Accelerating convolutional neural networks: Exploring FPGA-based architectures and challenges

This article proposes an uninterrupted processing technique for the convolutional neural network (CNN) accelerator. It primarily allows the CNN accelerator to simultaneously perform both processing element (PE) operation and data fetching that reduces its latency and enhances the achievable throughput. Corresponding to the suggested technique, this work also presents a low latency VLSI-architecture of the CNN accelerator using the new random access line-buffer (RALB)-based design of PE array. Subsequently, the proposed CNN-accelerator architecture has been further optimized by reusing the local data in PE array, incurring better energy conservation. Our CNN accelerator has been hardware implemented on Zynq-UltraScale + MPSoC-ZCU102 FPGA board, and it operates at a maximum clock frequency of 340 MHz, consuming 4.11 W of total power. In addition, the suggested CNN accelerator with 864 PEs delivers a peak throughput of 587.52 GOPs and an adequate energy efficiency of 142.95 GOPs/W. Comparison of aforementioned implementation results with the literature has shown that our CNN accelerator delivers 33.42% higher throughput and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$6.24\times $ </tex-math></inline-formula> better energy efficiency than the state-of-the-art work. Eventually, the field-programmable gate array (FPGA) prototype of the proposed CNN accelerator has been functionally validated using the real-world test setup for the detection of object from input image, using the GoogLeNet neural network.

We empirically evaluate an undervolting technique, i.e., underscaling the\ncircuit supply voltage below the nominal level, to improve the power-efficiency\nof Convolutional Neural Network (CNN) accelerators mapped to Field Programmable\nGate Arrays (FPGAs). Undervolting below a safe voltage level can lead to timing\nfaults due to excessive circuit latency increase. We evaluate the\nreliability-power trade-off for such accelerators. Specifically, we\nexperimentally study the reduced-voltage operation of multiple components of\nreal FPGAs, characterize the corresponding reliability behavior of CNN\naccelerators, propose techniques to minimize the drawbacks of reduced-voltage\noperation, and combine undervolting with architectural CNN optimization\ntechniques, i.e., quantization and pruning. We investigate the effect of\nenvironmental temperature on the reliability-power trade-off of such\naccelerators. We perform experiments on three identical samples of modern\nXilinx ZCU102 FPGA platforms with five state-of-the-art image classification\nCNN benchmarks. This approach allows us to study the effects of our\nundervolting technique for both software and hardware variability. We achieve\nmore than 3X power-efficiency (GOPs/W) gain via undervolting. 2.6X of this gain\nis the result of eliminating the voltage guardband region, i.e., the safe\nvoltage region below the nominal level that is set by FPGA vendor to ensure\ncorrect functionality in worst-case environmental and circuit conditions. 43%\nof the power-efficiency gain is due to further undervolting below the\nguardband, which comes at the cost of accuracy loss in the CNN accelerator. We\nevaluate an effective frequency underscaling technique that prevents this\naccuracy loss, and find that it reduces the power-efficiency gain from 43% to\n25%.\n

Convolutional neural networks (CNNs) have achieved great success in many applications. Recently, various FPGA-based accelerators have been proposed to improve the performance of CNNs. However, current most FPGA-based methods use single bit-width selection for all CNN layers, which lead to very low resource utilization efficiency and difficulty in further performance improvement. In this paper, we propose a new approach utilizing bit-width partitioning of FPGA DSP resources to improve the performance and resource utilization efficiency of CNN accelerator. Moreover, we use optimization approach to find the optimal allocation plan for DSP resources. On a Xilinx Virtex-7 FPGA, our design approach achieves performance over the state-of-the-art FPGA-based CNN accelerators from 5.48x to 7.25x and by 6.21x on average, when we evaluate the popular CNNs.

Convolutional neural network (CNN) is widely applied to image recognition with high recognition accuracy. CNN has a wider implementation in general-purpose processors and can be accelerated on FPGA. CNN has a unique way of computing, but general-purpose processors are not efficient for CNN and cannot meet energy efficiency requirements. And the previous studies on FPGA did not involve an energy-efficient implementation on FPGA. We innovatively propose energy efficiency models and implement high energy efficiency CNN on FPGA. We implemented the LeNet-5 network model on the GENESYS 2 board and compared it to the traditional processor and previous studies. By comparison, the computing throughput of CPU, GPU and FPGA are 3.831GFLOPS, 27.143GFLOPS and 19.61GFLOPS respectively, and their powers are 32.15W, 52W, 4.152W respectively. The final energy efficiency (GFLOPS/W) is 0.119GFLOPS/W, 0.522 GFLOPS/W, 4.723 GFLOPS/W, so the energy efficiency of FPGA are far superior to that of CPU and GPU. Since the energy efficiency we achieved on FPGA is also higher than that of FPL2009 and FPGA2015, and we have achieved good experimental results in energy efficiency.

Convolutional Neural Network (CNN) has achieved great success in various fields of machine vision, such as image classification and recognition, image segmentation, and video analysis. In specific applications, it is often necessary to customize the network structure by altering image size, convolution kernel size, pooling size, network architecture, and the number of network layers. These customizations pose significant challenges to the architecture of CNN accelerator especially in real-time systems, and as the number of these network variables increases, the internal convolution calculations become extremely demanding. The inference speed and energy consumption of CNN are becoming more and more important, which requires the design of the accelerator to adapt to different network architecture and run efficiently. Field-programmable gate arrays (FPGAs) is an ideal choice for a CNN accelerator due to its high programmability and low power consumption. To address the aforementioned challenges, we present NRUCA, a novel network reconfigurable universal CNN accelerator based on FPGA. On one hand, we have designed a flexible architecture that can dynamically configure CNN’s network structure parameters. These parameters can be sent to the FPGA via a configuration file, enabling the same design to run different CNN networks and allowing to modify on the fly. This forms the basis of the architecture’s adaptability presented in this paper. On the other hand, we employ a multi-channel parallel and efficient pipeline matrix multiplication architecture to implement the convolution and fully-connected layers, which constitute the majority of the computational load in CNNs. By utilizing an innovative inverted matrix multiplication algorithm and an interleaving cache data method, we reduce the internal cache required for multi-channel data. Furthermore, most of the intermediate calculation data does not need to be output to DDR, significantly improving the operation efficiency of the accelerator. We also fully leverage the FPGA chip architecture to compile multiple calculation kernels, which can be flexibly scheduled and combined to cater to various application scenarios. Based on the flexible and efficient overall design architecture, the layout of various resources of FPGA is balanced, which enables the designed FPGA project to compile four acceleration kernels with the running clock up to 300MHz in Xilinx Vitis and test them on Xilinx Alveo U250. We verified the adaptability and acceleration capabilities of our FPGA project using LeNet, AlexNet, VGG11, VGG13, VGG16, and VGG19 networks. Experiments demonstrated that our FPGA architecture achieves 33X and 35X energy savings compared to the Intel Xeon 5220R CPU, and 1.08X and 1.05X energy savings compared to the Nvidia Tesla P100 GPU, when accelerating AlexNet and VGG16, respectively.

In recent years, convolutional neural networks (CNNs) based machine learning algorithms have been widely applied in computer vision applications. However, for large-scale CNNs, the computation-intensive, memory-intensive and resource-consuming features have brought many challenges to CNN implementations. This work proposes an end-to-end FPGA-based CNN accelerator with all the layers mapped on one chip so that different layers can work concurrently in a pipelined structure to increase the throughput. A methodology which can find the optimized parallelism strategy for each layer is proposed to achieve high throughput and high resource utilization. In addition, a batch-based computing method is implemented and applied on fully connected layers (FC layers) to increase the memory bandwidth utilization due to the memory-intensive feature. Further, by applying two different computing patterns on FC layers, the required on-chip buffers can be reduced significantly. As a case study, a state-of-the-art large-scale CNN, AlexNet, is implemented on Xilinx VC709. It can achieve a peak performance of 565.94 GOP/s and 391 FPS under 156MHz clock frequency which outperforms previous approaches.

Convolutional neural network (CNN) accelerators have achieved great success from cloud to edge scenarios. However, given the trend towards even larger and deeper neural network models, it remains a challenging problem to efficiently process these CNNs especially on edge devices with limited energy budget. Accordingly, reducing the energy consumption is of paramount importance for sustainable CNN accelerators. In this paper, we propose AdaPrune, a novel pruning technique that reduces model size and computation to achieve performance improvement and energy savings for CNN accelerators. Unlike previous pruning techniques that sacrifice either computational regularity or accuracy, AdaPrune maintains both by customizing CNN pruning for the underlying accelerators to maximally leverage the sparsity benefits. AdaPrune consists of two techniques: input channel group pruning and output channel group pruning. By analyzing the weight fetching patterns of sparse CNN accelerators, AdaPrune adaptively switches between the two techniques to guarantee that the zeros are evenly distributed in each fetching group. In doing so, the pruned network structure preserves customized computational regularity for the underlying accelerators, thereby boosting the performance and energy efficiency. We evaluate AdaPrune on three sparse CNN accelerators with different spatial tiling strategies. The experimental results show that AdaPrune achieves up to 1.6&#x00D7; performance speedup, and 1.5&#x00D7; energy savings compared to unstructured pruning.

Convolution neural networks (CNNs) as one of today's main flavor of deep learning techniques dominate in various image recognition tasks. As the model size of modern CNNs continues to grow, neural network compression techniques have been proposed to prune the redundant neurons and synapses. However, prior techniques disconnect the software neural networks compression and hardware acceleration, which fail to balance multiple design parameters, including sparsity, performance, hardware area cost, and efficiency. More concretely, prior unstructured pruning techniques achieve high sparsity at the expense of extra performance overhead, while prior structured pruning techniques relying on strict sparse patterns lead to low sparsity and extra hardware cost. In this article, we propose OMNI, a framework for accelerating sparse CNNs on hardware accelerators. The innovation of OMNI stems from that it uses hardware amenable on-chip memory partition patterns to seamlessly engage the software CNN model compression and hardware CNN acceleration. To accelerate the compute-intensive convolution kernel, a promising hardware optimization approach is memory partition, which divides the original weight kernels into several groups so that the different hardware processing elements can simultaneously access the weight. We exploit the memory partition patterns including block, cyclic, or hybrid as a means of CNN compression patterns. Our software CNN model compression balances the sparsity across different groups and our hardware accelerator employs hardware parallelization coordinately with the sparse patterns, leading to a desirable compromise between sparsity and performance. We further develop performance models to help the designers to quickly identify the pattern factors subject to an area constraint. Last, we evaluate our design on application specific integrated circuit (ASIC) and field-programmable gate array (FPGA) platform. Experiments demonstrate that OMNI achieves 3.4×- 6.2× speedup for the modern CNNs, over a comparably ideal dense CNN accelerator. OMNI shows 114.7× energy efficiency improvement compared with GPU platform. OMNI is also evaluated on Xilinx ZC706 and ZCU102 FPGA platforms, achieving 41.5 GOP/s and 125.3 GOP/s, respectively.

Convolutional neural network(CNN) and related deep learning algorithms represent the state-of-art ability in several computer vision tasks, such as image classification and video analysis. However, model complexity and massive computational operations restrain CNN deployment on embedded systems with limited computing capability and low power budgets including smartphones, drones and robots. In this paper, we propose a scalable FPGA-based CNN hardware accelerator for embedded systems based on 8-bit fixed point approximation of a hardware-friendly CNN model named SqueezeNet v1.1. Proposed FPGA accelerator is implemented with OpenCL framework and achieves 1.9x energy efficiency compared to previous work. Besides, performance and resource occupation of proposed FPGA accelerator are capable of being customized to fit in different FPGA platforms by changing a single hyperparameter.

With the ever-increasing compute demands of artificial intelligence (AI) workloads, there is extensive interest in leveraging field-programmable gate-arrays (FPGAs) to quickly deploy hardware accelerators for the latest convolutional neural networks (CNNs). Recent FPGA architectures are also evolving to better serve the needs of AI, but accelerators need extensive re-design to leverage these new features. The Stratix 10 NX chip by Intel is a new FPGA that replaces traditional DSP blocks with in-fabric AI tensor blocks that provide 15x more multipliers and up to 143 TOPS of performance, at the cost of lower precision (INT8) and significant restrictions on how many operands can be fed to the multipliers from the programmable routing. In this paper, we explore different CNN accelerator structures to leverage the tensor blocks, considering the various tensor block modes, operand bandwidth restrictions, and on-chip memory restrictions. We incorporate the most performant techniques into HPIPE, a layer-pipelined and sparse-aware CNN accelerator for FPGAs. We enhance HPIPE's software compiler to restructure the CNN computations and on-chip memory layout to take advantage of the additional multipliers offered by the new tensor block architecture, while also avoiding stalls due to data loading restrictions. We achieve cycle-by-cycle speedups in tensor mode of up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8}.\mathbf{3}\mathbf{x}$</tex> for Mobilenet-v1 versus the original HPIPE design using conventional DSPs. On the FPGA, we achieve a throughput of 28,541 and 29,429 images/s on Mobilenet-v1 and Mobilenet-v2 respectively, outperforming all previous FPGA accelerators by at least 4.0x, including one on an AI-optimized Xilinx chip. We also outperform NVIDIA's V100 GPU, a machine learning targeted GPU on a similar process node with a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1}.\mathbf{7}\mathbf{x}$</tex> larger die size, by up to 17x with a batch size of one and 1.3x with NVIDIA's largest reported batch size of 128.

Convolutional Neural Network (CNN) accelerators are popular specialized platforms for efficient CNN processing. As semiconductor manufacturing technology scales down to nano scale, process variation dramatically affects the chip's quality. Process variation causes delay variation within the chip due to transistor parameter differences. CNN accelerators adopt a large number of Processing Elements (PEs) for parallel computing, which are highly susceptible to process variation effects. Fast CNN processing desires consistent performance among PEs, otherwise the processing speed is limited by the slowest PE within the chip. In this work, we first quantitatively model and analyze the impact of process variation on CNN accelerator's operating frequency. We further analyze the utilization of CNN accelerator and the characteristics of CNN models. We then leverage the PE underutilization to propose a sub-matrix reformation mechanism and leverage the pixel similarity of images to propose a weight transfer technique. Both techniques are able to tolerate the low-frequency PEs, and achieve performance improvement at chip level. Evaluation results show our techniques are able to achieve significant processing speed improvement with negligible accuracy loss.

Field programmable gate array (FPGA) is widely considered as a promising platform for convolutional neural network (CNN) acceleration. However, the large numbers of parameters of CNNs cause heavy computing and memory burdens for FPGA-based CNN implementation. To solve this problem, this paper proposes an optimized compression strategy, and realizes an accelerator based on FPGA for CNNs. Firstly, a reversed-pruning strategy is proposed which reduces the number of parameters of AlexNet by a factor of 13× without accuracy loss on the ImageNet dataset. Peak-pruning is further introduced to achieve better compressibility. Moreover, quantization gives another 4× with negligible loss of accuracy. Secondly, an efficient storage technique, which aims for the reduction of the whole overhead cache of the convolutional layer and the fully connected layer, is presented respectively. Finally, the effectiveness of the proposed strategy is verified by an accelerator implemented on a Xilinx ZCU104 evaluation board. By improving existing pruning techniques and the storage format of sparse data, we significantly reduce the size of AlexNet by 28×, from 243 MB to 8.7 MB. In addition, the overall performance of our accelerator achieves 9.73 fps for the compressed AlexNet. Compared with the central processing unit (CPU) and graphics processing unit (GPU) platforms, our implementation achieves 182.3× and 1.1× improvements in latency and throughput, respectively, on the convolutional (CONV) layers of AlexNet, with an 822.0× and 15.8× improvement for energy efficiency, separately. This novel compression strategy provides a reference for other neural network applications, including CNNs, long short-term memory (LSTM), and recurrent neural networks (RNNs).

Edge AI systems are increasingly being adopted in a wide range of application domains. These systems typically deploy Convolutional Neural Network (CNN) models on edge devices to perform inference, while relying on the cloud for model training. This is due to the high computational and memory demands of conventional model training, which exceeds the capabilities of resource-constrained edge devices running on tight power budgets. The dependency on the cloud for training is not suitable in many applications, where new objects or environmental conditions different from the ones present during training, are frequently encountered. In such applications, continual learning of new knowledge on the edge device becomes a necessity to avoid performance bottlenecks due to round-trip communication delays, network connectivity, and the available bandwidth. In this thesis, we propose Field-Programmable Gate Array (FPGA) based accelerator architecture and optimization strategies for a new paradigm of machine learning algorithms that is capable of continual learning. The proposed methods will enable edge FPGA systems to perform on-device deep continual learning for object classification. Specifically, the proposed methods aim to achieve real-time learning on-device, while providing for a high degree of scalability to learn a large number of classes. We first propose a FPGA accelerator for a Self-Organizing Neural Network (SONN), that can perform class-incremental continual learning in a streaming manner when combined with a CNN. The SONN model performs unsupervised learning from embedding features extracted from the CNN model by dynamically growing neurons and connections. We introduce design optimization strategies and runtime scheduling techniques to optimize resource usage, latency, and energy consumption. Experimental results based on Core50 dataset for continuous object recognition from video sequences, demonstrated that the proposed FPGA architecture outperforms CPU and GPU based counterparts in terms of latency and power. However, the SONN model grows proportionally to the classes learnt, which limits its scalability to learn a large number of classes efficiently. Next, we propose a FPGA accelerator for a Streaming Linear Discriminant Analysis (SLDA) model to overcome the scalability limitations of SONN. Similar to the SONN, SLDA performs continual learning from embedding features extracted from a CNN in a streaming manner. SLDA is highly scalable for learning a large number of classes as the network does not grow dynamically to accommodate new knowledge. We propose several design and runtime optimizations to minimize resource usage, latency, and energy consumption. Additionally, we introduce a new variant of SLDA and discuss the accuracy-efficiency trade-offs using popular datasets for continual learning, CoRE50 and CUB200. The results demonstrate that the proposed SLDA accelerator outperforms CPU and GPU counterparts, in terms of latency and energy efficiency. Finally, we demonstrate a full on-chip deep continual learning pipeline on FPGA, by integrating the proposed SLDA accelerator with Xilinx DPU, a programmable CNN accelerator IP. The design is implemented on a Xilinx Zynq Ultrascale+ MPSoC. In order to overcome the large performance bottleneck due to the communication overhead between the ARM processing system (PS) and programmable logic (PL), we implemented a Linux device driver that facilitates efficient memory mapping between PS-PL. The experimental results based on CoRE50 dataset show that the proposed pipeline is capable of performing continual learning at nearly the same latency as the inference pipeline, with only a marginal increase in energy consumption. Our results clearly demonstrate the viability of deploying real-time deep continual learning on edge AI systems that are equipped with FPGA accelerators.

Convolutional neural network (CNN) has become a successful algorithm in the region of artificial intelligence and a strong candidate for many computer vision algorithms. But the computation complexity of CNN is much higher than traditional algorithms. With the help of GPU acceleration, CNN-based applications are widely deployed in servers. However, for embedded platforms, CNN-based solutions are still too complex to be applied. Various dedicated hardware designs on field-programmable gate arrays (FPGAs) have been carried out to accelerate CNNs, while few of them explore the whole design flow for both fast deployment and high power efficiency. In this paper, we investigate state-of-the-art CNN models and CNN-based applications. Requirements on memory, computation and the flexibility of the system are summarized for mapping CNN on embedded FPGAs. Based on these requirements, we propose Angel-Eye, a programmable and flexible CNN accelerator architecture, together with data quantization strategy and compilation tool. Data quantization strategy helps reduce the bit-width down to 8-bit with negligible accuracy loss. The compilation tool maps a certain CNN model efficiently onto hardware. Evaluated on Zynq XC7Z045 platform, Angel-Eye is $6 {\times }$ faster and $5{\times }$ better in power efficiency than peer FPGA implementation on the same platform. Applications of VGG network, pedestrian detection and face alignment are used to evaluate our design on Zynq XC7Z020. NIVIDA TK1 and TX1 platforms are used for comparison. Angel-Eye achieves similar performance and delivers up to $16 {\times }$ better energy efficiency.

Convolutional neural network (CNN) based machine learning requires a highly parallel as well as low power consumption (including leakage power) hardware accelerator. In this paper, we will present a digital ReRAM crossbar based CNN accelerator that can achieve significantly higher throughput and lower power consumption than state-of-arts. The CNN is trained with binary constraints on both weights and activations such that all operations become bitwise. With further use of 1-bit comparator, the bitwise CNN model can be naturally realized on a digital ReRAM-crossbar device. A novel sneak-path-free ReRAM-crossbar is further utilized for large-scale realization. Simulation experiments show that the bitwise CNN accelerator on the digital ReRAM crossbar achieves 98.3% and 91.4% accuracy on MNIST and CIFAR-10 benchmarks, respectively. Moreover, it has a peak throughput of 792GOPS at the power consumption of 6.3mW, which is 18.86 times higher throughput and 44.1 times lower power than CMOS CNN (non-binary) accelerators.