most available routing protocols employ distributed structure that is complex and has high energy consumption. Compared with these distributed protocols, a novel centralized routing protocol with a central node is proposed. And all kinds of routing management of the whole WSN systems are centralized in the central node. The centralized management strategy reduces routing overhead and energy consumption of the whole network. Finally, a WSN system with Raspberry Pi and OpenWrt is designed to evaluate the new method. Experiment results show that it is very efficient.

Neural network based keyword spotting (KWS) system is a critical component for user interaction in current smart devices. Although small-footprint networks have been widely explored to reduce deployment overhead, low-precision input feature representation still lacks in-depth research. In this letter, an error-diffusion based speech feature quantization method is proposed. Specifically, our algorithm adapts image processing to quantize the input speech feature maps in arbitrary bits. Experiments show that in the 10-keyword KWS task, our 3-bit representation only brings a 0.45% average accuracy drop compared to the full-precision log-Mel spectrograms while others drop over 3%. In the 2 keywords task, our 3-bit representation produces no significant differences, while 1-bit quantization only leads to an average of 1.7% accuracy drop and is even capable of handling similar keywords and imbalanced data distribution. The result proves our method, to the best of our knowledge, is the first practical method that supports as low as 1-bit quantization for single-channel speech features in small-footprint KWS. In addition, we analyze the impact of error-diffusion directions and conclude that time-direction diffusion is more suitable for temporal convolutional networks.

With the rise of IoT and edge computing, deploying neural networks (NNs) on low-power edge computing devices is drawing more and more attention. In NNs, convolutional layers take up the majority of the computing cycles, especially when NNs are implemented on ARM processors [1]. Therefore, it is necessary to optimize the convolutional implementation on ARM Cortex-M MCUs. This paper proposes an efficient im2row-based fast convolution algorithm with two innovations. First, a novel im2row method for reusing the data of adjacent convolutional windows is presented. This method utilizes a reusable im2row buffer for data reuse, significantly reducing the amount of data copied during im2row and improving efficiency. Second, in algorithm implementation, a q7_t to q15_t data type extension technique that avoids data reordering is employed. This technique eliminates data reordering instructions, thus reducing the runtime of the algorithm. We evaluate our algorithm in separate convolutional layers and NNs. The results for convolutional layers show that, compared to baseline, the proposed algorithm speeds up the convolutional layer by an average of 1.42x, and the maximum speedup is up to 2.9x. Experiments on different NNs demonstrate that our algorithm can speed up the overall NN by up to 2.15x.

Keyword spotting system (KWS) is considered as an essential human-machine interface in multifarious edge devices. To extend the standby time of battery-driven devices, KWS system is required to work with ultra-low power consumption. However, conventional KWS systems requires time-frequency transform for feature extractor (FE), leading to high power consumption along with high-latency. This paper proposes a lightweight FE architecture without time-frequency transform, named Time-domain Filter Bank (TFBank), to tremendously reduce power consumption and latency. A post-framing method is also proposed to eliminate redundant computation between contiguous frames and approximate computing is used to mitigate power overload caused by multiplying. Simulation results show that TFBank consumes 387nW with 0.8V55nm. The system latency is 0.5ms and the memory size usage is 1.175Kb, achieving up to 16-96× speedup and 4-10× memory saving compared to the state-of-the-art designs.