Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “gpu”

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.

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

At least 397 records · Page 22
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Cholla-MHD: An Exascale-capable Magnetohydrodynamic Extension to the Cholla Astrophysical Simulation Code

Abstract We present an extension of the massively parallel, GPU native, astrophysical hydrodynamics code Cholla to magnetohydrodynamics (MHD). Cholla solves the ideal MHD equations in their Eulerian form on a static Cartesian mesh utilizing the Van Leer + constrained transport integrator, the HLLD Riemann solver, and reconstruction methods at second and third order. Cholla’s MHD module can perform ≈260 million cell updates per GPU-second on an NVIDIA A100 while using the HLLD Riemann solver and second order reconstruction. The inherently parallel nature of GPUs combined with increased memory in new hardware allows Cholla’s MHD module to perform simulations with resolutions ∼500 3 cells on a single high-end GPU (e.g., an NVIDIA A100 with 80 GB of memory). We employ GPU direct Message Passing Interface to attain excellent weak scaling on the exascale supercomputer Frontier, while using 74,088 GPUs and simulating a total grid size of over 7.2 trillion cells. A suite of test problems highlights the accuracy of Cholla’s MHD module and demonstrates that zero magnetic divergence in solutions is maintained to round off error. We also present new testing and CI tools using GoogleTest, GitHub Actions, and Jenkins that have made development more robust and accurate and ensure reliability in the future.

Astronomy & Astrophysics↗
DOE OSTI2022

The ICON-A model for direct QBO simulations on GPUs (version icon-cscs:baf28a514)

Abstract. Classical numerical models for the global atmosphere, as used for numerical weather forecasting or climate research, have been developed for conventional central processing unit (CPU) architectures. This hinders the employment of such models on current top-performing supercomputers, which achieve their computing power with hybrid architectures, mostly using graphics processing units (GPUs). Thus also scientific applications of such models are restricted to the lesser computer power of CPUs. Here we present the development of a GPU-enabled version of the ICON atmosphere model (ICON-A), motivated by a research project on the quasi-biennial oscillation (QBO), a global-scale wind oscillation in the equatorial stratosphere that depends on a broad spectrum of atmospheric waves, which originates from tropical deep convection. Resolving the relevant scales, from a few kilometers to the size of the globe, is a formidable computational problem, which can only be realized now on top-performing supercomputers. This motivated porting ICON-A, in the specific configuration needed for the research project, in a first step to the GPU architecture of the Piz Daint computer at the Swiss National Supercomputing Centre and in a second step to the JUWELS Booster computer at the Forschungszentrum Jülich. On Piz Daint, the ported code achieves a single-node GPU vs. CPU speedup factor of 6.4 and allows for global experiments at a horizontal resolution of 5 km on 1024 computing nodes with 1 GPU per node with a turnover of 48 simulated days per day. On JUWELS Booster, the more modern hardware in combination with an upgraded code base allows for simulations at the same resolution on 128 computing nodes with 4 GPUs per node and a turnover of 133 simulated days per day. Additionally, the code still remains functional on CPUs, as is demonstrated by additional experiments on the Levante compute system at the German Climate Computing Center. While the application shows good weak scaling over the tested 16-fold increase in grid size and node count, making also higher resolved global simulations possible, the strong scaling on GPUs is relatively poor, which limits the options to increase turnover with more nodes. Initial experiments demonstrate that the ICON-A model can simulate downward-propagating QBO jets, which are driven by wave–mean flow interaction.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

GSoFa: Scalable Sparse Symbolic LU Factorization on GPUs

Decomposing a matrix $\mathbf {A}$ into a lower matrix $\mathbf {L}$ and an upper matrix $\mathbf {U}$, which is also known as LU decomposition, is an essential operation in numerical linear algebra. For a sparse matrix, LU decomposition often introduces more nonzero entries in the $\mathbf {L}$ and $\mathbf {U}$ factors than in the original matrix. A symbolic factorization step is needed to identify the nonzero structures of $\mathbf {L}$ and $\mathbf {U}$ matrices. Attracted by the enormous potentials of the Graphics Processing Units (GPUs), an array of efforts have surged to deploy various LU factorization steps except for the symbolic factorization, to the best of our knowledge, on GPUs. This article introduces gSoFa, the first GPU-based symbolic factorization design with the following three optimizations to enable scalable LU symbolic factorization for nonsymmetric pattern sparse matrices on GPUs. First, here we introduce a novel fine-grained parallel symbolic factorization algorithm that is well suited for the Single Instruction Multiple Thread (SIMT) architecture of GPUs. Second, we tailor supernode detection into a SIMT friendly process and strive to balance the workload, minimize the communication and saturate the GPU computing resources during supernode detection. Third, we introduce a three-pronged optimization to reduce the excessive space consumption problem faced by multi-source concurrent symbolic factorization. Taken together, gSoFa achieves up to 31× speedup from 1 to 44 Summit nodes (6 to 264 GPUs) and outperforms the state-of-the-art CPU project, on average, by 5×. Notably, gSoFa also achieves up to 47 percent of the peak memory throughput of a V100 GPU in the Summit Supercomputer.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

A massively parallel time-domain coupled electrodynamics–micromagnetics solver

We present a high-performance coupled electrodynamics–micromagnetics solver for full physical modeling of signals in microelectronic circuitry. The overall strategy couples a finite-difference time-domain approach for Maxwell’s equations to a magnetization model described by the Landau–Lifshitz–Gilbert equation. The algorithm is implemented in the Exascale Computing Project software framework, AMReX, which provides effective scalability on manycore and GPU-based supercomputing architectures. Furthermore, the code leverages ongoing developments of the Exascale Application Code, WarpX, which is primarily being developed for plasma wakefield accelerator modeling. Our temporal coupling scheme provides second-order accuracy in space and time by combining the integration steps for the magnetic field and magnetization into an iterative sub-step that includes a trapezoidal temporal discretization for the magnetization. The performance of the algorithm is demonstrated by the excellent scaling results on NERSC multicore and GPU systems, with a significant (59×) speedup on the GPU using a node-by-node comparison. We demonstrate the utility of our code by performing simulations of an electromagnetic waveguide and a magnetically tunable filter.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Performance portable ice-sheet modeling with MALI

High-resolution simulations of polar ice sheets play a crucial role in the ongoing effort to develop more accurate and reliable Earth system models for probabilistic sea-level projections. These simulations often require a massive amount of memory and computation from large supercomputing clusters to provide sufficient accuracy and resolution; therefore, it has become essential to ensure performance on these platforms. Many of today’s supercomputers contain a diverse set of computing architectures and require specific programming interfaces in order to obtain optimal efficiency. In an effort to avoid architecture-specific programming and maintain productivity across platforms, the ice-sheet modeling code known as MPAS-Albany Land Ice (MALI) uses high-level abstractions to integrate Trilinos libraries and the Kokkos programming model for performance portable code across a variety of different architectures. In this article, we analyze the performance portable features of MALI via a performance analysis on current CPU-based and GPU-based supercomputers. The analysis highlights not only the performance portable improvements made in finite element assembly and multigrid preconditioning within MALI with speedups between 1.26 and 1.82x across CPU and GPU architectures but also identifies the need to further improve performance in software coupling and preconditioning on GPUs. We perform a weak scalability study and show that simulations on GPU-based machines perform 1.24–1.92x faster when utilizing the GPUs. The best performance is found in finite element assembly, which achieved a speedup of up to 8.65x and a weak scaling efficiency of 82.6% with GPUs. We additionally describe an automated performance testing framework developed for this code base using a changepoint detection method. The framework is used to make actionable decisions about performance within MALI. We provide several concrete examples of scenarios in which the framework has identified performance regressions, improvements, and algorithm differences over the course of 2 years of development.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Robustness of Deep Learning Classification to Adversarial Input on GPUs: Asynchronous Parallel Accumulation Is a Source of Vulnerability

The ability of machine learning (ML) classification models to resist small, targeted input perturbations—known as adversarial attacks—is a key measure of their safety and reliability. We show that floating-point non associativity (FPNA) coupled with asynchronous parallel programming on GPUs is sufficient to result in misclassification, without any perturbation to the input. Additionally, we show that this misclassification is particularly significant for inputs close to the decision boundary and that standard adversarial robustness results may be overestimated up to 4.6 when not considering machine-level details. We first study a linear classifier, before focusing on standard Graph Neural Network (GNN) architectures and datasets used in robustness assessments. We develop a novel black-box attack using Bayesian optimization to discover external workloads that can change the instruction scheduling which bias the output of reductions on GPUs and reliably lead to misclassification. Motivated by these results, we present a new learnable permutation (LP) gradient-based approach to learning floating-point operation orderings that lead to misclassifications. The LP approach provides a worst-case estimate in a computationally efficient manner, avoiding the need to run identical experiments tens of thousands of times over a potentially large set of possible GPU states or architectures. Finally, using instrumentation-based testing, we investigate parallel reduction ordering across different GPU architectures under external background workloads, when utilizing multi-GPU virtualization, and when applying power capping. Our results demonstrate that parallel reduction ordering varies significantly across architectures under the first two conditions, substantially increasing the search space required to fully test the effects of this parallel scheduler-based vulnerability. These results and the methods developed here can help to include machine-level considerations into adversarial robustness assessments, which can make a difference in safety and mission critical applications.

Shanmugavelu, Sanjif [Maxeler Technologies, a Groq↗
DOE OSTI2022

Optimal size of the block in block GMRES on GPUs: computational model and experiments

The block version of GMRES (BGMRES) is most advantageous over the single right hand side (RHS) counterpart when the cost of communication is high while the cost of floating point operations is not. This is the particular case on modern graphics processing units (GPUs), while it is generally not the case on traditional central processing units (CPUs). Here, in this paper, experiments on both GPUs and CPUs are shown that compare the performance of BGMRES against GMRES as the number of RHS increases, with a particular focus on GPU performance. The experiments indicate that there are many cases in which BGMRES is slower than GMRES on CPUs, but faster on GPUs. Furthermore, when varying the number of RHS on the GPU, there is an optimal number of RHS where BGMRES is clearly most advantageous over GMRES. A computational model for the GPU is developed using hardware specific parameters, providing insight towards how the qualitative behavior of BGMRES changes as the number of RHS increase, and this model also helps explain the phenomena observed in the experiments.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Porting Fragmentation Methods to Graphical Processing Units Using an OpenMP Application Programming Interface: Offloading the Fock Build for Low Angular Momentum Functions

Here, a framework to offload four-index two-electron repulsion integrals to graphical processing units (GPUs) using OpenMP is discussed. The method has been applied to the Fock build for low angular momentum s and p functions in both the restricted Hartree–Fock (RHF) and in the effective fragment molecular orbital (EFMO) framework. Benchmark calculations for the GPU code for the pure RHF method show an increasing speedup relative to the existing OpenMP CPU code in GAMESS from 1.04 to 52× for clusters of 70–569 water molecules. The parallel efficiency on 24 NVIDIA V100 GPU boards also increases when increasing the system size: from 75 to 94% for water clusters that contain 303–1120 molecules. In the EFMO framework, the GPU Fock build shows a high linear scalability up to 4608 V100s with a parallel efficiency of 96% for calculations on a solvated mesoporous silica nanoparticle system with ~67,000 basis functions.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

Performance of CUDA Unified Memory in CMS Heterogeneous Pixel Reconstruction

The management of separate memory spaces of CPUs and GPUs brings an additional burden to the development of software for GPUs. To help with this, CUDA unified memory provides a single address space that can be accessed from both CPU and GPU. The automatic data transfer mechanism is based on page faults generated by the memory accesses. This mechanism has a performance cost, that can be with explicit memory prefetch requests. Various hints on the inteded usage of the memory regions can also be given to further improve the performance. The overall effect of unified memory compared to an explicit memory management can depend heavily on the application. In this paper we evaluate the performance impact of CUDA unified memory using the heterogeneous pixel reconstruction code from the CMS experiment as a realistic use case of a GPU-targeting HEP reconstruction software. We also compare the programming model using CUDA unified memory to the explicit management of separate CPU and GPU memory spaces.

Kortelainen, Matti J.↗
DOE OSTI2024

A new portable random number generator wrapper library

Random number generator is an important component of many scientific projects. Many projects are written using programming models (like OpenMP and SYCL) to target different architectures. However, some programming models do not provide a random number generator. In this work, we introduce our random number generator wrapper. It is a header-only library that supports three distributions of random numbers: uniform, normal, and poisson. On the GPU backend, it wraps the cuRAND and rocRAND library, and supports various random number engines. It also wraps random123, a counterbased random number generator, on both CPU and GPU. With this library, we can generate random numbers with a few lines of code and target both GPU and multi-thread CPU with the same code. We also investigate the performance and scalability of this wrapper on different architectures with different engines and the number of cores.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

Ground and excited state gradients with end-to-end differentiable semiempirical quantum chemistry

Accurate and efficient gradients of molecular energy with respect to nuclear degrees of freedom are essential for geometry optimization and molecular dynamics, including simulations that go beyond the Born–Oppenheimer regime. A common approach involves deriving analytical formulas for new electronic structure methods, which is often conceptually difficult and requires tedious coding. Here, we implement analytical, semi-numerical, and automatic differentiation (AD)-based gradient pathways for semiempirical Hamiltonian models in the PYSEQM software package, leveraging both graphics processing unit (GPU) and central processing unit (CPU) architectures. We further extend these capabilities to excited states calculated using the configuration interaction singles and time-dependent Hartree–Fock ansätze. We benchmark wall time, peak memory usage, and accuracy across three molecular families of varying chemical complexity, including systems of up to a thousand atoms. For ground-state simulations, analytical and AD gradients achieve near-identical GPU runtimes, while semi-numerical gradients are slower on GPU but remain competitive on CPU. For excited states, both analytical and custom AD approaches using implicit differentiation show similar performance and low memory requirements, whereas gradients with full AD are memory-limited. AD gradients match analytical ones in accuracy across all tested systems, aided by a quaternion-based diatomic frame rotation for two-center quantities that ensures smooth energy surfaces. Overall, automatic differentiation emerges as a practical alternative to analytical gradients in semiempirical quantum chemistry, offering high accuracy while allowing seamless integration in AI-driven workflows and popular packages, such as PyTorch and JAX. Our results provide actionable guidance for selecting optimal gradient strategies in large-scale ground- and excited-state molecular dynamics simulations.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2026

Machine learning for seismic low-frequency extrapolation

The cycle-skipping problem that plagues full waveform inversion (FWI) can be at least partially mitigated if low frequencies (which encode the kinematics of wave propagation in seismic data) are recorded. However, seismic sources and receivers are band-limited, so seismic data does not generally include signals down to 0 Hz. To improve our ability to solve the seismic inverse problem, one can synthesize this missing low-frequency (LF) content from the recorded high-frequency (HF) data using machine learning (ML) models. Deep learning models such as convolutional neural networks (CNNs) demonstrate impressive ability to perform low frequency extrapolation. However, such models require powerful hardware (GPU machines) and careful training. We assess the extrapolation capabilities of three different ML models that do not require GPU machines, namely, random forest, Gaussian process regression and gradient boosting, on both synthetic and real data. Experimental results on two synthetic data sets (generated from a low velocity lens embedded in a homogeneous medium, and the Marmousi model) demonstrate that FWI applied to the extrapolated data consistently improves inversion accuracy relative to FWI applied to the original data sets that do not contain low frequencies. Application of low-frequency extrapolation to real data from the Northwest Shelf of Australia demonstrates that tree-based ML models such as gradient boosting can outperform CNNs in terms of both accuracy and computational cost on non-GPU architectures.

58 GEOSCIENCES↗
DOE OSTI2023

Machine Learning Models for Network Traffic Classification in Programmable Logic

Network traffic classification via machine learning on network packet payloads has emerged as an active area of research for network security due to the high accuracy machine learning models have achieved in classifying payloads. For effective deployment as part of network security, these machine learning models must not only classify malicious packet payloads accurately, they must also identify anomalous payloads and perform inference at speeds generally faster than 10,000 packets per second to be effective. This work explores the in- ference speeds and accuracy of several neural network models implemented in programmable logic on various field programmable gate arrays (FPGA) including the Xilinx VC1902 and Xilinx Zynq Ultrascale+. This work also presents the design and performance of both an autoencoder and variational autoencoder programmed on the FPGA for identifying anomalous packet payloads. The performance benefits of the FPGA implementation for this type of packet payload inspection driven by machine learning are compared against graphics processing unit (GPU) inference implementations run on two state-of-the-art datacenter GPU devices, the NVIDIA V100 and A100. The model accuracy difference between the FPGA and GPU implementations was found to be 4% or less while the Xilinx VC1902 outperformed both the NVIDIA V100 and A100 for inference speeds on all the models explored except the variational autoencoder.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

Toward Energy-Efficient HPC: Insights from Power Profiling a Cloud-Resolving Earth System Model

Power is a fundamental constraint as supercomputing advances to exascale. Efficient operation within strict power budgets requires application-aware power management based on a detailed understanding of application-level power behavior. This work analyzes the Energy Exascale Earth System Model (E3SM) atmosphere component, SCREAM, on Perlmutter (NERSC) and Frontier (OLCF). We characterize power variation across inputs, concurrency levels, and power caps, evaluate the energy impact of code optimizations, and attribute energy within the code using a newly developed GPU energy model. Results show that SCREAM’s peak power remains stable during its core execution phase and decreases gradually as concurrency increases. Power capping experiments reveal a performance–energy "sweet spot". On Perlmutter, limiting GPU power to 50% of thermal design power (TDP) achieves up to 15% energy savings with a 7% performance penalty. On Frontier, a 40% TDP cap yields up to 10% energy savings with less than 10% performance loss. Code optimizations reduce SCREAM energy by shortening run time without increasing power. Modeling reveals a critical insight: data movement accounts for approximately 70% of SCREAM’s GPU energy. This fundamentally shifts the optimization focus from FLOPS to data transfer reduction for this class of applications, offering the most impactful strategy for improving energy efficiency. This work establishes a foundation for practical, application-aware power management at exascale.

Zhao, Zhengji [Lawrence Berkeley National Laborato↗
DOE OSTI2022

Evaluating Unified Memory Performance in HIP

Heterogeneous unified memory management between a CPU and a GPU is a major challenge in GPU computing. Recently, unified memory (UM) has been supported by software and hardware components on AMD computing platforms. The support could simplify the complexities of memory management. In this paper, we attempt to have a better understanding of UM by evaluating the performance of UM programs on an AMD MI100 GPU. More specifically, we evaluate data migration using UM against other data transfer techniques for the overall performance of an application, assess the impacts of three commonly used optimization techniques on the kernel execution time of a vector add sample, and compare the performance and productivity of selected benchmarks with and without UM. The performance overhead associated with UM is not trivial, but it can improve programming productivity by reducing lines of code for scientific applications. We aim to present early results and feedback on the UM performance to the vendor.

Jin, Zheming↗
DOE OSTI2022

Understanding Strong Scaling on GPUs Using Empirical Performance Saturation Size

The roofline model provides a concise overview of the maximum performance capabilities of a given computer system through a combination of peak memory bandwidth and compute performance rates. The increasing complexity of scheduling and cache in recent GPUs, however, has introduced complicated performance variability that is not captured by arithmetic intensity alone. This work examines the effect of problem size and GPU launch configurations on roofline performance for V100, A100, MI100, and MI250X graphics processing units. We introduce an extended roofline model that takes problem size into account, and find that strong scaling on GPUs can be characterized by saturation problem sizes as additional key metrics. Saturation problem sizes break up a plot of GPU performance vs. problem size into three distinct performance regimes– size-limited, cache-bound, and DRAM-bound. With our extended roofline model, we are able to provide a robust view of these performance regimes across recent GPU architectures.

Eberius, David↗
DOE OSTI2022

Scalable Deep Learning-Based Microarchitecture Simulation on GPUs

Cycle-accurate microarchitecture simulators are essential tools for designers to architect, estimate, optimize, and manufacture new processors that meet specific design expectations. However, conventional simulators based on discrete-event methods often require an exceedingly long time-to-solution for the simulation of applications and architectures at full complexity and scale. Given the excitement around wielding the machine learning (ML) hammer to tackle various architecture problems, there have been attempts to employ ML to perform architecture simulations, such as Ithemal and SimNet. However, the direct application of existing ML approaches to architecture simulation may be even slower due to overwhelming memory traffic and stringent sequential computation logic. This work proposes the first graphics processing unit (GPU)-based microarchitecture simulator that fully unleashes the potential of GPUs to accelerate state-of-the-art ML-based simulators. First, considering the application traces are loaded from central processing unit (CPU) to GPU for simulation, we introduce various designs to reduce the data movement cost between CPUs and GPUs. Second, we propose a parallel simulation paradigm that partitions the application trace into sub-traces to simulate them in parallel with rigorous error analysis and effective error correction mechanisms. Combined, this scalable GPU-based simulator outperforms by orders of magnitude the traditional CPU-based simulators and the state-of-the-art ML-based simulators, i.e., SimNet and Ithemal.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Evaluating the Performance of Integer Sum Reduction in SYCL on GPUs

SYCL is a promising programming model for heterogeneous computing—allowing a single-source code to target devices from multiple vendors. One significant task performed on these accelerators is a primitive operation for integer sum reduction. This paper presents several SYCL implementations of integer sum reduction—using atomic functions, shared local memory, vectorized memory accesses and parameterized workload sizes—to compare the performance and maturity of SYCL against open-source vendor-specific implementations of the same reduction. For a sufficiently large number of integers, tuning the parameters of our SYCL implementations achieves 1.4X speedup over the open-source implementations on an Intel UHD630 integrated GPU. The SYCL reduction is 3% faster than the templated reduction in Thrust, and 0.3% faster than the device reduction in CUB on an Nvidia P100 GPU. The SYCL reduction is 1.9% faster than the templated reduction in Thrust, and 0.4% faster than the device reduction in CUB on an Nvidia V100 GPU.

Jin, Zheming↗