GPU Processing Research Articles

Achieving Large-Scale HBM Integration with Adaptive Pad Stacks on Molded Embedded Bridge Die Interposers

Chiplets and heterogeneous integration are now common vernacular in the semiconductor industry. One of the more pressing issues facing device designers and system architects in this new reality is the insatiable appetite for high-bandwidth memory tightly integrated with GPU, CPU and other chiplet processing devices. This ever-increasing drive for higher levels of integration is resulting in module sizes that now exceed the bounds of the industry’s current de-facto standard silicon interposer solutions. With module dimensions exceeding 50 mm on a side, the industry is transitioning to organic substrates or molded fan-out interposers with embedded bridge die. These embedded bridge technologies provide a practical roadmap for significant module size growth to 100mm per side and beyond. As module sizes continue to grow, the number of embedded bridge die similarly increase. An emerging device currently in design now exceeds 20 bridge die. During assembly of the molded interposer, manufacturing tolerances to achieve the true positions required for all the bridge die become incredibly challenging to achieve in a cost-effective high-volume process. Within this paper, an innovative solution, Adaptive Pad Stacks, will be presented which increases process tolerances by up to 10X versus the conventional approach. For example, the ± 1.5 µm true position required for 35 µm pitch bridge die interfaces can be achieved with a die attach specification allowing up to ± 25 µm. Adaptive Pad Stacks utilize Adaptive Patterning® and mask-less laser direct imaging photolithography to not only deliver high yield on large scale interposers, but also remove the historical challenge with reticle stitching.

Nearest data processing in GPU

Memory wall is known as one of the most critical bottlenecks in processors, rooted in the long memory access delay. With the advent of emerging memory-intensive applications such as image processing, the memory wall problem has become even more critical. Near data processing (NDP) has been introduced as an astonishing solution where instead of moving data from the main memory, instructions are offloaded to the cores integrated with the main memory level. However, in NDP, instructions that are to be offloaded, are statically selected at the compilation time prior to run-time. In addition, NDP ignores the benefit of offloading instructions into the intermediate memory hierarchy levels. We propose Nearest Data Processing (NSDP) which introduces a hierarchical processing approach in GPU. In NSDP, each memory hierarchy level is equipped with processing cores capable of executing instructions. By analyzing the instruction status at run-time, NSDP dynamically decides whether an instruction should be offloaded to the next level of memory hierarchy or be processed at the current level. Depending on the decision, either data is moved upward to the processing core or the instruction is moved downward to the data storage unit. With this approach, the data movement rate has been reduced, on average, by 47 % over the baseline. Consequently, NSDP has been able to improve the system performance, on average, by 37 % and reduce the power consumption, on average, by 18 %.

Refactoring the elastic–viscous–plastic solver from the sea ice model CICE v6.5.1 for improved performance

Abstract. This study focuses on the performance of the elastic–viscous–plastic (EVP) dynamical solver within the sea ice model, CICE v6.5.1. The study has been conducted in two steps. First, the standard EVP solver was extracted from CICE for experiments with refactored versions, which are used for performance testing. Second, one refactored version was integrated and tested in the full CICE model to demonstrate that the new algorithms do not significantly impact the physical results. The study reveals two dominant bottlenecks, namely (1) the number of Message Parsing Interface (MPI) and Open Multi-Processing (OpenMP) synchronization points required for halo exchanges during each time step combined with the irregular domain of active sea ice points and (2) the lack of single-instruction, multiple-data (SIMD) code generation. The standard EVP solver has been refactored based on two generic patterns. The first pattern exposes how general finite differences on masked multi-dimensional arrays can be expressed in order to produce significantly better code generation by changing the memory access pattern from random access to direct access. The second pattern takes an alternative approach to handle static grid properties. The measured single-core performance improvement is more than a factor of 5 compared to the standard implementation. The refactored implementation of strong scales on the Intel® Xeon® Scalable Processors series node until the available bandwidth of the node is used. For the Intel® Xeon® CPU Max series, there is sufficient bandwidth to allow the strong scaling to continue for all the cores on the node, resulting in a single-node improvement factor of 35 over the standard implementation. This study also demonstrates improved performance on GPU processors.

Generating synthetic data in biomedical imaging by designing GANs

Recent advances in deep learning techniques have made medical analysis available with enhanced accuracy and efficiency, where brain tumor classification is automatically identified in an influential role. Hence, one of the synthesized approaches of an innovative idea to use GANs in this paper development is the synthesis of T1-weighted and post-contrast ischemic stroke brain MRIs to increase performance in the classification of the mentioned diseases according to deep learning. This paper, therefore, has the following objective: to evaluate the efficiency of GAN-generated images in learning deep in the transfer learning models and the performance in both tumor and non-tumor brain images. We use the two main architectures of GAN in our process: Vanilla and Deep Convolutional GAN (DCGAN). Details of the three major deep transfer learning models below portray the Convolutional Neural Network (CNN), MobileNetV2, and ResNet152v2. This learned weight would become a pre-trained representation of the models combined with the augmented dataset for feature extraction and classification purposes. I.e., where transfer learning is applied in the models, it is way more accessible for those architectures of the neural network to tap into the knowledge learned by the former from large-scale datasets and adapt it for tasks at hand on classifying brain tumors. Concerning training and validation, Python programming language integrated with the Keras deep learning framework was employed to implement the indicated operations. In terms of training, GPU processing power was available to allow the model to learn faster. In this regard, this was incorporated with the GPU processing by using the NVIDIA GeForce RTX 2060 GPU. Both Vanilla GAN and DCGAN have counterparts when generating images.

GPU implementation of the Frenet Path Planner for embedded autonomous systems: A case study in the F1tenth scenario

Autonomous vehicles are increasingly utilized in safety-critical and time-sensitive settings like urban environments and competitive racing. Planning maneuvers ahead is pivotal in these scenarios, where the onboard compute platform determines the vehicle’s future actions. This paper introduces an optimized implementation of the Frenet Path Planner, a renowned path planning algorithm, accelerated through GPU processing. Unlike existing methods, our approach expedites the entire algorithm, encompassing path generation and collision avoidance. We gauge the execution time of our implementation, showcasing significant enhancements over the CPU baseline (up to 22x of speedup). Furthermore, we assess the influence of different precision types (double, float, half) on trajectory accuracy, probing the balance between completion speed and computational precision. Moreover, we analyzed the impact on the execution time caused by the use of Nvidia Unified Memory and by the interference caused by other processes running on the same system. We also evaluate our implementation using the F1tenth simulator and in a real race scenario. The results position our implementation as a strong candidate for the new state-of-the-art implementation for the Frenet Path Planner algorithm.

RESPAN: an accurate, unbiased and automated pipeline for analysis of dendritic morphology and dendritic spine mapping.

Accurate and unbiased reconstructions of neuronal morphology, including quantification of dendritic spine morphology and distribution, are widely used in neuroscience but remain a major roadblock for large-scale analysis. Traditionally, spine analysis has required labor-intensive manual annotation, which is prone to human error and impractical for large 3D datasets. Previous automated tools for reconstructing neuronal morphology and quantitative dendritic spine analysis face challenges in generating accurate results and, following close inspection, often require extensive manual correction. While recent tools leveraging deep learning approaches have substantially increased accuracy, they lack functionality and useful outputs, necessitating additional tools to perform a complete analysis and limiting their utility. In this paper, we describe Restoration Enhanced SPine And Neuron (RESPAN) analysis, a new comprehensive pipeline developed as an open-source, easily deployable solution that harnesses recent advances in deep learning and GPU processing. Our approach demonstrates high accuracy and robustness, validated extensively across a range of imaging modalities for automated dendrite and spine mapping. It also offers extensive visual and tabulated data outputs, including detailed morphological and spatial metrics, dendritic spine classification, and 3D renderings. Additionally, RESPAN includes tools for validating results, ensuring scientific rigor and reproducibility.

SimLOD: Simultaneous LOD Generation and Rendering for Point Clouds

LOD construction is typically implemented as a preprocessing step that requires users to wait before they are able to view the results in real time. We propose an incremental LOD generation approach for point clouds that allows us to simultaneously load points from disk, update an octree-based level-of-detail representation, and render the intermediate results in real time while additional points are still being loaded from disk. LOD construction and rendering are both implemented in CUDA and share the GPU's processing power, but each incremental update is lightweight enough to leave enough time to maintain real-time frame rates. Our approach is able to stream points from an SSD and update the octree on the GPU at rates of up to 580 million points per second (~9.3GB/s) on an RTX 4090 and a PCIe 5.0 SSD. Depending on the data set, our approach spends an average of about 1 to 2 ms to incrementally insert 1 million points into the octree, allowing us to insert several million points per frame into the LOD structure and render the intermediate results within the same frame.

Energy efficiency and performance analysis of a legacy atomic scale materials modeling simulator (VASP)

This work tackles the performance and energy consumption analysis of a legacy scientific application, the VASP (Vienna Ab-initio Simulation Package), an application commonly used by physicists and chemists for modeling materials at the atomic scale. Many of these scientific applications have been implemented in Fortran, where energy metrics instrumentation is not straightforward. We obtained performance figures (execution time and energy consumption) by instrumenting the source code using EML. This energy measurement library has been modified to introduce Fortran interfaces for these metrics. The analysis was carried out using different matrix algebra libraries, parallelization techniques, and hardware platforms, emphasizing on the MPI, OpenMP, and CUDA parallel implementations of the algorithms used in VASP. We employ various material specifications (atomic structures) and molecular sizes of a silicon-based crystal to create a set of benchmarks for these specifications, leading to some recommendations for final users regarding performance improvements. The proposed benchmarking technique assists the user in selecting the right combination of problem size, compilers, and parallelization options available in VASP. For a given system platform, the user will be able to determine not only the architecture to use (GPU or multicore processors), but also the appropriate library and parallelization according to the atomic structure and molecular size.

Volumetric beamforming in real-time using commodity hardware

Ultrasound imaging is widely used in medicine for its safety and affordability. However, it demands large transducer arrays because the image resolution is proportional to the number of elements, N, typically around 128 for 2D imaging. Three-dimensional imaging requires N2 audio channels (≈350 GB/s of data) if the equivalent 2D matrix array is used, which is practically impossible to process. To solve this issue, row-column arrays (RCAs) aggregate rows and columns of elements, reducing data rate and processing demands by a factor of N. A novel dual-stage beamforming algorithm further lowers the beamforming operations by N/2, with negligible impact on the image quality. For N = 128, the processing is 8192 times faster than with a matrix array, and it is hypothesized the 3D RCA beamforming can be done in real-time using a commodity graphics card. The beamforming rate of an NVIDIA RTX 4090 GPU was measured for in-vivo data from a rat kidney, achieving 1394 full volumes per second, which is over 150 times faster than previous implementations. Combining RCAs with the new beamforming algorithm and GPU processing thus enables volumetric beamforming to be done affordably at the bedside in real-time using a standard scanner and PC.

Modeling resonant ultrasound of synthetic polycrystalline microstructures produced by Dream3D

Resonant ultrasound spectroscopy (RUS) is established for determining elastic constants of homogeneous solids. Often applied to polycrystalline materials like metals, RUS presumes homogeneity based on the microstructure's scale. This presumption is challenged with large grain sizes and higher-order modes, complicating homogeneity testing in real samples due to the need for 3D metallurgical analysis. An alternative is RUS modeling on digitized synthetic microstructures. This presentation investigates the RUS variational model with Rayleigh-Ritz solutions on microstructures created using Dream3D. It starts with a proof for solutions that minimize the Lagrangian in heterogeneous solids. Subsequently, Rayleigh-Ritz volume integrals are numerically evaluated for synthetic polycrystalline microstructures, as spatial heterogeneity precludes analytical solutions. No homogeneity assumption is made. The 3D integrals, requiring advanced GPU processing for a detailed mesh, lead to forward model results. These will be compared against common homogenization schemes, highlighting new insights when using RUS for material characterization of polycrystalline materials.

Real‐time generation of spherical panoramic video using an omnidirectional multi‐camera system

AbstractThis paper presents a novel method for real‐time generation of seamless spherical panoramic videos from an omnidirectional multi‐camera system (OMS). Firstly, a multi‐view video alignment model called spherical projection constrained thin‐plate spline (SP‐TPS) was established and estimated using an approximately symmetrical seam‐line, maintaining the structure inconsistency around the seam‐line. Then, a look‐up table was designed to support real‐time processing of video re‐projection, video dodging and seam‐line updates. In the table, the overlapping areas in OMS multi‐view videos, the seam‐lines between spherical panoramas and OMS multi‐view videos and the pixel coordinate mapping relationship between spherical panoramas and OMS multi‐view videos were pre‐stored as a whole. Finally, a spherical panoramic video was outputted in real‐time through look‐up table computation under an ordinary GPU processor. The experiments were conducted on multi‐view video taken by “1 + 4” and “1 + 7” OMS, respectively. Experimental results demonstrate that compared with four state‐of‐the‐art methods reported in the literature and two bits of commercial software for video stitching, the proposed method excels in eliminating visual artefacts and demonstrates superior adaptability to scenes with varying depths of field. Assuming that OMS is not movable in the scene, this method can generate seamless spherical panoramic videos with a resolution of 8 K in real time, which is of great value to the surveillance field.

Committed Ice Loss in the European Alps Until 2050 Using a Deep‐Learning‐Aided 3D Ice‐Flow Model With Data Assimilation

AbstractModeling the short‐term (<50 years) evolution of glaciers is difficult because of issues related to model initialization and data assimilation. However, this timescale is critical, particularly for water resources, natural hazards, and ecology. Using a unique record of satellite remote‐sensing data, combined with a novel optimisation and surface‐forcing‐calculation method within the framework of the deep‐learning‐based Instructed Glacier Model, we are able to ameliorate initialization issues. We thus model the committed evolution of all glaciers in the European Alps up to 2050 using present‐day climate conditions, assuming no future climate change. We find that the resulting committed ice loss exceeds a third of the present‐day ice volume by 2050, with multi‐kilometer frontal retreats for even the largest glaciers. Our results show the importance of modeling ice dynamics to accurately retrieve the ice‐thickness distribution and to predict future mass changes. Thanks to high‐performance GPU processing, we also demonstrate our method's global potential.

ChromaX: a fast and scalable breeding program simulator.

ChromaX is a Python library that enables the simulation of genetic recombination, genomic estimated breeding value calculations, and selection processes. By utilizing GPU processing, it can perform these simulations up to two orders of magnitude faster than existing tools with standard hardware. This offers breeders and scientists new opportunities to simulate genetic gain and optimize breeding schemes. The documentation is available at https://chromax.readthedocs.io. The code is available at https://github.com/kora-labs/chromax.

Phasor identifier: A cloud-based analysis of phasor-FLIM data on Python notebooks

This paper introduces an innovative approach utilizing Google Colaboratory for the versatile analysis of phasor fluorescence lifetime imaging microscopy (FLIM) data collected from various samples (e.g., cuvette, cells, tissues) and in various input file formats. In fact, phasor-FLIM widespread adoption has been hampered by complex instrumentation and data analysis requirements. We mean to make advanced FLIM analysis more accessible to researchers through a cloud-based solution that 1) harnesses robust computational resources, 2) eliminates hardware limitations, and 3) supports both CPU and GPU processing. We envision a paradigm shift in FLIM data accessibility and potential, aligning with the evolving field of artificial intelligence-driven FLIM analysis. This approach simplifies FLIM data handling and opens doors for diverse applications, from studying cellular metabolism to investigating drug encapsulation, benefiting researchers across multiple domains. The comparative analysis of freely distributed FLIM tools highlights the unique advantages of this approach in terms of adaptability, scalability, and open-source nature.

Energy efficiency in edge TPU vs. embedded GPU for computer-aided medical imaging segmentation and classification

In this work, we evaluate the energy usage of fully embedded medical diagnosis aids based on both segmentation and classification of medical images implemented on Edge TPU and embedded GPU processors. We use glaucoma diagnosis based on color fundus images as an example to show the possibility of performing segmentation and classification in real time on embedded boards and to highlight the different energy requirements of the studied implementations.Several other works develop the use of segmentation and feature extraction techniques to detect glaucoma, among many other pathologies, with deep neural networks. Memory limitations and low processing capabilities of embedded accelerated systems (EAS) limit their use for deep network-based system training. However, including specific acceleration hardware, such as NVIDIA's Maxwell GPU or Google's Edge TPU, enables them to perform inferences using complex pre-trained networks in very reasonable times.In this study, we evaluate the timing and energy performance of two EAS equipped with Machine Learning (ML) accelerators executing an example diagnostic tool developed in a previous work. For optic disc (OD) and cup (OC) segmentation, the obtained prediction times per image are under 29 and 43 ms using Edge TPUs and Maxwell GPUs, respectively. Prediction times for the classification subsystem are lower than 10 and 14 ms for Edge TPUs and Maxwell GPUs, respectively. Regarding energy usage, in approximate terms, for OD segmentation Edge TPUs and Maxwell GPUs use 38 and 190 mJ per image, respectively. For fundus classification, Edge TPUs and Maxwell GPUs use 45 and 70 mJ, respectively.

Fast neural network for TV super resolution scaling‐up system

AbstractIn this paper, we propose a modified architecture aimed at reducing the computational demands of the generative adversarial network for super‐resolution image generation. To achieve this, we embedded depth‐wise and point‐wise convolution into the convolution layer, effectively decreasing operational complexity and improving the overall network structure. For training and validation, we utilized a dataset consisting of 900 image pairs with resolutions of 480 × 270 and 1920 × 1080. Our experimental results demonstrated that the proposed method can reduce computational operators by 63% compared to the original network, while still maintaining the quality of super‐resolution images. To enable real‐time implementation, the architecture with light model subsequently deployed it on a GPU processor, allowing for efficient scaling of TV signals for 16× resolution expansion. Our experiments showed that the peak signal‐to‐noise ratio (PSNR) reached approximately 28 dB, and the processing rate ranged from 6 to 14 frames per second. The network effectively produced output with 16 times greater resolution without introducing any blurring and obvious artifact.

Robust Holographic Reconstruction by Deep Learning with One Frame

A robust method is proposed to reconstruct images with only one hologram in digital holography by introducing a deep learning (DL) network. The U-net neural network is designed according to DL principles and trained by the image data set collected using phase-shifting digital holography (PSDH). The training data set was established by collecting thousands of reconstructed images using PSDH. The proposed method can complete the holography reconstruction with only a single hologram and then benefits the space bandwidth product and relaxes the storage loads of PSDH. Compared with the results of PSDH, the results of deep learning are immune to most disturbances, including reference tilt, phase-shift errors, and speckle noise. Assisted by a GPU processor, the proposed reconstruction method can reduce the consumption time to almost one percent of the time needed by two-step PSDH. This method is expected to be capable of holography imaging with a single hologram, with high capacity, efficiently in the digital holography applications.

The Design of Triple Store and Query Processing on GPU for Large Scale Resource Description Framework Data

The Resource Description Framework (RDF) is commonly used as a standard for data interchange on the web. Due to the current big data era, its size is prone to increase drastically. In order to speed up the large RDF data query, we propose a novel RDF data representation along with the RDF query algorithm utilizing GPU processing. We also present the representation that is suitable for RDF data and GPU processing, the indexing approach, the querying process in the GPU and other techniques that increase the efficiency such as pre-upload filtering, ID assignment by using the term similarity of term vector, and dimensional reduction to transform back to term ID. The experiments show that the developed framework utilizes the storage only 1/6 of the original one and can reduce the querying time. The speedup obtained can be up to 29.57 when compared with the RDF-3X system and 45.23 when compared to using gStore, a graph data store.

A Hybrid Modified Deep Learning Architecture for Intrusion Detection System with Optimal Feature Selection

With the exponentially evolving trends in technology, IoT networks are vulnerable to serious security issues, allowing intruders to break into networks without authorization and manipulate the data. Their actions can be recognized and avoided by using a system that can detect intrusions. This paper presents a hybrid intelligent system and inverted hour-glass-based layered network classifier for feature selection and classification processes, respectively. To accomplish this task, three different datasets have been utilized in the proposed model for identifying old and new attacks. Moreover, a hybrid optimization feature selection technique has been implemented for selecting only those features that can enhance the accuracy of the detection rate. Finally, the classification is performed by using the inverted hour-glass-based layered network model in which data are up-sampled with the increase in the number of layers for effective training. Data up-sampling is performed when small subset of datapoints are observed for any class, which in turn helps in improving the accuracy of the proposed model. The proposed model demonstrated an accuracy of 99.967%, 99.567%, and 99.726% for NSL-KDD, KDD-CUP99, and UNSW NB15 datasets, respectively, which is significantly better than the traditional CNID model. These results demonstrate that our model can detect different attacks with high accuracy and is expected to show good results for new datasets as well. Additionally, to reduce the computational cost of the proposed model, we have implemented it on CPU-based core i3 processors, which are much cheaper than GPU processors.

An application for the earthquake spectral and source parameters and prediction using adaptive neuro fuzzy inference system and machine learning

Parameters related to earthquake origins can be broken down into two broad classes: source location and source dimension. Scientists use distance curves versus average slowness to approximate the epicentre of an earthquake. The shape of curves is the complex function to the epicentral distance, the geological structures of Earth, and the path taken by seismic waves. Brune’s model for source is fitted to the measured seismic wave’s displacement spectrum in order to estimate the source’s size by optimising spectral parameters. The use of ANFIS to determine earthquake magnitude has the potential to significantly alter the playing field. ANFIS can learn like a person using only the data that has already been collected, which improves predictions without requiring elaborate infrastructure. For this investigation’s FIS development, we used a machine with Python 3x running on a core i5 from the 11th generation and an NVIDIA GEFORCE RTX 3050ti GPU processor. Moreover, the research demonstrates that presuming a large number of inputs to the membership function is not necessarily the best option. The quality of inferences generated from data might vary greatly depending on how that data is organised. Subtractive clustering, which does not necessitate any type of normalisation, can be used for prediction of earthquakes magnitude with a high degree of accuracy. This study has the potential to improve our ability to foresee quakes larger than magnitude 5. A solution is not promised to the practitioner, but the research is expected to lead in the right direction. Using Brune’s source model and high cut-off frequency factor, this article suggests using machine learning techniques and a Brune Based Application (BBA) in Python. Application accept input in the Sesame American Standard Code for Information Interchange Format (SAF). An application calculates the spectral level of low frequency displacement (Ω0), the corner frequency at which spectrum decays with a rate of 2(fc), the cut-off frequency at which spectrum again decays (fmax), and the rate of decay above fmax on its own (N). Seismic moment, stress drop, source dimension, etc. have all been estimated using spectral characteristics, and scaling laws. As with the maximum frequency, fmax, its origin can be determined through careful experimentation and study. At some sites, the moment magnitude was 4.7 0.09, and the seismic moment was in the order of (107 0.19) 1023. (dyne.cm). The stress reduction is 76.3 11.5 (bars) and the source-radius is (850.0 38.0) (m). The ANFIS method predicted pretty accurately as the residuals were distributed uniformly near to the centrelines. The ANFIS approach made fairly accurate predictions, as evidenced by the fact that the residuals were distributed consistently close to the centerlines. The R2, RMSE, and MAE indices demonstrate that the ANFIS accuracy level is superior to that of the ANN.