Use Of Graphics Processing Units Research Articles

Active and passive cooling techniques of graphical processing units in automotive applications - a review

Abstract Graphical Processing Units (GPU) have become a vital source for data processing in many advanced or smart technologies. They have proven to be faster and more efficient than the Central Processing Unit (CPU) in computing large data sets. The automotive industry has adopted GPUs for Advanced Driver Assistance Systems (ADAS), but other systems like electric vehicles and autonomous vehicles will also employ GPUs soon. Although GPUs are faster at computing large data sets, there is the challenge of a substantial increase in core temperature during the process. This review paper primarily focuses on Active and Passive cooling methods for GPUs in automotive. The paper commences by addressing the role and need for GPUs in the automotive industry and explores the challenges faced by them and the factors that lead to an increase in temperature. The paper further reviews various cooling methods used by Original Equipment Manufacturers (OEM) in the automotive industry. The outcomes of the review are a detailed understanding of the use of GPUs in automotive, a brief interpretation of heating issues in GPUs and what causes them, a comprehensive understanding of the cooling methods used to reduce the core temperature, and a perception of an anticipated technology to resolve the heating problem.

Run-3 Commissioning of CMS Online HLT reconstruction using GPUs

The software-based High-Level Trigger (HLT) of CMS reduces the data readout rate from 100 kHz (obtained from Level 1 trigger) to around 5 kHz. It makes use of all detector subsystems and runs a streamlined version of CMS reconstruction. Run-2 of the LHC saw the reconstruction algorithms run on a CPU farm. However, the need to have increased computational power as we approach the high luminosity phase of LHC demands the use of Graphical Processing Units (GPUs) to reign in the cost, size and power consumption of the HLT farm. Parallelization of the reconstruction algorithms, on top of the multi-threading functionality introduced in Run-2, allowed parts of the Hadronic Calorimeter (HCAL), Electromagnetic Calorimeter (ECAL) and Pixel Tracker reconstruction to be offloaded to NVIDIA GPUs. In order to ensure the reproducibility of physics results on any machine, the HLT configuration was designed to run seamlessly with and without GPUs, that is, the algorithms were automatically offloaded to a GPU when one was available and otherwise fell back to running on the CPU. This contribution will describe the development of GPU-based algorithms for the HLT and the challenges they presented, along with the comprehensive validation and commissioning activity undertaken by CMS to ensure the successful operations of the new HLT farm.

Analysis of Integration Technologies for High-Speed Analog Neuromorphic Photonics

While the use of graphic processing units fueled the success of artificial intelligence models, their future evolution will likely require overcoming the speed and energy efficiency limitations of current implementations with the use of specialized neuromorphic hardware. In this scenario, neuromorphic photonic processors have recently proved to be a feasible solution. In this paper, we first discuss basic analog photonic processing elements based on Mach-Zehnder modulators and assess their effective bit resolution. Then we evaluate how different photonic integration technologies affect the performance and the scalability of analog optical processors, in order to provide a clearer path toward real-world implementations of such engines. To this aim, we focus our analysis on the silicon on insulator (SOI), lithium niobate on insulator (LNOI), and indium phosphide (InP) platforms. In particular, we have numerically evaluated the performance of the Photonic Electronic Multiply- Accumulate Neuron (PEMAN) and its tensorial version, both based on Mach-Zehnder modulators, with the three technologies in terms of resolution, energy efficiency, and footprint efficiency. LNOI modulators achieve the best resolution at high speed, with 4.3 bits at 56 GMAC/s for the single PEMAN and 3.6 bits at 224 GMAC/S for the tensorial version. The energy consumption in InP and LNOI platforms is the lowest, accounting for just 13.2 pJ/MAC and 4.6 pJ/MAC for the single and tensorial PEMAN, respectively. Nonetheless, SOI devices outperform the others in terms of footprint efficiency, reaching 18.6 GMAC/s/mm <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> in the single-neuron version and 29.6 GMAC/s/mm <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> in the tensorial version.

A 2D Hydraulic Simulation Model Including Dynamic Piping and Overtopping Dambreach

Numerical simulation of unsteady free surface flows using depth averaged equations that consider the presence of initial discontinuities has been often reported for situations dealing with dam break flow. The usual approach is to assume a sudden removal of the gate at the dam location. Additionally, in order to prevent any kind of dam risk in earthen dams, it is very interesting to include the possibility of a progressive dam breach leading to dam overtopping or dam piping so that predictive hydraulic models benefit the global analysis of the water flow. On the other hand, when considering a realistic large domain with complex topography, a fine spatial discretization is mandatory. Hence, the number of grid cells is usually very large and, therefore, it is necessary to use parallelization techniques for the calculation, with the use of Graphic Processing Units (GPU) being one of the most efficient, due to the leveraging of thousands of processors within a single device. The aim of the present work is to describe an efficient GPU-based 2D shallow water flow solver (RiverFlow2D-GPU) supplied with the formulation of internal boundary conditions to represent dynamic dam failure processes. The results obtained indicate that it is able to develop a transient flow analysis taking into account several scenarios. The efficiency of the model is proven in two complex domains, leading to &gt;76× faster simulations compared with the traditional CPU computation.

ReachingBot: An automated and scalable benchtop device for highly parallel Single Pellet Reach-and-Grasp training and assessment in mice

BackgroundThe single pellet reaching and grasp (SPRG) task is a behavioural assay widely used to study motor learning, control and recovery after nervous system injury in animals. The manual training and assessment of the SPRG is labour intensive and time consuming and has led to the development of multiple devices which automate the SPRG task. New methodHere, using robotics, computer vision, and machine learning analysis of videos, we describe a device that can be left unattended, presents pellets to mice, and, using two supervised learning algorithms, classifies the outcome of each trial with an accuracy of greater than 94% without the use of graphical processing units (GPUs). Our devices can also be operated using our cross-platform Graphical User Interface (GUI). ResultsWe show that these devices train and assess mice in parallel. 21 out of 30 mice retrieved > 40% of pellets successfully following the training period. Following ischaemic stroke; some mice showed large persistent deficits whilst others showed only transient deficits. This highlights the heterogeneity in reaching outcomes following stroke. Comparison with existing method(s)Current state-of-the-art desktop methods either still require supervision, manual classification of trial outcome, or expensive locally-installed hardware such as graphical processing units (GPUs). ConclusionsReachingBots successfully automated SPRG training and assessment and revealed the heterogeneity in reaching outcomes following stroke. We conjecture that reach-and-grasp is represented in motor cortex bilaterally but with greater asymmetry in some mice than in others.

Efficient GPU Cloud architectures for outsourcing high-performance processing to the Cloud

AbstractThe world is becoming increasingly dependant in computing intensive applications. The appearance of new paradigms, such as Internet of Things (IoT), and advances in technologies such as Computer Vision (CV) and Artificial Intelligence (AI) are creating a demand for high-performance applications. In this regard, Graphics Processing Units (GPUs) have the ability to provide better performance by allowing a high degree of data parallelism. These devices are also beneficial in specialized fields of manufacturing industry such as CAD/CAM. For all these applications, there is a recent tendency to offload these computations to the Cloud, using a computing offloading Cloud architecture. However, the use of GPUs in the Cloud presents some inefficiencies, where GPU virtualization is still not fully resolved, as our research on what main Cloud providers currently offer in terms of GPU Cloud instances shows. To address these problems, this paper first makes a review of current GPU technologies and programming techniques that increase concurrency, to then propose a Cloud computing outsourcing architecture to make more efficient use of these devices in the Cloud.

How Fast is Fast Enough? Industry Perspectives on the Use of Large-eddy Simulation in Wind Energy

The use of graphics processing units (GPUs) has facilitated unprecedented performance gains for computational fluids dynamics in recent years. In many industries this has enabled the integration of large-eddy simulation (LES) in the engineering practice. Flow modelling in the wind industry though still primarily relies on models with significantly lower fidelity. This paper seeks to investigate the reasons why wind energy applications of LES are still an exception in the industrial practice. On that account, we present a survey among industry experts on the matter. The survey shows that the large runtimes and computational costs of LES are still seen as a main obstacle. However, other reasons such as a lack of expertise and user experience, the need for more validation, and lacking trust in the potential benefits of LES reveal that computational efficiency is not the only concern. Lastly, we present an exemplary simulation of a generic offshore wind farm using a GPU-resident Lattice Boltzmann LES framework. The example shows that the runtime requirements stated by a large part of the respondents can already now be fulfilled with reasonable hardware effort.

Parallelization Strategies for Graph-Code-Based Similarity Search

The volume of multimedia assets in collections is growing exponentially, and the retrieval of information is becoming more complex. The indexing and retrieval of multimedia content is generally implemented by employing feature graphs. Feature graphs contain semantic information on multimedia assets. Machine learning can produce detailed semantic information on multimedia assets, reflected in a high volume of nodes and edges in the feature graphs. While increasing the effectiveness of the information retrieval results, the high level of detail and also the growing collections increase the processing time. Addressing this problem, Multimedia Feature Graphs (MMFGs) and Graph Codes (GCs) have been proven to be fast and effective structures for information retrieval. However, the huge volume of data requires more processing time. As Graph Code algorithms were designed to be parallelizable, different paths of parallelization can be employed to prove or evaluate the scalability options of Graph Code processing. These include horizontal and vertical scaling with the use of Graphic Processing Units (GPUs), Multicore Central Processing Units (CPUs), and distributed computing. In this paper, we show how different parallelization strategies based on Graph Codes can be combined to provide a significant improvement in efficiency. Our modeling work shows excellent scalability with a theoretical speedup of 16,711 on a top-of-the-line Nvidia H100 GPU with 16,896 cores. Our experiments with a mediocre GPU show that a speedup of 225 can be achieved and give credence to the theoretical speedup. Thus, Graph Codes provide fast and effective multimedia indexing and retrieval, even in billion-scale use cases.

GPU acceleration of the ATLAS calorimeter clustering algorithm

Given the upcoming High-Luminosity LHC Upgrade, the performance requirements for the trigger systems associated with the LHC experiments will increase due to the larger volume of data to be processed. One of the possibilities that the ATLAS Collaboration is evaluating for upgrading the software-based portion of its trigger system is the use of Graphical Processing Units as hardware accelerators. The present work focuses on the GPU acceleration of the Topological Clustering algorithm, which is used to reconstruct calorimeter showers by grouping cells according to their signal-to-noise ratio. A more GPU parallelizable version of the Topological Clustering, called Topo-Automaton Clustering, was implemented within AthenaMT, the software framework of the ATLAS trigger, and its results were compared to those of the standard CPU algorithm to ensure physical validity is maintained. Time measurements suggest an average improvement of the event processing time by a factor between 3.5 and 5.5 (depending on the kind of the event), though less than 20% of that time corresponds to the algorithm itself, suggesting that the main bottleneck lies in data transfers and conversions.

Generation of thermal scattering laws with the CINEL code

The thermal scattering laws (TSL) take into account the crystalline structure and atomic motions of isotopes bound in materials. This paper presents the CINEL code, which was developed to generate temperature-dependent TSL for solid, liquid and free gas materials of interest for nuclear reactors. CINEL is able to calculate TSL from the phonon density of states (PDOS) of materials under the Gaussian-Incoherent approximations. The PDOS can be obtained by using theoretical approaches (e.g., ab initio density functional theory and molecular dynamics) or experimental results. In this work, the PDOS presented in the ENDF/BVIII.0 and NJOY-NCrystal libraries were used for numerical validation purposes. The CINEL results are in good agreement with those reported in these databases, even in the specific cases of TSL with the newly mixed elastic format. The coding flexibility offered by Python using the JupyterLab interface allowed to investigate limits of physical models reported in the literature, such as a four-site model for UO2, anharmonic behaviors of oxygen atoms bound in a Fm3m structure, texture in Zry4 samples and jump corrections in a roto-translational diffusion model for liquid water. The use of graphic processing units (GPU) is a necessity to perform calculations in a few minutes. The performances of the CINEL code is illustrated with the results obtained on actinide oxides having a Fm3m structure (UO2, ThO2, NpO2 and PuO2), low enriched fuel (UMo), cladding (Zry4) and moderators (H2O with a specific emphasis on ice).

Irregular alignment of arbitrarily long DNA sequences on GPU

The use of Graphics Processing Units to accelerate computational applications is increasingly being adopted due to its affordability, flexibility and performance. However, achieving top performance comes at the price of restricted data-parallelism models. In the case of sequence alignment, most GPU-based approaches focus on accelerating the Smith-Waterman dynamic programming algorithm due to its regularity. Nevertheless, because of its quadratic complexity, it becomes impractical when comparing long sequences, and therefore heuristic methods are required to reduce the search space. We present GPUGECKO, a CUDA implementation for the sequential, seed-and-extend sequence-comparison algorithm, GECKO. Our proposal includes optimized kernels based on collective operations capable of producing arbitrarily long alignments while dealing with heterogeneous and unpredictable load. Contrary to other state-of-the-art methods, GPUGECKO employs a batching mechanism that prevents memory exhaustion by not requiring to fit all alignments at once into the device memory, therefore enabling to run massive comparisons exhaustively with improved sensitivity while also providing up to 6x average speedup w.r.t. the CUDA acceleration of BLASTN.

Implementation of Beeman's algorithm to calculate execution time on GPU using CUDA

&#x0D; Graphics processing units (GPUs) are microprocessors designed to the operation of display and manipulation of graphics data. . Currently, these graphics processor are found on all graphics hardware and have become very important instruments for parallel computing. GPUs are practical tools for the development of several fields like decoding and encoding, solving differential equations. Their advantages are increase in performance, faster data processing and reduced power consumption. It is simple to program a GPU with CUDA C to run parallel calculations. But it is necessary to have an understanding of the architectural aspects of the GPU and CUDA C. This paper, we will describe and implement Beeman's algorithm on GPU and CPU using CUDA C to solve the differential equation of charged particles in an electromagnetic field. Our goal is to evaluate the performances of the implementation on GPU and CPU processors and to deduce the efficiency of the use of GPUs.&#x0D;

Near-wall modeling of forests for atmosphere boundary layers using lattice Boltzmann method on GPU

In this paper, the simulation and modeling of the turbulent atmospheric boundary layers (ABLs) in the presence of forests are studied using a lattice Boltzmann method with large eddy simulation, which was implemented in the open-source program GASCANS with the use of Graphic Processing Units (GPU). A method of modeling forests in the form of body forces injected near the wall is revisited, while the effects of leaf area density (LAD) on the model accuracy is further addressed. Since a uniform cell size is applied throughout the computational domain, the wall-normal height of the near-wall cells is very large, theoretically requiring a wall function to model the boundary layer. However, the wall function is disregarded here when the forest is modeled. This approximation is validated based on the comparison with previous experimental and numerical data. It concludes that for the ABL conditions specified in this study as well as a large body of literature, the forest forces overwhelm the wall friction so that the modeling of the latter effect is trivial. Constant and varying LAD profiles across the forest zone are defined with the same total leaf area despite the varying one being studied previously. It is found that the two LAD profiles provide consistent predictions. The present forest modeling can therefore be simplified with the use of the constant LAD without degrading the model accuracy remarkably.

Comparison between Weather Signal Simulation in Frequency and Time Domain

Abstract Simulated weather time series are often used in engineering and research practice to assess radar systems behavior and/or to evaluate the performance of novel techniques. There are two main approaches to simulating weather time series. One is based on summing individual returns from a large number of distributed weather particles to create a cumulative return. The other is aimed at creating simulated random signals based on the predetermined values of radar observables and is of interest herein. So far, several methods to simulate weather time series, using the latter approach, have been proposed. All of these methods are based on applying the inverse discrete Fourier transform to the spectral model with added random fluctuations. To meet the desired simulation accuracy, such an approach typically requires generating the number of samples that is larger than the base sample number due to the discrete Fourier transform properties. In that regard, a novel method to determine simulation length is proposed. It is based on a detailed theoretical development that demonstrates the exact source of errors incurred by this approach. Furthermore, a simple method for time series simulation that is based on the autocorrelation matrix exists. This method neither involves manipulations in the spectral domain nor requires generating the number of samples larger than the base sample number. Herein, this method is suggested for weather time series simulation and its accuracy and efficiency are analyzed and compared to the spectral-based approach. Significance Statement All research articles published so far on the topic of weather time series simulation propose the use of inverse discrete Fourier transform (IDFT) when based on the desired Doppler moment values. Herein, a detailed theoretical development that demonstrates the exact source of errors incurred by this approach is presented. Also, a novel method to determine the simulation length that is based on the theoretical error computation is proposed. As an alternative, a computationally efficient general method (not using IDFT) previously developed for the simulation of sequences with desired properties is suggested for weather time series simulation. It is demonstrated that the latter method produces accurate results within overall shorter computational times. Moreover, it is shown that the use of graphics processing unit (GPU), ubiquitous in modern computers, significantly reduces computational times compared to the sole use of central processing unit (CPU) for all simulation-related calculations.

En route for molecular dynamics simulation of a living cell

Creating an in silico all-atom whole-cell model for molecular dynamics (MD) simulation is one of the best ways to quantitatively understand the basic structure and function of cells in terms of the laws of physics and chemistry. The heavy use of graphics processing units (GPUs), the exponential growth of supercomputing power, and the emergence of MD simulation-specific supercomputers lay the groundwork for the MD simulation of molecular machinery. Moreover, the involvement of artificial intelligence (AI) will not only improve the accuracy of the simulation but also significantly accelerate the sampling efficiency. However, several underlying critical puzzles prevent in silico all-atom whole-cell modeling, which is the holy grail of MD simulation. From this perspective, we briefly reviewed the accomplishments of present techniques and hardware as well as provided insight to address the challenge of MD simulation of a living cell. With the rapid advancements in computational hardware, AI, and experimental cell biology, it would be possible to achieve this overarching goal.

GPU acceleration and data fitting: Agent-based models of viral infections can now be parameterized in hours

For many years, infectious disease modelers have used agent-based models to study the spread of viruses, but the models were too computationally intensive to fully replicate even in vitro experiments. Now, with technological advancements and accessible software, agent-based models can be used to their full potential. In this work, we created an agent-based model that expresses viral transmission and diffusion, can manipulate and track individual cells, and can be fit to real experimental data in a timely manner due to acceleration of computation with graphics processing units (GPUs). The use of GPUs allows simulations to run on desktop computers in a few minutes or hours, while still simulating an accurate number of cells to replicate infection experiments. This model can now be used to study in-host infections quickly, so that in the event of an outbreak or epidemic a treatment plan and course of action can be developed in less time.

Andrew Ng, AI Minimalist: The Machine-Learning Pioneer Says Small is the New Big

Andrew Ng has serious street cred in artificial intelligence. He pioneered the use of graphics processing units (GPUs) to train deep learning models in the late 2000s with his students at Stanford University, cofounded Google Brain in 2011, and then served for three years as chief scientist for Baidu, where he helped build the Chinese tech giant&#x0027;s AI group. So when he says he has identified the next big shift in artificial intelligence, people listen. And that&#x0027;s what he told IEEE Spectrum in an exclusive Q&#x0026;A. &#x2022; Ng&#x0027;s current efforts are focused on his company, Landing AI, which built a platform called LandingLens to help manufacturers improve visual inspection with computer vision. He has also become something of an evangelist for what he calls the data-centric AI movement, which he says can yield &#x201C;small data&#x201D; solutions to big issues in AI, including model efficiency, accuracy, and bias.

An Iterative Least Squares Method for Proton CT Image Reconstruction.

Clinically useful proton Computed Tomography images will rely on algorithms to find the three-dimensional proton stopping power distribution that optimally fits the measured proton data. We present a least squares iterative method with many features to put proton imaging into a more quantitative framework. These include the definition of a unique solution that optimally fits the protons, the definition of an iteration vector that takes into account proton measurement uncertainties, the definition of an optimal step size for each iteration individually, the ability to simultaneously optimize the step sizes of many iterations, the ability to divide the proton data into arbitrary numbers of blocks for parallel processing and use of graphical processing units, and the definition of stopping criteria to determine when to stop iterating. We find that it is possible, for any object being imaged, to provide assurance that the image is quantifiably close to an optimal solution, and the optimization of step sizes reduces the total number of iterations required for convergence. We demonstrate the use of these algorithms on real data.

Polygon-based computer-generated holography: a review of fundamentals and recent progress [Invited

In this review paper, we first provide comprehensive tutorials on two classical methods of polygon-based computer-generated holography: the traditional method (also called the fast-Fourier-transform-based method) and the analytical method. Indeed, other modern polygon-based methods build on the idea of the two methods. We will then present some selective methods with recent developments and progress and compare their computational reconstructions in terms of calculation speed and image quality, among other things. Finally, we discuss and propose a fast analytical method called the fast 3D affine transformation method, and based on the method, we present a numerical reconstruction of a computer-generated hologram (CGH) of a 3D surface consisting of 49,272 processed polygons of the face of a real person without the use of graphic processing units; to the best of our knowledge, this represents a state-of-the-art numerical result in polygon-based computed-generated holography. Finally, we also show optical reconstructions of such a CGH and another CGH of the Stanford bunny of 59,996 polygons with 31,724 processed polygons after back-face culling. We hope that this paper will bring out some of the essence of polygon-based computer-generated holography and provide some insights for future research.

Multimodel System for Driver Distraction Detection and Elimination

On average 3,700 people lose their lives on roads every day due to car accidents as a result of drivers&#x2019; distraction. In this research, a proposed hybrid approach is presented. The approach is based on deep learning to detect the driver&#x2019;s actions and eliminate the driver&#x2019;s distraction as a packed solution. The detection is performed by analyzing the driver&#x2019;s actions and his head pose. The elimination is made by using voice commands that are based on trigger words, speech to text, and text classification models to access the car&#x2019;s functions such as the air-condition, radio, etc. The results of the driver&#x2019;s actions classification showed 94.1% accuracy on the AUC benchmark database for driver distraction achieving the state-of-the-art accuracy on this benchmark. The results of the command to text classification is 95.19% while the results of the head pose estimation show a 6.21-degree MAE in face angles detection. With using our car commands dataset, the domain of the speech recognition output is more focused on car commands. The previously mentioned algorithms are beneficial to the safety of the driver. He can use his voice to operate the car accessories. His alert state is monitored and he is warned through an alarm if a distraction is detected. However, this research is not concerned with the detection of retinal abnormalities such as sleeping with eyes open. The results of real-time testing show 0.080 second response time for the driver&#x2019;s behavior classification and command following with the use of graphical processing units.