Search NASA⌕ Search

SEARCH · Search NASA

Results for “parallel processing (computers)”

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 1,153 records · Page 64
All sourcesNASA NTRSDOE OSTI
NASA NTRS2010

Adaptive Numerical Algorithms in Space Weather Modeling

Space weather describes the various processes in the Sun-Earth system that present danger to human health and technology. The goal of space weather forecasting is to provide an opportunity to mitigate these negative effects. Physics-based space weather modeling is characterized by disparate temporal and spatial scales as well as by different physics in different domains. A multi-physics system can be modeled by a software framework comprising of several components. Each component corresponds to a physics domain, and each component is represented by one or more numerical models. The publicly available Space Weather Modeling Framework (SWMF) can execute and couple together several components distributed over a parallel machine in a flexible and efficient manner. The framework also allows resolving disparate spatial and temporal scales with independent spatial and temporal discretizations in the various models. Several of the computationally most expensive domains of the framework are modeled by the Block-Adaptive Tree Solar wind Roe Upwind Scheme (BATS-R-US) code that can solve various forms of the magnetohydrodynamics (MHD) equations, including Hall, semi-relativistic, multi-species and multi-fluid MHD, anisotropic pressure, radiative transport and heat conduction. Modeling disparate scales within BATS-R-US is achieved by a block-adaptive mesh both in Cartesian and generalized coordinates. Most recently we have created a new core for BATS-R-US: the Block-Adaptive Tree Library (BATL) that provides a general toolkit for creating, load balancing and message passing in a 1, 2 or 3 dimensional block-adaptive grid. We describe the algorithms of BATL and demonstrate its efficiency and scaling properties for various problems. BATS-R-US uses several time-integration schemes to address multiple time-scales: explicit time stepping with fixed or local time steps, partially steady-state evolution, point-implicit, semi-implicit, explicit/implicit, and fully implicit numerical schemes. Depending on the application, we find that different time stepping methods are optimal. Several of the time integration schemes exploit the block-based granularity of the grid structure. The framework and the adaptive algorithms enable physics based space weather modeling and even forecasting.

Toth, Gabor↗
DOE OSTI2026

Quality Control of Silicon Sensor Modules for Particle Detectors

The High-Luminosity Large Hadron Collider (HL-LHC) will produce a higher rate of particle collisions than the current Large Hadron Collider (LHC), requiring significant upgrades to the Compact Muon Solenoid (CMS) to handle the increased amount of data. An important upgrade is the Phase-2 Outer Tracker Upgrade, which consists of 13,000 silicon sensor modules made of two parallel silicon sensors and readout electronics. These modules undergo careful quality control checks both during and after module assembly to ensure precise and reliable detector performance. This project focuses on precision testing for quality control of silicon sensor modules at Fermilab. Hands-on work includes visual inspection, current-voltage testing, module testing, and ultraviolet (UV) light exposure of modules showing abnormal current-voltage behavior. The ultraviolet exposure process improves the abnormal sensor readout data by placing the selected sensor side of the module directly under the UV light inside a controlled box. In addition to laboratory testing and ultraviolet experiments, I developed a Python-based data tool that connects to a module database and allows selected testing conditions and module information to be retrieved and displayed efficiently. These different testing procedures, experimental processes, and computational tools support the broader goal of identifying module issues and improving modules that will be used in the CMS Outer Tracker Phase-2 Upgrade.

Siddiqui, Hooriya [DuPage Coll.; Fermilab] (ORCID:↗
DOE OSTI2024

Evolution of the SLATE linear algebra library

SLATE (Software for Linear Algebra Targeting Exascale) is a distributed, dense linear algebra library targeting both CPU-only and GPU-accelerated systems, developed over the course of the Exascale Computing Project (ECP). While it began with several documents setting out its initial design, significant design changes occurred throughout its development. In some cases, these were anticipated: an early version used a simple consistency flag that was later replaced with a full-featured consistency protocol. In other cases, performance limitations and software and hardware changes prompted a redesign. Sequential communication tasks were parallelized; host-to-host MPI calls were replaced with GPU device-to-device MPI calls; more advanced algorithms such as Communication Avoiding LU and the Random Butterfly Transform (RBT) were introduced. Early choices that turned out to be cumbersome, error prone, or inflexible have been replaced with simpler, more intuitive, or more flexible designs. Applications have been a driving force, prompting a lighter weight queue class, nonuniform tile sizes, and more flexible MPI process grids. Of paramount importance has been building a portable library that works across several different GPU architectures – AMD, Intel, and NVIDIA – while keeping a clean and maintainable codebase. Here we explore the evolving design choices and their effects, both in terms of performance and software sustainability.

Gates, Mark↗
NASA NTRS1998

Cloud Scaling Properties and Cloud Parameterization

Cloud liquid and cloud traction variability is studied as a function of horizontal scale in the ECMWF forecast model during several 10-day runs at the highest available model resolution, recently refined from approximately 60 km (T213) down to approximately 20 km (T639). At higher resolutions, model plane-parallel albedo biases are reduced, so that models may be tuned to have larger, more realistic, cloud liquid water amounts, However, the distribution of cloud liquid assumed -within- each gridbox, for radiative and thermodynamic computations, depends on ad hoc assumptions that are not necessarily consistent with observed scaling properties, or with scaling properties produced by the model at larger scales. To study the larger-scale cloud properties, ten locations on the Earth are chosen to coincide with locations having considerable surface data available for validation, and representing a variety of climatic regimes, scaling exponents are determined from a range or scales down to model resolution, and are re-computed every three hours, separately for low, medium and high clouds, as well as column-integrated cloudiness. Cloud variability fluctuates in time, due to diurnal, synoptic and other' processes, but scaling exponents are found to be relatively stable. various approaches are considered for applying computed cloud scaling to subgrid cloud distributions used for radiation, beyond simple random or maximal overlap now in common use. Considerably more work is needed to compare model cloud scaling with observations. This will be aided by increased availability of high-resolution surface, aircraft and satellite data, and by the increasing resolution of global models,

Cahalan, R. F.↗
NASA NTRS2000

NPSS on NASA's IPG: Using CORBA and Globus to Coordinate Multidisciplinary Aeroscience Applications

Within NASA's High Performance Computing and Communication (HPCC) program, the NASA Glenn Research Center is developing an environment for the analysis/design of aircraft engines called the Numerical Propulsion System Simulation (NPSS). The vision for NPSS is to create a "numerical test cell" enabling full engine simulations overnight on cost-effective computing platforms. To this end, NPSS integrates multiple disciplines such as aerodynamics, structures, and heat transfer and supports "numerical zooming" between O-dimensional to 1-, 2-, and 3-dimensional component engine codes. In order to facilitate the timely and cost-effective capture of complex physical processes, NPSS uses object-oriented technologies such as C++ objects to encapsulate individual engine components and CORBA ORBs for object communication and deployment across heterogeneous computing platforms. Recently, the HPCC program has initiated a concept called the Information Power Grid (IPG), a virtual computing environment that integrates computers and other resources at different sites. IPG implements a range of Grid services such as resource discovery, scheduling, security, instrumentation, and data access, many of which are provided by the Globus toolkit. IPG facilities have the potential to benefit NPSS considerably. For example, NPSS should in principle be able to use Grid services to discover dynamically and then co-schedule the resources required for a particular engine simulation, rather than relying on manual placement of ORBs as at present. Grid services can also be used to initiate simulation components on parallel computers (MPPs) and to address inter-site security issues that currently hinder the coupling of components across multiple sites. These considerations led NASA Glenn and Globus project personnel to formulate a collaborative project designed to evaluate whether and how benefits such as those just listed can be achieved in practice. This project involves firstly development of the basic techniques required to achieve co-existence of commodity object technologies and Grid technologies; and secondly the evaluation of these techniques in the context of NPSS-oriented challenge problems. The work on basic techniques seeks to understand how "commodity" technologies (CORBA, DCOM, Excel, etc.) can be used in concert with specialized "Grid" technologies (for security, MPP scheduling, etc.). In principle, this coordinated use should be straightforward because of the Globus and IPG philosophy of providing low-level Grid mechanisms that can be used to implement a wide variety of application-level programming models. (Globus technologies have previously been used to implement Grid-enabled message-passing libraries, collaborative environments, and parameter study tools, among others.) Results obtained to date are encouraging: we have successfully demonstrated a CORBA to Globus resource manager gateway that allows the use of CORBA RPCs to control submission and execution of programs on workstations and MPPs; a gateway from the CORBA Trader service to the Grid information service; and a preliminary integration of CORBA and Grid security mechanisms. The two challenge problems that we consider are the following: 1) Desktop-controlled parameter study. Here, an Excel spreadsheet is used to define and control a CFD parameter study, via a CORBA interface to a high throughput broker that runs individual cases on different IPG resources. 2) Aviation safety. Here, about 100 near real time jobs running NPSS need to be submitted, run and data returned in near real time. Evaluation will address such issues as time to port, execution time, potential scalability of simulation, and reliability of resources. The full paper will present the following information: 1. A detailed analysis of the requirements that NPSS applications place on IPG. 2. A description of the techniques used to meet these requirements via the coordinated use of CORBA and Globus. 3. A description of results obtained to date in the first two challenge problems.

Lopez, Isaac↗
NASA NTRS2003

Evaluation of Alternate Concepts for Synthetic Vision Flight Displays With Weather-Penetrating Sensor Image Inserts During Simulated Landing Approaches

A simulation study was conducted in 1994 at Langley Research Center that used 12 commercial airline pilots repeatedly flying complex Microwave Landing System (MLS)-type approaches to parallel runways under Category IIIc weather conditions. Two sensor insert concepts of 'Synthetic Vision Systems' (SVS) were used in the simulated flights, with a more conventional electro-optical display (similar to a Head-Up Display with raster capability for sensor imagery), flown under less restrictive visibility conditions, used as a control condition. The SVS concepts combined the sensor imagery with a computer-generated image (CGI) of an out-the-window scene based on an onboard airport database. Various scenarios involving runway traffic incursions (taxiing aircraft and parked fuel trucks) and navigational system position errors (both static and dynamic) were used to assess the pilots' ability to manage the approach task with the display concepts. The two SVS sensor insert concepts contrasted the simple overlay of sensor imagery on the CGI scene without additional image processing (the SV display) to the complex integration (the AV display) of the CGI scene with pilot-decision aiding using both object and edge detection techniques for detection of obstacle conflicts and runway alignment errors.

Parrish, Russell V.↗
DOE OSTI2025

Ensemble Simulation Techniques and Fast Randomized Algorithms

The major goals of the project were to develop and analyze new ensemble simulation techniques, including trajectory stratification and preconditioned MCMC techniques, as well as develop fast numerical linear algebra techniques closely related to ensemble simulation ideas. The trajectory stratification techniques involve simulating in parallel short trajectory fragments of a Markov process confined to a specific region of space‐time and then patching together the statistics gathered to assemble estimates of very general dynamical properties. We have also developed this approach for rare event simulation and extended the techniques to applications requiring a more general framework (such as electronic structure calculations). The preconditioned MCMC techniques involve simulating multiple Markov chains in parallel and then using information from the ensemble to speed the mixing of each individual chain. The fast randomized linear algebra methods are motivated by the diffusion Monte Carlo technique, but are applicable to finding the dominant eigenvalue of (almost) general matrices. For most non‐negative matrices, the schemes result in an error (compared to the power method) that is constant in the dimension of the problem. For more general matrices, we see a very clear sublinear cost trend in computational tests.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2025

Self-Consistent Relativistic Electron Scattering using the Sherlock Scattering Model for X-ray Diagnostics

We present on a new, self-consistent, arbitrary-temperature Romberg integration scheme for modeling electron scattering in materials in a LANL Lagrangian Shock Hydro (LSH) code. Electron beam-target interactions are fundamental to a wide range of scientific and technological applications. When high-energy electron beams hit their target, they may scatter, deposit energy, or ionize the source. These processes govern the behavior and outcomes in nanotechnology manufacturing, electron microscopy, and modern X-ray diagnostics. Simulating these interactions is essential for interpreting experimental results, predicting material responses, and designing efficient tools and experiments. At Los Alamos, this is done using a LSH code, which is a multi-dimension, multi-material, massively parallel, multi-physics code used to simulate applications from asteroid impacts to electron beam interactions. By effectively and efficiently modeling the way that electrons scatter from the beam we can bolster these simulations and more accurately predict experimental outcomes. The model currently implemented in the LSH of interest is based on work by Papp and does not self-consistently preserve momentum in the slightly relativistic regime; here we adopt a model proposed by Braams and Karney and implement a Romberg integration scheme to compute the diffusion tensor. In this paper we will provide background on the Braams-Karney diffusion tensor as well as the Romberg integration scheme we employed to numerically solve for it. We will show that our integration scheme is accurate in solving for the set of scalar potentials used to re-express the diffusion tensor in differential form, and in solving for the diffusion coefficients in the larger LSH code. By using this diffusion tensor rather than the existing Papp one, and numerically integrating it with a Romberg method, we produce much more accurate, self-consistent results.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2025

DIMPLES: Distributed Influence Maximization for Pandemic pLanning on Exascale Systems

We study exascale parallel algorithms for the selection of intervention or monitoring strategies in massive realistic socio-technical networks through scalable Influence Maximization (InfMax) algorithms. We employ novel techniques to enable efficient scaling on up to 8k nodes of OLCF Frontier, with 65k AMD GPUs and 458k AMD CPU cores. Current state-of-the-art InfMax tools are limited to networks with only a few million actors (vertices) and a few hundred million interactions (edges). By overcoming these limitations, we show that our approach is capable of processing a realistic social contact network of the United States with 285 million nodes and about 8 billion edges. This two orders-of-magnitude improvement over the previous state-of-the-art is obtained by leveraging algorithmic advancements for the InfMax problem and designing several problem-specific approaches to overlap communication with computation, improve GPU efficiency, and lower the application’s memory requirements. We evaluate strong scaling for computing 10k most influential seeds using up to 8k nodes of an exascale system, and weak scaling from 128 to 8k system nodes for seed sets ranging from 625 to 40k seeds. We achieve the fastest-known runtime of 25 minutes while performing 48 million diffusion simulations totaling 2.31 petabytes to identify 40k influential seeds using 8k nodes, and take 5.75 minutes to identify 10k seeds while using 4k nodes.

Minutoli, Marco [Pacific Northwest National Labora↗
DOE OSTI2024

Boundary of the Distribution of Solar Wind Proton Beta versus Temperature Anisotropy

Abstract The frequency distribution of solar wind protons, measured in the vicinity of Earth’s orbit, is customarily plotted in ( β ∥ , T ⊥ / T ∥ ) phase space. Here, T ⊥ / T ∥ is the ratio of perpendicular and parallel temperatures, and β ∥ = 8 π nT ∥ / B 2 is the ratio of parallel thermal energy to background magnetic field energy, the so-called “parallel beta,” with ⊥ and ∥ denoting directions with respect to the ambient magnetic field. Such a frequency distribution, plotted as a two-dimensional histogram, forms a peculiar rhombic shape defined with an outer boundary in the said phase space. Past studies reveal that the threshold conditions for temperature anisotropy–driven plasma instability partially account for the boundary on the high- β ∥ side. The low- β ∥ side remains largely unexplained despite some efforts. Work by Vafin et al. recently showed that certain contours of collisional relaxation frequency, ν pp , when parameterized by T ⊥ / T ∥ and β ∥ , could match the overall shape of the left-hand boundary, thus suggesting that the collisional relaxation process might be closely related to the formation of the left-hand boundary. The present paper extends the analysis by Vafin et al. and carries out the dynamical computation of the collisional relaxation process for an ensemble of initial proton states with varying degrees of anisotropic temperatures. The final states of the relaxed protons are shown to closely match the observed boundary to the left of the ( β ∥ , T ⊥ / T ∥ ) phase space. When coupled with a similar set of calculations for the ensemble in the collective instability regime, it is found that the combined collisional/collective effects provide the baseline explanation for the observation.

Astronomy & Astrophysics↗
NASA NTRS1991

Electronic neural network for dynamic resource allocation

A VLSI implementable neural network architecture for dynamic assignment is presented. The resource allocation problems involve assigning members of one set (e.g. resources) to those of another (e.g. consumers) such that the global 'cost' of the associations is minimized. The network consists of a matrix of sigmoidal processing elements (neurons), where the rows of the matrix represent resources and columns represent consumers. Unlike previous neural implementations, however, association costs are applied directly to the neurons, reducing connectivity of the network to VLSI-compatible 0 (number of neurons). Each row (and column) has an additional neuron associated with it to independently oversee activations of all the neurons in each row (and each column), providing a programmable 'k-winner-take-all' function. This function simultaneously enforces blocking (excitatory/inhibitory) constraints during convergence to control the number of active elements in each row and column within desired boundary conditions. Simulations show that the network, when implemented in fully parallel VLSI hardware, offers optimal (or near-optimal) solutions within only a fraction of a millisecond, for problems up to 128 resources and 128 consumers, orders of magnitude faster than conventional computing or heuristic search methods.

Thakoor, A. P.↗
NASA NTRS2010

Application of the FUN3D Unstructured-Grid Navier-Stokes Solver to the 4th AIAA Drag Prediction Workshop Cases

FUN3D Navier-Stokes solutions were computed for the 4th AIAA Drag Prediction Workshop grid convergence study, downwash study, and Reynolds number study on a set of node-based mixed-element grids. All of the baseline tetrahedral grids were generated with the VGRID (developmental) advancing-layer and advancing-front grid generation software package following the gridding guidelines developed for the workshop. With maximum grid sizes exceeding 100 million nodes, the grid convergence study was particularly challenging for the node-based unstructured grid generators and flow solvers. At the time of the workshop, the super-fine grid with 105 million nodes and 600 million elements was the largest grid known to have been generated using VGRID. FUN3D Version 11.0 has a completely new pre- and post-processing paradigm that has been incorporated directly into the solver and functions entirely in a parallel, distributed memory environment. This feature allowed for practical pre-processing and solution times on the largest unstructured-grid size requested for the workshop. For the constant-lift grid convergence case, the convergence of total drag is approximately second-order on the finest three grids. The variation in total drag between the finest two grids is only 2 counts. At the finest grid levels, only small variations in wing and tail pressure distributions are seen with grid refinement. Similarly, a small wing side-of-body separation also shows little variation at the finest grid levels. Overall, the FUN3D results compare well with the structured-grid code CFL3D. The FUN3D downwash study and Reynolds number study results compare well with the range of results shown in the workshop presentations.

Lee-Rausch, Elizabeth M.↗
DOE OSTI2026

Tula: Optimizing Time, Cost, and Generalization in Distributed Large-Batch Training

Distributed training increases the number of batches processed per iteration either by scaling-out (adding more nodes) or scaling-up (increasing the batch-size). However, the largest configuration does not necessarily yield the best performance. Horizontal scaling introduces additional communication overhead, while vertical scaling is constrained by computation cost and device memory limits. Thus, simply increasing the batch-size leads to diminishing returns: training time and cost decrease initially but eventually plateaus, creating a knee-point in the time/cost vs. batch-size pareto curve. The optimal batch-size therefore depends on the underlying model, data and available compute resources. Large batches also suffer from worse model quality due to the well-known “generalization gap”. In this paper, we present Tula, an online service that automatically optimizes time, cost, and convergence quality for large-batch training of convolutional models. It combines parallel-systems modeling with statistical performance prediction to identify the optimal batchsize. Tula predicts training time and cost within 7.5−14% error across multiple models, and achieves up to 20× overall speedup and improves test accuracy by ≈9% on average over standard large-batch training on various vision tasks, thus successfully mitigating the generalization gap and accelerating training at the same time.

Tyagi, Sahil [ORNL] (ORCID:0009000783144745)↗
NASA NTRS1996

On-board Neural Processor Design for an Intelligent Multi-sensor Microspacecraft

A compact VLSI neural processor based on the Optimization Cellular Neural Network (OCNN)has been under development to provide a wide range of support for an intelligent remote sensing microspacecraft which requires both high bandwidth communication and high-performance computing for on-board data analysis, thematic data reduction, synergy of multiple types of sensors, and other smart-sensor functions. The OCNN architecture is a programmable multi-dimensional array of neurons which are locally connected with their local neurons. The OCNN operation theory, architecture, design and implementation, prototype chip, and system applications have been investigated in detail and presented in this paper.

array↗
NASA NTRS1998

High Rate Digital Demodulator ASIC

The architecture of High Rate (600 Mega-bits per second) Digital Demodulator (HRDD) ASIC capable of demodulating BPSK and QPSK modulated data is presented in this paper. The advantages of all-digital processing include increased flexibility and reliability with reduced reproduction costs. Conventional serial digital processing would require high processing rates necessitating a hardware implementation in other than CMOS technology such as Gallium Arsenide (GaAs) which has high cost and power requirements. It is more desirable to use CMOS technology with its lower power requirements and higher gate density. However, digital demodulation of high data rates in CMOS requires parallel algorithms to process the sampled data at a rate lower than the data rate. The parallel processing algorithms described here were developed jointly by NASA's Goddard Space Flight Center (GSFC) and the Jet Propulsion Laboratory (JPL). The resulting all-digital receiver has the capability to demodulate BPSK, QPSK, OQPSK, and DQPSK at data rates in excess of 300 Mega-bits per second (Mbps) per channel. This paper will provide an overview of the parallel architecture and features of the HRDR ASIC. In addition, this paper will provide an over-view of the implementation of the hardware architectures used to create flexibility over conventional high rate analog or hybrid receivers. This flexibility includes a wide range of data rates, modulation schemes, and operating environments. In conclusion it will be shown how this high rate digital demodulator can be used with an off-the-shelf A/D and a flexible analog front end, both of which are numerically computer controlled, to produce a very flexible, low cost high rate digital receiver.

Ghuman, Parminder↗
NASA NTRS2007

Extravehicular Mobility Unit Penetration Probability from Micrometeoroids and Orbital Debris: Revised Analytical Model and Potential Space Suit Improvements

The NASA Extravehicular Mobility Unit (EMU) micrometeoroid and orbital debris protection ability has recently been assessed against an updated, higher threat space environment model. The new environment was analyzed in conjunction with a revised EMU solid model using a NASA computer code. Results showed that the EMU exceeds the required mathematical Probability of having No Penetrations (PNP) of any suit pressure bladder over the remaining life of the program (2,700 projected hours of 2 person spacewalks). The success probability was calculated to be 0.94, versus a requirement of >0.91, for the current spacesuit s outer protective garment. In parallel to the probability assessment, potential improvements to the current spacesuit s outer protective garment were built and impact tested. A NASA light gas gun was used to launch projectiles at test items, at speeds of approximately 7 km per second. Test results showed that substantial garment improvements could be made, with mild material enhancements and moderate assembly development. The spacesuit s PNP would improve marginally with the tested enhancements, if they were available for immediate incorporation. This paper discusses the results of the model assessment process and test program. These findings add confidence to the continued use of the existing NASA EMU during International Space Station (ISS) assembly and Shuttle Operations. They provide a viable avenue for improved hypervelocity impact protection for the EMU, or for future space suits.

Chase, Thomas D.↗
DOE OSTI2026

HARMONY: Large-Scale Architecture Search for Efficient Hybrid Language Models

As large language models scale to trillions of parameters, their computational and memory requirements present critical challenges for efficient training and deployment. While Mixture of Experts (MoE) architectures enable efficient scaling through sparse parameter activation, and state-space models like Mamba offer linear-time complexity, principled methods for combining these paradigms remain undeveloped. We introduce HARMONY (Hybrid Architecture Research for Mamba, Optimized with Neural efficiencY), a multi-objective evolutionary neural architecture search framework for discovering efficient hybrid language models that integrate Transformer attention mechanisms, Mixture-of-Experts routing, and Mamba state-space components. Through large-scale distributed search using 16,384 MI250X GPUs on the Frontier supercomputer, HARMONY explores a comprehensive design space encompassing six attention variants (MHA, MQA, GQA, MLA, SWA, and Mamba-2), variable MoE configurations with both routed and shared experts, and extensive Mamba hyperparameters. Our framework discovers heterogeneous architectures that balance training performance with computational efficiency through multi-objective optimization incorporating latency penalties and fitness-based selection. Analysis of discovered architectures reveals that optimal hybrid designs favor heterogeneous component mixing rather than homogeneous patterns, with Mamba-2 and Multi-Head Latent Attention (MLA) emerging as preferred mechanisms. Discovered architectures demonstrate superior training efficiency: our best configuration achieves a final perplexity of 1.0874 with 2.38B parameters while processing 4,320 tokens/second, outperforming significantly larger manually designed models. Full-scale evaluation shows HARMONY's top architectures achieve better loss trajectories than equivalently-sized models using state-of-the-art configurations including Mixtral, Jamba, and Samba. Additionally, we demonstrate 91% weak scaling efficiency when training discovered 36B-parameter models across 1,024 GPUs. HARMONY is released as an open framework with comprehensive tools for building and training hybrid models using expert-data-pipeline parallelism, democratizing access to automated architecture design for next-generation language models.

Herron, Emily [ORNL] (ORCID:0000000273008172)↗
NASA NTRS1993

On the spectral stability of time integration algorithms for a class of constrained dynamics problems

Incomplete field formulations have recently been the subject of intense research because of their potential in coupled analysis of independently modeled substructures, adaptive refinement, domain decomposition, and parallel processing. This paper discusses the design and analysis of time-integration algorithms for these formulations and emphasizes the treatment of their inter-subdomain constraint equations. These constraints are shown to introduce a destabilizing effect in the dynamic system that can be analyzed by investigating the behavior of the time-integration algorithm at infinite and zero frequencies. Three different approaches for constructing penalty-free unconditionally stable second-order accurate solution procedures for this class of hybrid formulations are presented, discussed and illustrated with numerical examples. The theoretical results presented in this paper also apply to a large family of nonlinear multibody dynamics formulations. Some of the algorithms outlined herein are important alternatives to the popular technique consisting of transforming differential/algebraic equations into ordinary differential equations via the introduction of a stabilization term that depends on arbitrary constants and that influences the computed so1ution.

Farhat, Charbel↗