Tesla K40 Research Articles

New Adversarial Image Detection Based on Sentiment Analysis.

Deep neural networks (DNNs) are vulnerable to adversarial examples, while adversarial attack models, e.g., DeepFool, are on the rise and outrunning adversarial example detection techniques. This article presents a new adversarial example detector that outperforms state-of-the-art detectors in identifying the latest adversarial attacks on image datasets. Specifically, we propose to use sentiment analysis for adversarial example detection, qualified by the progressively manifesting impact of an adversarial perturbation on the hidden-layer feature maps of a DNN under attack. Accordingly, we design a modularized embedding layer with the minimum learnable parameters to embed the hidden-layer feature maps into word vectors and assemble sentences ready for sentiment analysis. Extensive experiments demonstrate that the new detector consistently surpasses the state-of-the-art detection algorithms in detecting the latest attacks launched against ResNet and Inception neutral networks on the CIFAR-10, CIFAR-100, and SVHN datasets. The detector only has about 2 million parameters and takes less than 4.6 ms to detect an adversarial example generated by the latest attack models using a Tesla K80 GPU card.

Reconfigurable origami hologram based on deep neural networks.

Reconfigurable and multifunctional metasurfaces are becoming indispensable in a variety of applications due to their capability to execute diverse functions across various states. However, many of these metasurfaces incorporate complex active components, thereby escalating structural complexity and bulk volume. In this research, we propose a reconfigurable passive hologram based solely on an origami structure, enabling the successful generation of holograms depicting the 'Z' and 'L' illuminated by a right-hand circular polarization (RHCP) wave in two distinct states: planar and zigzag configuration, respectively. The transformation between the 2D planar metasurface and the 3D zigzag structure with slant angles of 35 is achieved solely through mechanically stretching and compressing the origami metasurface. The phases on the origami metasurface are trained through a deep neural network which operates on the NVIDIA Tesla k80 GPU, with the total training process costing 11.88 s after 100 epochs. The reconfigurable scheme proposed in this research provides flexibility and ease of implementation in the fields of imaging and data processing.

BrutNet: A novel approach for violence detection and classification using DCNN with GRU

AbstractAutomatic Violence Detection and Classification (AVDC) with deep learning has garnered significant attention in computer vision research. This paper presents a novel approach for combining a custom Deep Convolutional Neural Network (DCNN) with a Gated Recurrent Unit (GRU) in developing a new AVDC model called BrutNet. Specifically, a time‐distributed DCNN (TD‐DCNN) is developed to generate a compact 2D representation with 512 spatial features per frame from a set of equally‐spaced frames of dimension 16090 in short video segments. Further to leverage the temporal information, a GRU layer is utilised, generating a condensed 1D vector that enables binary classification of violent or non‐violent content through multiple dense layers. Overfitting is addressed by incorporating dropout layers with a rate of 0.5, while the hidden and output layers employ rectified linear unit (ReLU) and sigmoid activations, respectively. The model is trained on the NVIDIA Tesla K80 GPU through Google Colab, demonstrating superior performance compared to existing models across various video datasets, including hockey fights, movie fights, AVD, and RWF‐2000. Notably, the model stands out by requiring only 3.416 million parameters and achieving impressive test accuracies of 97.62%, 100%, 97.22%, and 86.43% on the respective datasets. Thus, BrutNet exhibits the potential to emerge as a highly efficient and robust AVDC model in support of greater public safety, content moderation and censorship, computer‐aided investigations, and law enforcement.

Empowering edge devices: FPGA‐based 16‐bit fixed‐point accelerator with SVD for CNN on 32‐bit memory‐limited systems

AbstractConvolutional neural networks (CNNs) are now often used in deep learning and computer vision applications. Its convolutional layer accounts for most calculations and should be computed fast in a local edge device. Field‐programmable gate arrays (FPGAs) have been adequately explored as promising hardware accelerators for CNNs due to their high performance, energy efficiency, and reconfigurability. This paper developed an efficient FPGA‐based 16‐bit fixed‐point hardware accelerator unit for deep learning applications on the 32‐bit low‐memory edge device (PYNQ‐Z2 board). Additionally, singular value decomposition is applied to the fully connected layer for dimensionality reduction of weight parameters. The accelerator unit was designed for all five layers and employed eight processing elements in convolution layers 1 and 2 for parallel computations. In addition, array partitioning, loop unrolling, and pipelining are the techniques used to increase the speed of calculations. The AXI‐Lite interface was also used to communicate between IP and other blocks. Moreover, the design is tested with grayscale image classification on MNIST handwritten digit dataset and color image classification on the Tumor dataset. The experimental results show that the proposed accelerator unit implementation performs faster than the software‐based implementation. Its inference speed is 89.03% more than INTEL 3‐core CPU, 86.12% higher than Haswell 2‐core CPU, and 82.45% more than NVIDIA Tesla K80 GPU. Furthermore, the throughput of the proposed design is 4.33GOP/s, which is better than the conventional CNN accelerator architectures.

Speeding up the GENGA N-body integrator on consumer-grade graphics cards

Context. Graphics processing unit (GPU) computing has become popular due to the enormous calculation potential that can be harvested from a single card. The N-body integrator Gravitational ENcounters with GPU Acceleration (GENGA) is built to harvest the computing power from such cards, but it suffers a severe performance penalty on consumer-grade Nvidia GPUs due to their artificially truncated double precision performance. Aims. We aim to speed up GENGA on consumer-grade cards by harvesting their high single-precision performance. Methods. We modified GENGA to have the option to compute the mutual long-distance forces between bodies in single precision and tested this with five experiments. First, we ran a high number of simulations with similar initial conditions of on average 6600 fully self-gravitating planetesimals in both single and double precision to establish whether the outcomes were statistically different. These simulations were run on Tesla K20 cards. We supplemented this test with simulations that (i) began with a mixture of planetesimals and planetary embryos, (ii) planetesimal-driven giant planet migration, and (iii) terrestrial planet formation with a dissipating gas disc. All of these simulations served to determine the accuracy of energy and angular momentum conservation under various scenarios with single and double precision forces. Second, we ran the same simulation beginning with 40 000 self-gravitating planetesimals using both single and double precision forces on a variety of consumer-grade and Tesla GPUs to measure the performance boost of computing the long-range forces in single precision. Results. We find that there are no statistical differences when simulations are run with the gravitational forces in single or double precision that can be attributed to the force prescription rather than stochastic effects. The accumulations in uncertainty in energy are almost identical when running with single or double precision long-range forces. However, the uncertainty in the angular momentum using single rather than double precision long-range forces is about two orders of magnitude greater, but still very low. Running the simulations in single precision on consumer-grade cards decreases running time by a factor of three and becomes within a factor of three of a Tesla A100 GPU. Additional tuning speeds up the simulation by a factor of two across all types of cards. Conclusions. The option to compute the long-range forces in single precision in GENGA when using consumer-grade GPUs dramatically improves performance at a little penalty to accuracy. There is an additional environmental benefit because it reduces energy usage.

Fixed-point FPGA Implementation of the FFT Accumulation Method for Real-time Cyclostationary Analysis

The spectral correlation density (SCD) is an important tool in cyclostationary signal detection and classification. Even using efficient techniques based on the fast Fourier transform (FFT), real-time implementations are challenging because of the high computational complexity. A key dimension for computational optimization lies in minimizing the wordlength employed. In this article, we analyze the relationship between wordlength and signal-to-quantization noise in fixed-point implementations of the SCD function. A canonical SCD estimation algorithm, the FFT accumulation method (FAM) using fixed-point arithmetic, is studied. We derive closed-form expressions for SQNR and compare them at wordlengths ranging from 14 to 26 bits. The differences between the calculated SQNR and bit-exact simulations are less than 1 dB. Furthermore, an HLS-based FPGA design is implemented on a Xilinx Zynq UltraScale+ XCZU28DR-2FFVG1517E RFSoC. Using less than 25% of the logic fabric on the device, it consumes 7.7 W total on-chip power and has a power efficiency of 12.4 GOPS/W, which is an order of magnitude improvement over an Nvidia Tesla K40 graphics processing unit (GPU) implementation. In terms of throughput, it achieves 50 MS/sec, which is a speedup of 1.6 over a recent optimized FPGA implementation.

ReSTiNet: An Efficient Deep Learning Approach to Improve Human Detection Accuracy

Human detection is an important task in computer vision. It is one of the most important tasks in global security and safety monitoring. In recent days, Deep Learning has improved human detection technology. Despite modern techniques, there are very few optimal techniques to construct networks with a small size, deep architecture, and fast training time while maintaining accuracy. ReSTiNet is a novel small convolutional neural network that overcomes the problems of network size, detection speed, and accuracy. The developed ReSTiNet contains fire modules by evaluating their number and position in the network to minimize the model parameters and network size. To improve the detection speed and accuracy of ReSTiNet, the residual block within the fire modules is carefully designed to increase the feature propagation and maximize the information flow in the network. The developed approach compresses the well-known Tiny-YOLO architecture while improving the following features: (i) small model size, (ii) faster detection speed, (iii) resolution of overfitting, and (iv) better performance than other compact networks such as SqueezeNet and MobileNet in terms of mAP on the Pascal VOC and MS COCO datasets.ReSTiNet is 10.7 MB, five times smaller than Tiny-YOLO. On Tesla k80, mAP is 27.3% for MS COCO and 63.74% for PASCAL VOC. The validation of the proposed ReSTiNet model has been done on INRIA person dataset using the Tesla K80.•All the necessary steps, algorithms, and mathematical formulas for building the net- work are provided.•The network is small in size but has a faster detection speed with high accuracy.

A transfer learning‐based system for grading breast invasive ductal carcinoma

AbstractBreast carcinoma is a sort of malignancy that begins in the breast. Breast malignancy cells generally structure a tumour that can routinely be seen on an x‐ray or felt like a lump. Despite advances in screening, treatment, and observation that have improved patient endurance rates, breast carcinoma is the most regularly analyzed malignant growth and the subsequent driving reason for malignancy mortality among ladies. Invasive ductal carcinoma is the most boundless breast malignant growth with about 80% of all analyzed cases. It has been found from numerous types of research that artificial intelligence has tremendous capabilities, which is why it is used in various sectors, especially in the healthcare domain. In the initial phase of the medical field, mammography is used for diagnosis, and finding cancer in the case of a dense breast is challenging. The evolution of deep learning and applying the same in the findings are helpful for earlier tracking and medication. The authors have tried to utilize the deep learning concepts for grading breast invasive ductal carcinoma using Transfer Learning in the present work. The authors have used five transfer learning approaches here, namely VGG16, VGG19, InceptionReNetV2, DenseNet121, and DenseNet201 with 50 epochs in the Google Colab platform which has a single 12GB NVIDIA Tesla K80 graphical processing unit (GPU) support that can be used up to 12 h continuously. The dataset used for this work can be openly accessed from http://databiox.com. The experimental results that the authors have received regarding the algorithm's accuracy are as follows: VGG16 with 92.5%, VGG19 with 89.77%, InceptionReNetV2 with 84.46%, DenseNet121 with 92.64%, DenseNet201 with 85.22%. From the experimental results, it is clear that the DenseNet121 gives the maximum accuracy in terms of cancer grading, whereas the InceptionReNetV2 has minimal accuracy.

Intrinsic Performance of Monte Carlo Calibration-Free Algorithm for Laser-Induced Breakdown Spectroscopy.

The performance of the Monte Carlo (MC) algorithm for calibration-free LIBS was studied on the example of a simulated spectrum that mimics a metallurgical slag sample. The underlying model is that of a uniform, isothermal, and stationary plasma in local thermodynamical equilibrium. Based on the model, the algorithm generates from hundreds of thousands to several millions of simultaneous configurations of plasma parameters and the corresponding number of spectra. The parameters are temperature, plasma size, and concentrations of species. They are iterated until a cost function, which indicates a difference between synthetic and simulated slag spectra, reaches its minimum. After finding the minimum, the concentrations of species are read from the model and compared to the certified values. The algorithm is parallelized on a graphical processing unit (GPU) to reduce computational time. The minimization of the cost function takes several minutes on the GPU NVIDIA Tesla K40 card and depends on the number of elements to be iterated. The intrinsic accuracy of the MC calibration-free method is found to be around 1% for the eight elements tested. For a real experimental spectrum, however, the efficiency may turn out to be worse due to the idealistic nature of the model, as well as incorrectly chosen experimental conditions. Factors influencing the performance of the method are discussed.

ReSTiNet: On Improving the Performance of Tiny-YOLO-Based CNN Architecture for Applications in Human Detection

Human detection is a special application of object recognition and is considered one of the greatest challenges in computer vision. It is the starting point of a number of applications, including public safety and security surveillance around the world. Human detection technologies have advanced significantly in recent years due to the rapid development of deep learning techniques. Despite recent advances, we still need to adopt the best network-design practices that enable compact sizes, deep designs, and fast training times while maintaining high accuracies. In this article, we propose ReSTiNet, a novel compressed convolutional neural network that addresses the issues of size, detection speed, and accuracy. Following SqueezeNet, ReSTiNet adopts the fire modules by examining the number of fire modules and their placement within the model to reduce the number of parameters and thus the model size. The residual connections within the fire modules in ReSTiNet are interpolated and finely constructed to improve feature propagation and ensure the largest possible information flow in the model, with the goal of further improving the proposed ReSTiNet in terms of detection speed and accuracy. The proposed algorithm downsizes the previously popular Tiny-YOLO model and improves the following features: (1) faster detection speed; (2) compact model size; (3) solving the overfitting problems; and (4) superior performance than other lightweight models such as MobileNet and SqueezeNet in terms of mAP. The proposed model was trained and tested using MS COCO and Pascal VOC datasets. The resulting ReSTiNet model is 10.7 MB in size (almost five times smaller than Tiny-YOLO), but it achieves an mAP of 63.74% on PASCAL VOC and 27.3% on MS COCO datasets using Tesla k80 GPU.

GPU-Accelerated Computation of EM Scattering of a Time-Evolving Oceanic Surface Model II: EM Scattering of Actual Oceanic Surface

Based on marine environmental factors of different sea areas, a high-performance sea clutter time series modeling algorithm for the real sea surface is developed to study the amplitude mean and Doppler spectrum characteristics of sea clutter. The European Centre for Medium-Range Weather Forecasts (ECMWF) data set (ERA-Interim) and ESA’s soil moisture and ocean salinity (SMOS) data set are utilized to establish databases of different marine environmental factors. Combined with the mixed spectrum model, the geometric fine structure of wind-driven sea surface with swell superposition is established by using the double-superposition method (DSM) and comprehensively considering small-scale capillary ripples, large-scale gravity waves and swell. A triangle facet-based sea clutter series modeling algorithm is developed, in which the quasi-specular scattering based on a triangle and the scattering based on gravity wave modulation capillary spectrum are calculated, respectively, and compared with the measured results. For high-resolution radar, dynamic sea surface modeling and sea clutter calculation are very time consuming. In this paper, the Tesla K80 GPU manufactured by NVIDIA in Santa Clara, Computed Unified Device architecture (CUDA) high-performance parallel technology and some optimization strategies are adopted to improve the efficiency of sea clutter modeling. The results can be used to analyze the distribution characteristics of marine factors, the average amplitude and Doppler characteristics of sea clutter in different sea areas.

Investigating the effect of varying block size on power and energy consumption of GPU kernels

Power consumption is likely to remain a significant concern for exascale performance in the foreseeable future. In addition, graphics processing units (GPUs) have become an accepted architectural feature for exascale computing due to their scalable performance and power efficiency. In a recent study, we found that we can achieve a reasonable amount of power and energy savings based on the selection of algorithms. In this research, we suggest that we can save more power and energy by varying the block size in the kernel configuration. We show that we may attain more savings by selecting the optimum block size while executing the workload. We investigated two kernels on NVIDIA Tesla K40 GPU, a Bitonic Mergesort and Vector Addition kernels, to study the effect of varying block sizes on GPU power and energy consumption. The study should offer insights for upcoming exascale systems in terms of power and energy efficiency.

A New Framework for Accelerating Magnetic Resonance Imaging using Deep Learning along with HPC Parallel Computing Technologies

MRI (magnetic resource imaging) has played a vital role in emerging technologies because of its non-invasion principle. MR equipment is traditional procedure being used for imaging biological structures. In medical domain, MRI is a most important tool being used for staging in clinical diagnosis that has ability to furnish rich physiological and functional information and radiation and non-ionizing nature. However, MRI is highly demanding in several clinical applications. In this paper, we have proposed a novel deep learning based method that accelerates MRI using a huge number of MR images. In proposed method, we used supervised learning approach that performs network training of given datasets. It determines the required network parameters that afford an accurate reconstruction of under-sampled acquisitions. We also designed offline based neural network (NN) that was trained to discover the relationship between MR images and K-space. All the experiments were performed over advanced NVIDIA GPUs (Tesla k80 and GTX Titan) based computers. It was observed that the proposed model outperformed and attained <0.2% error rate. With our best knowledge, our method is the best approach that can be considered as leading model in future.

Translating the InChI: adapting neural machine translation to predict IUPAC names from a chemical identifier

We present a sequence-to-sequence machine learning model for predicting the IUPAC name of a chemical from its standard International Chemical Identifier (InChI). The model uses two stacks of transformers in an encoder-decoder architecture, a setup similar to the neural networks used in state-of-the-art machine translation. Unlike neural machine translation, which usually tokenizes input and output into words or sub-words, our model processes the InChI and predicts the IUPAC name character by character. The model was trained on a dataset of 10 million InChI/IUPAC name pairs freely downloaded from the National Library of Medicine’s online PubChem service. Training took seven days on a Tesla K80 GPU, and the model achieved a test set accuracy of 91%. The model performed particularly well on organics, with the exception of macrocycles, and was comparable to commercial IUPAC name generation software. The predictions were less accurate for inorganic and organometallic compounds. This can be explained by inherent limitations of standard InChI for representing inorganics, as well as low coverage in the training data.

Performance modeling of graphics processing unit application using static and dynamic analysis

SummaryGraphics processing units (GPUs) have become an integral part of high‐performance computing to achieve an exascale performance. Understanding and estimating GPU performance is crucial for developers to design performance‐driven as well as energy‐efficient applications for a given architecture. This work presents a model developed using a static analysis of CUDA code to predict the execution time of NVIDIA GPU kernels without the need for running it. Here a PTX code is statically analyzed to extract instruction features, control flow, and data dependence. We propose a scheduling algorithm that satisfies resource reservation constraints to schedule these instructions in threads across streaming multiprocessors (SMs). We use dynamic analysis to build a set of memory access penalty models and use these models in conjunction with the scheduling information to estimate the execution time of the code. We present the experimental results which support that this approach works across architectures of NVIDIA GPUs. We first tested our model on two Kepler machines, where the mean percentage error (MPE)/mean absolute percentage error (MAPE) was 8.88%/28.3% for Tesla K20 and 5.66%/29.4% for Quadro K4200. We further tested the model on Maxwell and Pascal architectures and recorded the MPEs/MAPEs to be 10.64%/47.8% and %/28.5%, respectively.

Effects of mesh loop modes on performance of unstructured finite volume GPU simulations

In unstructured finite volume method, loop on different mesh components such as cells, faces, nodes, etc is used widely for the traversal of data. Mesh loop results in direct or indirect data access that affects data locality significantly. By loop on mesh, many threads accessing the same data lead to data dependence. Both data locality and data dependence play an important part in the performance of GPU simulations. For optimizing a GPU-accelerated unstructured finite volume Computational Fluid Dynamics (CFD) program, the performance of hot spots under different loops on cells, faces, and nodes is evaluated on Nvidia Tesla V100 and K80. Numerical tests under different mesh scales show that the effects of mesh loop modes are different on data locality and data dependence. Specifically, face loop makes the best data locality, so long as access to face data exists in kernels. Cell loop brings the smallest overheads due to non-coalescing data access, when both cell and node data are used in computing without face data. Cell loop owns the best performance in the condition that only indirect access of cell data exists in kernels. Atomic operations reduced the performance of kernels largely in K80, which is not obvious on V100. With the suitable mesh loop mode in all kernels, the overall performance of GPU simulations can be increased by 15%-20%. Finally, the program on a single GPU V100 can achieve maximum 21.7 and average 14.1 speed up compared with 28 MPI tasks on two Intel CPUs Xeon Gold 6132.

A Comparative Study of Block Incomplete Sparse Approximate Inverses Preconditioning on Tesla K20 and V100 GPUs

Incomplete Sparse Approximate Inverses (ISAI) has shown some advantages over sparse triangular solves on GPUs when it is used for the incomplete LU based preconditioner. In this paper, we extend the single GPU method for Block–ISAI to multiple GPUs algorithm by coupling Block–Jacobi preconditioner, and introduce the detailed implementation in the open source numerical package PETSc. In the experiments, two representative cases are performed and a comparative study of Block–ISAI on up to four GPUs are conducted on two major generations of NVIDIA’s GPUs (Tesla K20 and Tesla V100). Block–Jacobi preconditioning with Block–ISAI (BJPB-ISAI) shows an advantage over the level-scheduling based triangular solves from the cuSPARSE library for the cases, and the overhead of setting up Block–ISAI and the total wall clock times of GMRES is greatly reduced using Tesla V100 GPUs compared to Tesla K20 GPUs.

Neural Network Inversion-Based Model for Predicting an Optimal Hardware Configuration

Deciding the number of processors that can efficiently speed-up solving a computationally intensive problem while perceiving efficient power consumption constitutes a major challenge to researcher in the HPC high performance computing realm. This paper exploits machine learning techniques to propose and implement a recommender system that recommends the optimal HPC architecture given the problem size. An approach for multi-objective function optimization based on neural network (neural network inversion) is employed. The neural network inversion approach is used for forward problem optimization. The objective functions in concern are maximizing the speedup and minimizing the power consumption. The recommendations of the proposed prediction systems achieved more than 89% accuracy for both validation and testing set. The experiments were conducted on 2500 CUDA core on Tesla K20 Kepler GPU Accelerator and Intel(R) Xeon(R) CPU E5-2695 v2.

A novel Ship detection method from SAR image with reduced false alarm

Many research works using deep learning techniques for automatic ship detection from SAR images have good detection accuracy. But the main problem in these methods is false detection, mostly due to speckle presence. Therefore, we propose a new deep learning model with a novel preprocessing stage to address this problem. We are introducing a deep learning architecture to detect and localize ships in the SAR image. First, generate a three-channel image from gray-scale SAR image. Then, this image is used to train the model to predict the ship’s position in the SAR image. We experimented on the public SAR ship detection dataset (SSDD) and Dataset of Ship Detection for Deep Learning under Complex Backgrounds (SDCD) to validate the proposed method’s feasibility. We used python 3.5 for coding with the Keras framework in the NVIDIA Tesla K80 GPU hardware platform. The experimental results indicated that our proposed method’s ship detection accuracy has increased with reduced false detection percentage.

Numerical implementation of biological kinetics multi-species model in the Azov Sea

Paper covers the research of biological kinetics processes based on a multi-species model of plankton and fish interaction of the Azov Sea at low and high size considering the V erhulst and Allee effects, competition for resources, taxis, catching, spatial distribution of biogenic matter and detritus. Discrete analogue of developed model problem of water ecology, included in a software complex, were calculated using schemes of increased order of accuracy considering the partial filling of computational cells. The system of grid equations of large dimension, arising at discretization, has been solved on the basis of a two-layer variational type method – the minimum corrections method having the maximum convergence rate. Effective parallel algorithms were developed for numerical implementation of biological kinetics problem and oriented on multiprocessor computer system and NVIDIA Tesla K80 graphics accelerator with the data storage format modification of. Due to it, the reproduction processes of biogeocenose populations have been analysed in real and accelerated time.