SEARCH · Engineering Papers

Results for “GPUs”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 91 records · Page 5

All sources NASA NTRS DOE OSTI

DOE OSTI2024

Accelerated Constrained Sparse Tensor Factorization on Massively Parallel Architectures

This study presents the first constrained sparse tensor factorization (cSTF) framework that optimizes and fully offloads computation to massively parallel GPU architectures, and the first performance characterization of cSTF on GPU architectures. In contrast to prior work on tensor factorization, where the matricized tensor times Khatri-Rao product (MTTKRP) is the primary performance bottleneck, our systematic analysis of the cSTF algorithm on GPUs reveals that adding constraints creates an additional bottleneck in the update operation for many real-world sparse tensors. While executing the update operation on the GPU brings significant speedup over its CPU counterpart, it remains a significant bottleneck. To further accelerate the update operation, we propose cuADMM, a new update algorithm that leverages algorithmic and code optimization strategies to minimize both computation and data movement on GPUs. As a result, our framework delivers significantly improved performance compared to prior state-of-the-art. On 10 real-world sparse tensors, our framework achieves geometric mean speedup of 5.1 × (max 41.59 ×) and 7.01 × (max 58.05 ×) on the NIVIDA A100 and H100 GPUs, respectively, over the state-of-the-art SPLATT library running on a 26-core Intel Ice Lake Xeon CPU.

Soh, Yongseok↗

DOE OSTI2025

Accelerating GNNs on GPU Sparse Tensor Cores through N:M Sparsity-Oriented Graph Reordering

Recent advancements in GPU hardware support have introduced the capability to leverage N:M sparse patterns for substantial performance gains. Graphs in Graph Neural Networks (GNNs) are typically sparse, but the sparsity is often irregular, not conforming to such sparse patterns. In this paper, we propose a novel graph reordering algorithm, the first of its kind, to reshape irregular graph data into the N:M structured sparse pattern at the tile level, allowing linear-algebra-based graph operations in GNNs to benefit from the N:M sparse hardware. The optimization is lossless, maintaining the accuracy of GNN. It can remove 98-100\% violations of the N:M sparse patterns at the vector level, and increase the proportion of conforming graphs in SuiteSparse collection from 5-9\% to 88.7-93.5\%. On A100 GPUs, the optimization accelerates Sparse Matrix Matrix (SpMM) by up to 43X (2.3X -- 7.5X on average) and speeds up the key graph operations in GNNs on real graphs by as much as 8.6X (3.5X on average).

artificial intelligence, graph neural networks↗

DOE OSTI2026

Did You Win the GPU Cloud Lottery? Benchmarking from TFLOPS to Tokens/$

Cloud GPUs are commonly assumed to deliver consistent performance for a given GPU model. This assumption does not always hold: cloud providers employ diverse system configurations and virtualization mechanisms, and GPUs themselves exhibit non-negligible manufacturing variability (the silicon lottery). In this work, we present a large-scale measurement study of GPU performance variability across 11 cloud providers, covering over 3,500 physical GPUs and 6,800 benchmark runs. Our hierarchical analysis shows that while execution-level variation stays below 9%, performance varies by up to 38% across devices and providers for the same GPU model. Regression analysis indicates that driver- and OS-related software factors contribute less than 1% of the variance; instead, silicon lottery effects dominate observed performance variation, and cloud providers further amplify them through persistent, systematic second-order effects.

Slynko, Platon [Silicon Data, New York, USA] (ORCI↗

DOE OSTI2024

Visualization at exascale: Making it all work with VTK-m

The VTK-m software library enables scientific visualization on exascale-class supercomputers. Exascale machines are particularly challenging for software development in part because they use GPU accelerators to provide the vast majority of their computational throughput. Algorithmic designs for GPUs and GPU-centric computing often deviate from those that worked well on previous generations of high-performance computers that relied on traditional CPUs. Fortunately, VTK-m provides scientific visualization algorithms for GPUs and other accelerators. VTK-m also provides a framework that simplifies the implementation of new algorithms and adds a porting layer to work across multiple processor types. This paper describes the main challenges encountered when making scientific visualization available at exascale. Here, we document the surprises and obstacles faced when moving from pre-exascale platforms to the final exascale designs and the performance on those systems including scaling studies on Frontier, an exascale machine with over 37,000 AMD GPUs. We also report on the integration of VTK-m with other exascale software technologies. Finally, we show how VTK-m helps scientific discovery for applications such as fusion and particle acceleration that leverage an exascale supercomputer.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2024

Portable, heterogeneous ensemble workflows at scale using libEnsemble

libEnsemble is a Python-based toolkit for running dynamic ensembles, developed as part of the DOE Exascale Computing Project. The toolkit utilizes a unique generator–simulator–allocator paradigm, where generators produce input for simulators, simulators evaluate those inputs, and allocators decide whether and when a simulator or generator should be called. The generator steers the ensemble based on simulation results. Generators may, for example, apply methods for numerical optimization, machine learning, or statistical calibration. libEnsemble communicates between a manager and workers. Flexibility is provided through multiple manager–worker communication substrates each of which has different benefits. These include Python’s multiprocessing, mpi4py, and TCP. Multisite ensembles are supported using Balsam or Globus Compute. We overview the unique characteristics of libEnsemble as well as current and potential interoperability with other packages in the workflow ecosystem. We highlight libEnsemble’s dynamic resource features: libEnsemble can detect system resources, such as available nodes, cores, and GPUs, and assign these in a portable way. These features allow users to specify the number of processors and GPUs required for each simulation; and resources will be automatically assigned on a wide range of systems, including Frontier, Aurora, and Perlmutter. Such ensembles can include multiple simulation types, some using GPUs and others using only CPUs, sharing nodes for maximum efficiency. We also describe the benefits of libEnsemble’s generator–simulator coupling, which easily exposes to the user the ability to cancel, and portably kill, running simulations based on models that are updated with intermediate simulation output. We demonstrate libEnsemble’s capabilities, scalability, and scientific impact via a Gaussian process surrogate training problem for the longitudinal density profile at the exit of a plasma accelerator stage. In conclusion, the study uses gpCAM for the surrogate model and employs either Wake-T or WarpX simulations, highlighting efficient use of resources that can easily extend to exascale.

Dynamic ensembles↗

DOE OSTI2026

A GPU-based compressible combustion solver for applications exhibiting disparate space and time scales

High-speed chemically active flows pose significant computational challenges due to their disparate space and time scales, with stiff chemistry often dominating simulation time. While modern scientific computing programs achieve exascale performance by leveraging graphics processing units (GPUs), existing GPU-based compressible combustion solvers face critical limitations in memory management, load balancing, and handling the highly localized nature of chemical reactions. To this end, we present a high-performance compressible reacting flow solver built on the AMReX framework and optimized for multi-GPU settings. Here, our approach addresses three GPU performance bottlenecks: memory access patterns through column-major storage optimization, computational workload variability via a bulk-sparse integration strategy for chemical kinetics, and multi-GPU load distribution for adaptive mesh refinement applications. The solver adapts existing matrix-based chemical kinetics formulations to multi-grid contexts. Using representative combustion applications, including 2D and 3D detonations and a 3D jet-in-crossflow configuration, we demonstrate 1.4–5× performance improvements over initial implementations on an in-house cluster of NVIDIA H100 GPUs, and near-ideal weak scaling on the Frontier supercomputer (Oak Ridge Leadership Computing Facility) with up to 1024 AMD Instinct MI250X GPUs. Roofline analysis reveals substantial improvements in arithmetic intensity for both convection (∼ 10 ×) and chemistry (∼ 4 ×) routines, confirming efficient utilization of GPU memory bandwidth and computational resources.

42 ENGINEERING↗

DOE OSTI2025

High temporal frequency data from a four turbine, blade-resolved wind farm simulation with ExaWind

The data was generated with ExaWind (https://github.com/Exawind) which couples AMR-Wind (https://github.com/Exawind/amr-wind/), Nalu-Wind (https://github.com/Exawind/nalu-wind), TIOGA (https://github.com/Exawind/tioga), and OpenFAST (https://github.com/OpenFAST/openfast). This is a large-scale simulation of a blade-resolved wind farm using the ExaWind software stack. ExaWind couples together a background flow solver, AMR-Wind, and a near-body solver, Nalu-Wind, through an overset technique from the TIOGA application. Another application, OpenFAST, handles the structural dynamics of the turbine blades and towers, which informs the fluid-structure interaction of the wind turbines with the flow solvers. This particular simulation includes four blade-resolved wind turbines operating in a turbulent atmospheric boundary layer. The AMR-Wind solver uses 500 million cells and is being solved on 256 AMD GPUs of the Oakridge Leadership Computing Facility Frontier supercomputer. Each turbine is assigned its own Nalu-Wind solver with over 13 million elements per turbine and solved using 448 CPU cores, for a total of 1792 CPU cores. For each node, 56 cores contain Nalu-Wind, while 8 cores correspond to AMR-Wind operations on the GPUs. Consequently, ExaWind is entirely utilizing the CPUs and the GPUs of the nodes concurrently. The data used in the visualization is full flow field data output from the simulation. It is lossy-compressed to a specific accuracy using ZFP and written to disk every 16 time-steps to enable real-time flow visualization. The flow fields are sampled at a high temporal frequency to enable real-time, 24fps visualization. The flow fields are sampled every 12 simulation time steps (every 0.04132s).

17 WIND ENERGY↗

DOE OSTI2024

Deployment of inference as a service at the US CMS Tier-2 data centers

Coprocessors, especially GPUs, will be a vital ingredient of data production workflows at the HL-LHC. At CMS, the GPU-as-a-service approach for production workflows is implemented by the SONIC project (Services for Optimized Network Inference on Coprocessors). SONIC provides a mechanism for outsourcing computationally demanding algorithms, such as neural network inference, to remote servers, where requests from multiple clients are intelligently distributed across multiple GPUs by a load-balancing service. This talk highlights the recent progress in deploying SONIC at selected U.S. CMS Tier-2 data centers. Using realistic CMS Run3 data processing workflows, such as those containing transformer-based algorithms, we demonstrate how SONIC is integrated into the production-like environment to enable accelerated inference offloading. We will present developments from both the client and server sides, including production job and data center configurations for NVIDIA and AMD GPUs. We will also present performance scaling benchmarks and discuss the challenges of operating SONIC in CMS production, such as server discovery, GPU saturation, fallback server logic, etc.

Holzman, Burt↗

DOE OSTI2026

TChem-atm (v2.0.0): scalable performance-portable multiphase atmospheric chemistry

We present TChem-atm, a performance-portable approach that enables efficient simulation of chemically detailed and multiphase atmospheric chemistry on modern heterogeneous computing architectures. Unlike previous efforts that rely on architecture-specific code or focus exclusively on gas-phase chemistry, TChem-atm supports fully coupled gas–aerosol systems with execution across CPUs, NVIDIA GPUs, and AMD GPUs through the Kokkos programming model. It integrates the flexible multiphase capabilities of the Community Atmospheric Model Chemistry Package (CAMP) with the high-performance kinetic routines of TChem, and includes automatic Jacobian construction with support for a range of stiff ODE solvers. In a proof-of-concept integration with the particle-resolved model PartMC, TChem-atm reproduces the existing PartMC–CAMP implementation within solver tolerances and delivers substantial GPU speedups, especially for large particle populations. Performance benchmarks reveal substantial speedups on GPU platforms, particularly for large particle populations, with consistent results across hardware backends. TChem-atm enables performance-portable execution across CPUs and GPUs, though optimal efficiency may require modest architecture-specific tuning (e.g., team and vector sizes), with up to a twofold improvement on the NVIDIA H100. It directly supports sectional and particle-resolved host models, while modal aerosol schemes require minor adaptation to provide particle-scale quantities such as representative diameters. By enabling chemically detailed, multiphase simulations with performance portability and host-model flexibility, TChem-atm facilitates the incorporation of advanced chemistry into atmospheric models.

Díaz-Ibarra, Oscar Homero [Sandia National Laborat↗

NASA NTRS2010

GPU Based Software Correlators - Perspectives for VLBI2010

Caused by historical separation and driven by the requirements of the PC gaming industry, Graphics Processing Units (GPUs) have evolved to massive parallel processing systems which entered the area of non-graphic related applications. Although a single processing core on the GPU is much slower and provides less functionality than its counterpart on the CPU, the huge number of these small processing entities outperforms the classical processors when the application can be parallelized. Thus, in recent years various radio astronomical projects have started to make use of this technology either to realize the correlator on this platform or to establish the post-processing pipeline with GPUs. Therefore, the feasibility of GPUs as a choice for a VLBI correlator is being investigated, including pros and cons of this technology. Additionally, a GPU based software correlator will be reviewed with respect to energy consumption/GFlop/sec and cost/GFlop/sec.

Hobiger, Thomas↗

NASA NTRSDate not supplied

Performance Optimization Methods for a Memory-Bound, Unstructured-Grid CFD Application on Massively Parallel GPU Platforms

Computational performance of the FUN3D unstructured-grid computational fluid dynamics (CFD) application on massively parallel GPU environments is memory-bound and highly dependent upon efficient reads from and atomic updates to the irregular cell-, edge-, and node-based data structures. In this talk, we present recent efforts into optimizing select performance-critical kernels on NVIDIA Tesla V100 and A100 GPUs and AMD CDNA MI100 GPUs. A novel use of L2 cache residency controls and asynchronous loads into on-chip shared memory are explored on the A100 GPU for the sparse iterative solver, which is dominated by mixed-precision, sparse matrix vector multiplication. Demonstrations show that these methods improve global memory bandwidth utilization by 13.5% on the A100 GPU. Several techniques are also presented that use registers and/or shared memory to facilitate array transposition and aggregation which combine to reduce the frequency and increase the cache efficiency of floating-point atomic updates to the irregular data structures. These methods are demonstrated to improve the kernel throughput by nearly 500% on select kernels on the AMD MI100 over atomic updates directly to global memory. Overall, both V100 and A100 GPUs outperformed the MI100 GPU on kernels dominated by double-precision atomic updates; however, the techniques demonstrated here reduced the performance gap and improved the MI100 performance.

GPU CPU unstructured CFD memory↗

NASA NTRSDate not supplied

GPU Acceleration of VULCAN-CFD

This work presents a comprehensive overview of recent advancements in the application of GraphicsProcessing Units (GPUs) to accelerate the NASA-developed VULCAN-CFD code for hypersonic flow sim-ulations. The unstructured solver in VULCAN-CFD is undergoing a significant rewrite from modern Fortranto C++, enabling its execution on both GPUs and CPUs through the utilization of Kokkos, a programmingmodel for performance portability developed by Sandia National Labs. The paper outlines some modifica-tions that were made in the original implementation of VULCAN-CFD in order to harness the computationalpower of GPUs. Finally, we demonstrate performance improvements achieved through GPU acceleration.The accelerated code throughput on one GPU is shown to match approximately 200 CPU cores for bothsingle species and multi-species reacting cases.

Matthew O'Connell↗

NASA NTRS2024

EdgeCortix SAKURA-I Machine-Learning, PCIe Accelerator SEE Heavy Ion Test Report

To enable autonomy in space, machine-learning and computer vision applications become invaluable for sensor processing. However, these algorithms are computationally complex and unfeasible for many embedded central processing units (CPUs) and usually require external coprocessors, such as graphics processing units (GPUs) or accelerators specific to the application, including application specific integrated circuits (ASICs). In power-constrained systems, GPUs tend to consume more power than is acceptable (>40W), so lower-power accelerators have shown promise to provide the performance needed under spacecraft constraints. For radiation engineers, developing methodologies that can properly test CPUs, GPUs, and accelerators, and enable comparisons between them remains a necessary complication to solve as the devices become more complex. The methodology in this test aims to be a start in developing a baseline single-event effect (SEE) test for client-device machine learning accelerators. This category of devices do not host their own operating system. This testing campaign is a continuation of a previous 200 MeV proton test performed in January 2024. This report covers two heavy ion tests of the SAKURA-I card: one in April 2024, and one in June 2024. Additional data was needed after the April test due to ion-range issues experienced at higher linear-energy transfers (LETs). These range issues are described in more detail in Section 8. This experiment characterizes SEEs and data error susceptibility of the EdgeCortix SAKURA-I machine-learning accelerator under heavy ions. The device was monitored for single event upsets (SEUs) and single event functional interrupts (SEFIs) at the Lawrence Berkeley National Laboratory’s 88-inch cyclotron. The SAKURA-I board accelerates machine-learning inference applications on a host computer through a PCIex16 connection. For the purposes of devising an end to end automated analysis workflow for this experiment, the YOLO-V5 and SSD300 objection-detection models, and the ResNet-50, EfficientNet, and MobileNetV2 image classification models were used as a representative suite of analytical machine-learning models.

Seth S Roffe↗

DOE OSTI2025

GSplit: Scaling Graph Neural Network Training on Large Graphs via Split-Parallelism

Graph neural networks (GNNs), an emerging class of machine learning models for graphs, have gained popularity for their superior performance in various graph analytical tasks. Mini-batch training is commonly used to train GNNs on large graphs, and data parallelism is the standard approach to scale mini-batch training across multiple GPUs. Data parallel approaches contain redundant work as subgraphs sampled by different GPUs contain significant overlap. To address this issue, we introduce a hybrid parallel mini-batch training paradigm called Split parallelism. Split parallelism avoids redundant work by splitting the sampling, loading, and training of each mini-batch across multiple GPUs. Split parallelism, however, introduces communication overheads that can be more than the savings from removing redundant work. We further present a lightweight partitioning algorithm that probabilistically minimizes these overheads. We implement spllit parllelism in GSplit and show that it outperforms state-of-the-art mini-batch training systems like DGL, Quiver, and P3.

Lim, Seung-Hwan [ORNL] (ORCID:0000000194616866)↗

DOE OSTI2025

A GPU ‐Accelerated 3D Unstructured Mesh Based Particle Tracking Code for Multi‐Species Impurity Transport Simulation in Fusion Tokamaks

ABSTRACT This paper presents the multi‐species global impurity transport capability developed in a GPU‐accelerated fully 3D unstructured mesh‐based code, GITRm, to simultaneously track multiple impurity species and handle interactions of these impurities with mixed‐material surfaces. Different computational approaches to model particle‐surface interaction or surface response have been developed and compared. Sheath electric field is taken into account by employing a fast distance‐to‐boundary calculation, which is carried out in parallel on distributed or partitioned meshes on multiple GPUs without the need for any inter‐process communication during the simulation. Several example cases, including two for the DIII‐D tokamak, that is, one with the SAS‐V divertor and the other with the collector probes, are used to demonstrate the utility of the current multi‐species capability. For the DIII‐D probe case, the capability of GITRm to resolve the spatial distribution of particles in localized regions, such as diagnostic probes, within non‐axisymmetric tokamak geometries is demonstrated. These simulations involve up to 320 million particles and utilize up to 48 GPUs.

Nath, Dhyanjyoti D. [Scientific Computation Resear↗

DOE OSTI2025

Characterizing GPU Energy Usage in Exascale-Ready Portable Science Applications

We characterize the GPU energy usage of two widely adopted exascale-ready applications representing two classes of particle and mesh solvers: (i) QMCPACK, a quantum Monte Carlo package, and (ii) AMReX-Castro, an adaptive mesh astrophysical code. We analyze power, temperature, utilization, and energy traces from double-/single (mixed)-precision benchmarks on NVIDIA’s A100 and H100 and AMD’s MI250X GPUs using queries in NVML and rocm_smi_lib, respectively. We explore application-specific metrics to provide insights on energy vs. performance trade-offs. Our results suggest that mixed-precision energy savings range between 6–25% on QMCPACK and 45% on AMReX-Castro. Also, we found gaps in the AMD tooling used on Frontier GPUs that need to be understood, while query resolutions on NVML have little variability between 1 ms-1 s. Overall, application level knowledge is crucial to define energy-cost/science-benefit opportunities for the codesign of future supercomputer architectures in the post-Moore era.

Godoy, William [ORNL] (ORCID:0000000225905178)↗

DOE OSTI2025

Massively parallel phase-field simulations targeting exascale

The interface thickness in the phase-field (PF) method limits its simulation scales. Consequently, large-scale PF simulations become prohibitively expensive for resolving the extremely fine microstructures that typically form during rapid solidification processing. This challenge is significant in predicting microstructure evolution in metal additive manufacturing and has been identified by the United States Department of Energy’s Exascale Computing Project. Here, to address this, we develop a multi-GPU and MPI-based massively parallel simulation code, utilizing state-of-the-art algorithms, software, and libraries, for large-scale three-dimensional (3D) PF simulations. We report the first GPU-parallel PF simulations on Frontier (currently the second TOP500 exascale cluster) and Summit machines, taking dendritic growth as an example problem. We evaluate the parallel performance of our implementation using scaling studies with more than 24 000 GPUs (among the largest known computations to date) and the acceleration performance using large-scale simulations of dendritic growth in 3D. Finally, massively parallel GPUs in these supercomputers enabled the first coupled multiscale simulations of laser melting and subsequent dendritic solidification on the scale of a full melt-pool, demonstrating the feasibility of performing PF simulations with a point total over 2 billion grid points within an acceptable time.

Exascale↗

DOE OSTI2024

Efficient distributed continual learning for steering experiments in real-time

Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), incremental training suffers from catastrophic forgetting (i.e., new patterns are reinforced at the expense of previously acquired knowledge). Training from scratch each time new training data becomes available would result in extremely long training times and massive data accumulation. Rehearsal-based continual learning has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs to achieve high accuracy, short runtime, and scalability. It leverages a set of buffers (local to each GPU) and uses several asynchronous techniques for updating these local buffers in an embarrassingly parallel fashion, all while handling the communication overheads necessary to augment input minibatches using unbiased, global sampling. We further propose a generalization of rehearsal buffers to support both classification and generative learning tasks, as well as more advanced rehearsal strategies (notably Dark Experience Replay, leveraging knowledge distillation). We illustrate this approach with a real-life HPC streaming application from the domain of ptychographic image reconstruction. Furthermore, we run extensive experiments on up to 128 GPUs of the ThetaGPU supercomputer to compare our approach with baselines representative of training-from-scratch (the upper bound in terms of accuracy) and incremental training (the lower bound). Results show that rehearsal-based continual learning achieves a top-5 validation accuracy close to the upper bound, while simultaneously exhibiting a runtime close to the lower bound.

Asynchronous data management↗