The convolutional neural network (CNN) is one of the most used deep learning models for image detection and classification, due to its high accuracy when compared to other machine learning algorithms. CNNs achieve better results at the cost of higher computing and memory requirements. Inference of convolutional neural networks is therefore usually done in centralized high-performance platforms. However, many applications based on CNNs are migrating to edge devices near the source of data due to the unreliability of a transmission channel in exchanging data with a central server, the uncertainty about channel latency not tolerated by many applications, security and data privacy, etc. While advantageous, deep learning on edge is quite challenging because edge devices are usually limited in terms of performance, cost, and energy. Reconfigurable computing is being considered for inference on edge due to its high performance and energy efficiency while keeping a high hardware flexibility that allows for the easy adaption of the target computing platform to the CNN model. In this paper, we described the features of the most common CNNs, the capabilities of reconfigurable computing for running CNNs, the state-of-the-art of reconfigurable computing implementations proposed to run CNN models, as well as the trends and challenges for future edge reconfigurable platforms.

Deep neural networks have recently shown great results in a vast set of image applications. The associated deep learning models are computationally very demanding and, therefore, several hardware solutions have been proposed to accelerate their computation. FPGAs have recently shown very good performances for these kind of applications and so it is considered a promising platform to accelerate the execution of deep learning algorithms. A common operation in these algorithms is multiply-accumulate (MACC) that is used to calculate dot-products. Since many dot products can be calculated in parallel, as long as memory bandwidth is available, it is very important to implement this operation very efficiently to increase the density of MACC units in an FPGA. In this paper, we propose an implementation of parallel MACC units in FPGA for dot-product operations with very high performance/area ratios using a mix of DSP blocks and LUTs. We consider fixed-point representations with 8 bits of size, but the method can be applied to other bit widths. The method allows us to achieve TOPs performances, even for low cost FPGAs.

Floating-point computing with more than one TFLOP of peak performance is already a reality in recent Field-Programmable Gate Arrays (FPGA). General-Purpose Graphics Processing Units (GPGPU) and recent many-core CPUs have also taken advantage of the recent technological innovations in integrated circuit (IC) design and had also dramatically improved their peak performances. In this paper, we compare the trends of these computing architectures for high-performance computing and survey these platforms in the execution of algorithms belonging to different scientific application domains. Trends in peak performance, power consumption and sustained performances, for particular applications, show that FPGAs are increasing the gap to GPUs and many-core CPUs moving them away from high-performance computing with intensive floating-point calculations. FPGAs become competitive for custom floating-point or fixed-point representations, for smaller input sizes of certain algorithms, for combinational logic problems and parallel map-reduce problems. © 2014 Technical University of Munich (TUM).

O aumento constante da densidade de integração e do desempenho dos circuitos integrados reconfiguráveis permitiram a implementação num único integrado de sistemas capazes de suportar os requisitos bastante exigentes dos sistemas embebidos nas áreas de multimédia e de telecomunicações. Além disso, a flexibilidade destes sistemas permite a utilização de metodologias de projecto baseadas em plataforma com reutilização de IP, com a consequente redução do tempo para o mercado. Neste trabalho, é proposta uma metodologia de co-síntese para o projecto de plataformas SoC configuráveis com uma arquitectura de rede intra-chip para execução de sistemas embebidos predominantemente de processamento de dados. A metodologia foi testada com o projecto de um algoritmo de compressão de imagem em FPGA que atingiu uma taxa de processamento de 800Mbps. Os resultados são bastante promissores, uma vez que foi possível integrar facilmente vários núcleos IP num único circuito reconfigurável e obter soluções de elevada qualidade.

Edge devices are becoming smarter with the integration of machine learning methods, such as deep learning, and are therefore used in many application domains where decisions have to be made without human intervention. Deep learning and, in particular, convolutional neural networks (CNN) are more efficient than previous algorithms for several computer vision applications such as security and surveillance, where image and video analysis are required. This better efficiency comes with a cost of high computation and memory requirements. Hence, running CNNs in embedded computing devices is a challenge for both algorithm and hardware designers. New processing devices, dedicated system architectures and optimization of the networks have been researched to deal with these computation requirements. In this paper, we improve the inference execution times of CNNs in low density FPGAs (Field-Programmable Gate Arrays) using fixed-point arithmetic, zero-skipping and weight pruning. The developed architecture supports the execution of large CNNs in FPGA devices with reduced on-chip memory and computing resources. With the proposed architecture, it is possible to infer an image in AlexNet in 2.9 ms in a ZYNQ7020 and 1.0 ms in a ZYNQ7045 with less than 1% accuracy degradation. These results improve previous state-of-the-art architectures for CNN inference.

Convolutional Neural Networks (CNNs) have improved several computer vision applications, like object detection and classification, when compared to other machine learning algorithms. Running these models in edge computing devices close to data sources is attracting the attention of the community since it avoids high-latency data communication of private data for cloud processing and permits real-time decisions turning these systems into smart embedded devices. Running these models is computationally very demanding and requires a large amount of memory, which are scarce in edge devices compared to a cloud center. In this paper, we proposed an architecture for the inference of pruned convolutional neural networks in any density FPGAs. A configurable block pruning method is proposed together with an architecture that supports the efficient execution of pruned networks. Also, pruning and batching are studied together to determine how they influence each other. With the proposed architecture, we run the inference of a CNN with an average performance of 322 GOPs for 8-bit data in a XC7Z020 FPGA. The proposed architecture running AlexNet processes 240 images/s in a ZYNQ7020 and 775 images/s in a ZYNQ7045 with only 1.2% accuracy degradation.

Financial and commercial applications depend on decimal arithmetic because they must produce results that match exactly those obtained by human calculations. Decimal multiplication is a frequently used operation in these applications and also in the design of decimal floating-point units. In this paper we propose a new architecture for parallel decimal multiplication that improves the area of previous decimal multipliers while keeping the best performances. A decimal adder [1] based on a mixed BCD/excess-6 representation of the operands is utilized. A new partial product generation unit is proposed based on a 5221 recoding of the multiplier digits. With the proposed multiplier, we are able to improve on state-of-the-art parallel decimal multipliers targeting LUT-6 FPGAs. Compared to previous decimal multipliers, implementation results for 2, 4, 8, 16, 32 and 34-digits show that the proposed multiplier achieves over 20% better area without performance degradation.

Deep learning is now present in a wide range of services and applications, replacing and complementing other machine learning algorithms. Performing training and inference of deep neural networks using the cloud computing model is not viable for applications where low latency is required. Furthermore, the rapid proliferation of the Internet of Things will generate a large volume of data to be processed, which will soon overload the capacity of cloud servers. One solution is to process the data at the edge devices themselves, in order to alleviate cloud server workloads and improve latency. However, edge devices are less powerful than cloud servers, and many are subject to energy constraints. Hence, new resource and energy-oriented deep learning models are required, as well as new computing platforms. This paper reviews the main research directions for edge computing deep learning algorithms.

Deep learning and, in particular, convolutional neural networks (CNN) achieve very good results on several computer vision applications like security and surveillance, where image and video analysis are required. These networks are quite demanding in terms of computation and memory and therefore are usually implemented in high-performance computing platforms or devices. Running CNNs in embedded platforms or devices with low computational and memory resources requires a careful optimization of system architectures and algorithms to obtain very efficient designs. In this context, Field Programmable Gate Arrays (FPGA) can achieve this efficiency since the programmable hardware fabric can be tailored for each specific network. In this paper, a very efficient configurable architecture for CNN inference targeting any density FPGAs is described. The architecture considers fixed-point arithmetic and image batch to reduce computational, memory and memory bandwidth requirements without compromising network accuracy. The developed architecture supports the execution of large CNNs in any FPGA devices including those with small on-chip memory size and logic resources. With the proposed architecture, it is possible to infer an image in AlexNet in 4.3 ms in a ZYNQ7020 and 1.2 ms in a ZYNQ7045.

Financial and commercial data are mostly represented in decimal format. To avoid errors introduced when converting some decimal fractions to binary, these data are processed with decimal arithmetic. Most processors only have hardwired binary arithmetic units. So, decimal operations are executed with slow software-based decimal arithmetic functions. For the fast execution of decimal operations, dedicated hardware units have been proposed and designed in FPGA. Decimal multiplication is found in most decimal-based applications and so its optimized design is very important for fast execution. In this paper two new parallel decimal multipliers in FPGA are proposed. These are based on a new decimal adder/subtractor also proposed in this paper. The new decimal multipliers improve state-of-the-art parallel decimal multipliers. Compared to previous architectures, implementation results show that the proposed multipliers achieve 26% better area and 12% better performance. Also, the new decimal multipliers reduce the area and performance gap to binary multipliers and are smaller for 32 digit operands.