The emerging mobile robot industry has spurred a flurry of interest in solving the simultaneous localization and mapping(SLAM)problem.However,existing SLAM platforms have difficulty in meeting the real-time and low-po...The emerging mobile robot industry has spurred a flurry of interest in solving the simultaneous localization and mapping(SLAM)problem.However,existing SLAM platforms have difficulty in meeting the real-time and low-pow-er requirements imposed by mobile systems.Though specialized hardware is promising with regard to achieving high per-formance and lowering the power,designing an efficient accelerator for SLAM is severely hindered by a wide variety of SLAM algorithms.Based on our detailed analysis of representative SLAM algorithms,we observe that SLAM algorithms advance two challenges for designing efficient hardware accelerators:the large number of computational primitives and ir-regular control flows.To address these two challenges,we propose a hardware accelerator that features composable com-putation units classified as the matrix,vector,scalar,and control units.In addition,we design a hierarchical instruction set for coping with a broad range of SLAM algorithms with irregular control flows.Experimental results show that,com-pared against an Intel x86 processor,on average,our accelerator with the area of 7.41 mm^(2) achieves 10.52x and 112.62x better performance and energy savings,respectively,across different datasets.Compared against a more energy-efficient ARM Cortex processor,our accelerator still achieves 33.03x and 62.64x better performance and energy savings,respec-tively.展开更多
Massive computational complexity and memory requirement of artificial intelligence models impede their deploy-ability on edge computing devices of the Internet of Things(IoT).While Power-of-Two(PoT)quantization is pro...Massive computational complexity and memory requirement of artificial intelligence models impede their deploy-ability on edge computing devices of the Internet of Things(IoT).While Power-of-Two(PoT)quantization is pro-posed to improve the efficiency for edge inference of Deep Neural Networks(DNNs),existing PoT schemes require a huge amount of bit-wise manipulation and have large memory overhead,and their efficiency is bounded by the bottleneck of computation latency and memory footprint.To tackle this challenge,we present an efficient inference approach on the basis of PoT quantization and model compression.An integer-only scalar PoT quantization(IOS-PoT)is designed jointly with a distribution loss regularizer,wherein the regularizer minimizes quantization errors and training disturbances.Additionally,two-stage model compression is developed to effectively reduce memory requirement,and alleviate bandwidth usage in communications of networked heterogenous learning systems.The product look-up table(P-LUT)inference scheme is leveraged to replace bit-shifting with only indexing and addition operations for achieving low-latency computation and implementing efficient edge accelerators.Finally,comprehensive experiments on Residual Networks(ResNets)and efficient architectures with Canadian Institute for Advanced Research(CIFAR),ImageNet,and Real-world Affective Faces Database(RAF-DB)datasets,indicate that our approach achieves 2×∼10×improvement in the reduction of both weight size and computation cost in comparison to state-of-the-art methods.A P-LUT accelerator prototype is implemented on the Xilinx KV260 Field Programmable Gate Array(FPGA)platform for accelerating convolution operations,with performance results showing that P-LUT reduces memory footprint by 1.45×,achieves more than 3×power efficiency and 2×resource efficiency,compared to the conventional bit-shifting scheme.展开更多
This paper describes our implementation of several neural networks built on a field programmable gate array (FPGA) and used to recognize a handwritten digit dataset—the Modified National Institute of Standards and Te...This paper describes our implementation of several neural networks built on a field programmable gate array (FPGA) and used to recognize a handwritten digit dataset—the Modified National Institute of Standards and Technology (MNIST) database. We also propose a novel hardware-friendly activation function called the dynamic Rectifid Linear Unit (ReLU)—D-ReLU function that achieves higher performance than traditional activation functions at no cost to accuracy. We built a 2-layer online training multilayer perceptron (MLP) neural network on an FPGA with varying data width. Reducing the data width from 8 to 4 bits only reduces prediction accuracy by 11%, but the FPGA area decreases by 41%. Compared to networks that use the sigmoid functions, our proposed D-ReLU function uses 24% - 41% less area with no loss to prediction accuracy. Further reducing the data width of the 3-layer networks from 8 to 4 bits, the prediction accuracies only decrease by 3% - 5%, with area being reduced by 9% - 28%. Moreover, FPGA solutions have 29 times faster execution time, even despite running at a 60× lower clock rate. Thus, FPGA implementations of neural networks offer a high-performance, low power alternative to traditional software methods, and our novel D-ReLU activation function offers additional improvements to performance and power saving.展开更多
Moving object detection including background subtraction and morphological processing is a critical research topic for video surveillance because of its high computational loading and power consumption. This paper pro...Moving object detection including background subtraction and morphological processing is a critical research topic for video surveillance because of its high computational loading and power consumption. This paper proposes a hardware design to accelerate the computation of background subtraction with low power consumption. A real-time background subtraction method is designed with a frame-buffer scheme and function partition to improve throughput, and implemented using Verilog HDL on FPGA. The design parallelizes the computations of background update and subtraction with a seven-stage pipeline. A stripe-based morphological processing and accounting for the completion of detected objects is devised. Simulation results for videos of VGA resolutions on a low-end FPGA device show 368 fps throughput for only the real-time background subtraction module, and 51 fps for the whole system, including off-chip memory access. Real-time efficiency with low power consumption and low resource utilization is thus demonstrated.展开更多
The Monte Carlo(MC)simulation is regarded as the gold standard for dose calculation in brachytherapy,but it consumes a large amount of computing resources.The development of heterogeneous computing makes it possible t...The Monte Carlo(MC)simulation is regarded as the gold standard for dose calculation in brachytherapy,but it consumes a large amount of computing resources.The development of heterogeneous computing makes it possible to substantially accelerate calculations with hardware accelerators.Accordingly,this study develops a fast MC tool,called THUBrachy,which can be accelerated by several types of hardware accelerators.THUBrachy can simulate photons with energy less than 3 MeV and considers all photon interactions in the energy range.It was benchmarked against the American Association of Physicists in Medicine Task Group No.43 Report using a water phantom and validated with Geant4 using a clinical case.A performance test was conducted using the clinical case,showing that a multicore central processing unit,Intel Xeon Phi,and graphics processing unit(GPU)can efficiently accelerate the simulation.GPU-accelerated THUBrachy is the fastest version,which is 200 times faster than the serial version and approximately 500 times faster than Geant4.The proposed tool shows great potential for fast and accurate dose calculations in clinical applications.展开更多
With the development of computer vision researches, due to the state-of-the-art performance on image and video processing tasks, deep neural network (DNN) has been widely applied in various applications (autonomous ve...With the development of computer vision researches, due to the state-of-the-art performance on image and video processing tasks, deep neural network (DNN) has been widely applied in various applications (autonomous vehicles, weather forecasting, counter-terrorism, surveillance, traffic management, etc.). However, to achieve such performance, DNN models have become increasingly complicated and deeper, and result in heavy computational stress. Thus, it is not sufficient for the general central processing unit (CPU) processors to meet the real-time application requirements. To deal with this bottleneck, research based on hardware acceleration solution for DNN attracts great attention. Specifically, to meet various real-life applications, DNN acceleration solutions mainly focus on issue of hardware acceleration with intense memory and calculation resource. In this paper, a novel resource-saving architecture based on Field Programmable Gate Array (FPGA) is proposed. Due to the novel designed processing element (PE), the proposed architecture </span><span style="font-family:Verdana;">achieves good performance with the extremely limited calculating resource. The on-chip buffer allocation helps enhance resource-saving performance on memory. Moreover, the accelerator improves its performance by exploiting</span> <span style="font-family:Verdana;">the sparsity property of the input feature map. Compared to other state-of-the-art</span><span style="font-family:Verdana;"> solutions based on FPGA, our architecture achieves good performance, with quite limited resource consumption, thus fully meet the requirement of real-time applications.展开更多
Graph processing has been widely used in many scenarios,from scientific computing to artificial intelligence.Graph processing exhibits irregular computational parallelism and random memory accesses,unlike traditional ...Graph processing has been widely used in many scenarios,from scientific computing to artificial intelligence.Graph processing exhibits irregular computational parallelism and random memory accesses,unlike traditional workloads.Therefore,running graph processing workloads on conventional architectures(e.g.,CPUs and GPUs)often shows a significantly low compute-memory ratio with few performance benefits,which can be,in many cases,even slower than a specialized single-thread graph algorithm.While domain-specific hardware designs are essential for graph processing,it is still challenging to transform the hardware capability to performance boost without coupled software codesigns.This article presents a graph processing ecosystem from hardware to software.We start by introducing a series of hardware accelerators as the foundation of this ecosystem.Subsequently,the codesigned parallel graph systems and their distributed techniques are presented to support graph applications.Finally,we introduce our efforts on novel graph applications and hardware architectures.Extensive results show that various graph applications can be efficiently accelerated in this graph processing ecosystem.展开更多
基金supported by the National Natural Science Foundation of China under Grant Nos.61925208,61906179,U19B2019,and U20A20227the Strategic Priority Research Program of Chinese Academy of Sciences under Grant No.XDB32050200+1 种基金Beijing Academy of Artificial Intelligence(BAAI),Chinese Academy of Sciences(CAS)Project for Young Scientists in Basic Research(YSBR-029)Youth Innovation Promotion Association CAS.
文摘The emerging mobile robot industry has spurred a flurry of interest in solving the simultaneous localization and mapping(SLAM)problem.However,existing SLAM platforms have difficulty in meeting the real-time and low-pow-er requirements imposed by mobile systems.Though specialized hardware is promising with regard to achieving high per-formance and lowering the power,designing an efficient accelerator for SLAM is severely hindered by a wide variety of SLAM algorithms.Based on our detailed analysis of representative SLAM algorithms,we observe that SLAM algorithms advance two challenges for designing efficient hardware accelerators:the large number of computational primitives and ir-regular control flows.To address these two challenges,we propose a hardware accelerator that features composable com-putation units classified as the matrix,vector,scalar,and control units.In addition,we design a hierarchical instruction set for coping with a broad range of SLAM algorithms with irregular control flows.Experimental results show that,com-pared against an Intel x86 processor,on average,our accelerator with the area of 7.41 mm^(2) achieves 10.52x and 112.62x better performance and energy savings,respectively,across different datasets.Compared against a more energy-efficient ARM Cortex processor,our accelerator still achieves 33.03x and 62.64x better performance and energy savings,respec-tively.
基金This work was supported by Open Fund Project of State Key Laboratory of Intelligent Vehicle Safety Technology by Grant with No.IVSTSKL-202311Key Projects of Science and Technology Research Programme of Chongqing Municipal Education Commission by Grant with No.KJZD-K202301505+1 种基金Cooperation Project between Chongqing Municipal Undergraduate Universities and Institutes Affiliated to the Chinese Academy of Sciences in 2021 by Grant with No.HZ2021015Chongqing Graduate Student Research Innovation Program by Grant with No.CYS240801.
文摘Massive computational complexity and memory requirement of artificial intelligence models impede their deploy-ability on edge computing devices of the Internet of Things(IoT).While Power-of-Two(PoT)quantization is pro-posed to improve the efficiency for edge inference of Deep Neural Networks(DNNs),existing PoT schemes require a huge amount of bit-wise manipulation and have large memory overhead,and their efficiency is bounded by the bottleneck of computation latency and memory footprint.To tackle this challenge,we present an efficient inference approach on the basis of PoT quantization and model compression.An integer-only scalar PoT quantization(IOS-PoT)is designed jointly with a distribution loss regularizer,wherein the regularizer minimizes quantization errors and training disturbances.Additionally,two-stage model compression is developed to effectively reduce memory requirement,and alleviate bandwidth usage in communications of networked heterogenous learning systems.The product look-up table(P-LUT)inference scheme is leveraged to replace bit-shifting with only indexing and addition operations for achieving low-latency computation and implementing efficient edge accelerators.Finally,comprehensive experiments on Residual Networks(ResNets)and efficient architectures with Canadian Institute for Advanced Research(CIFAR),ImageNet,and Real-world Affective Faces Database(RAF-DB)datasets,indicate that our approach achieves 2×∼10×improvement in the reduction of both weight size and computation cost in comparison to state-of-the-art methods.A P-LUT accelerator prototype is implemented on the Xilinx KV260 Field Programmable Gate Array(FPGA)platform for accelerating convolution operations,with performance results showing that P-LUT reduces memory footprint by 1.45×,achieves more than 3×power efficiency and 2×resource efficiency,compared to the conventional bit-shifting scheme.
文摘This paper describes our implementation of several neural networks built on a field programmable gate array (FPGA) and used to recognize a handwritten digit dataset—the Modified National Institute of Standards and Technology (MNIST) database. We also propose a novel hardware-friendly activation function called the dynamic Rectifid Linear Unit (ReLU)—D-ReLU function that achieves higher performance than traditional activation functions at no cost to accuracy. We built a 2-layer online training multilayer perceptron (MLP) neural network on an FPGA with varying data width. Reducing the data width from 8 to 4 bits only reduces prediction accuracy by 11%, but the FPGA area decreases by 41%. Compared to networks that use the sigmoid functions, our proposed D-ReLU function uses 24% - 41% less area with no loss to prediction accuracy. Further reducing the data width of the 3-layer networks from 8 to 4 bits, the prediction accuracies only decrease by 3% - 5%, with area being reduced by 9% - 28%. Moreover, FPGA solutions have 29 times faster execution time, even despite running at a 60× lower clock rate. Thus, FPGA implementations of neural networks offer a high-performance, low power alternative to traditional software methods, and our novel D-ReLU activation function offers additional improvements to performance and power saving.
文摘Moving object detection including background subtraction and morphological processing is a critical research topic for video surveillance because of its high computational loading and power consumption. This paper proposes a hardware design to accelerate the computation of background subtraction with low power consumption. A real-time background subtraction method is designed with a frame-buffer scheme and function partition to improve throughput, and implemented using Verilog HDL on FPGA. The design parallelizes the computations of background update and subtraction with a seven-stage pipeline. A stripe-based morphological processing and accounting for the completion of detected objects is devised. Simulation results for videos of VGA resolutions on a low-end FPGA device show 368 fps throughput for only the real-time background subtraction module, and 51 fps for the whole system, including off-chip memory access. Real-time efficiency with low power consumption and low resource utilization is thus demonstrated.
基金supported by the National Natural Science Foundation of China(No.11875036)。
文摘The Monte Carlo(MC)simulation is regarded as the gold standard for dose calculation in brachytherapy,but it consumes a large amount of computing resources.The development of heterogeneous computing makes it possible to substantially accelerate calculations with hardware accelerators.Accordingly,this study develops a fast MC tool,called THUBrachy,which can be accelerated by several types of hardware accelerators.THUBrachy can simulate photons with energy less than 3 MeV and considers all photon interactions in the energy range.It was benchmarked against the American Association of Physicists in Medicine Task Group No.43 Report using a water phantom and validated with Geant4 using a clinical case.A performance test was conducted using the clinical case,showing that a multicore central processing unit,Intel Xeon Phi,and graphics processing unit(GPU)can efficiently accelerate the simulation.GPU-accelerated THUBrachy is the fastest version,which is 200 times faster than the serial version and approximately 500 times faster than Geant4.The proposed tool shows great potential for fast and accurate dose calculations in clinical applications.
文摘With the development of computer vision researches, due to the state-of-the-art performance on image and video processing tasks, deep neural network (DNN) has been widely applied in various applications (autonomous vehicles, weather forecasting, counter-terrorism, surveillance, traffic management, etc.). However, to achieve such performance, DNN models have become increasingly complicated and deeper, and result in heavy computational stress. Thus, it is not sufficient for the general central processing unit (CPU) processors to meet the real-time application requirements. To deal with this bottleneck, research based on hardware acceleration solution for DNN attracts great attention. Specifically, to meet various real-life applications, DNN acceleration solutions mainly focus on issue of hardware acceleration with intense memory and calculation resource. In this paper, a novel resource-saving architecture based on Field Programmable Gate Array (FPGA) is proposed. Due to the novel designed processing element (PE), the proposed architecture </span><span style="font-family:Verdana;">achieves good performance with the extremely limited calculating resource. The on-chip buffer allocation helps enhance resource-saving performance on memory. Moreover, the accelerator improves its performance by exploiting</span> <span style="font-family:Verdana;">the sparsity property of the input feature map. Compared to other state-of-the-art</span><span style="font-family:Verdana;"> solutions based on FPGA, our architecture achieves good performance, with quite limited resource consumption, thus fully meet the requirement of real-time applications.
基金supported by the National Key Research and Development Program of China under Grant No.2023YFB4502300.
文摘Graph processing has been widely used in many scenarios,from scientific computing to artificial intelligence.Graph processing exhibits irregular computational parallelism and random memory accesses,unlike traditional workloads.Therefore,running graph processing workloads on conventional architectures(e.g.,CPUs and GPUs)often shows a significantly low compute-memory ratio with few performance benefits,which can be,in many cases,even slower than a specialized single-thread graph algorithm.While domain-specific hardware designs are essential for graph processing,it is still challenging to transform the hardware capability to performance boost without coupled software codesigns.This article presents a graph processing ecosystem from hardware to software.We start by introducing a series of hardware accelerators as the foundation of this ecosystem.Subsequently,the codesigned parallel graph systems and their distributed techniques are presented to support graph applications.Finally,we introduce our efforts on novel graph applications and hardware architectures.Extensive results show that various graph applications can be efficiently accelerated in this graph processing ecosystem.