Search NASA⌕ Search

SEARCH · Search NASA

Results for “hardware accelerators”

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 181 records · Page 10

SpF: Enabling Petascale Performance for Pseudospectral Dynamo Models

Pseudospectral (PS) methods possess a number of characteristics (e.g., efficiency, accuracy, natural boundary conditions) that are extremely desirable for dynamo models. Unfortunately, dynamo models based upon PS methods face a number of daunting challenges, which include exposing additional parallelism, leveraging hardware accelerators, exploiting hybrid parallelism, and improving the scalability of global memory transposes. Although these issues are a concern for most models, solutions for PS methods tend to require far more pervasive changes to underlying data and control structures. Further, improvements in performance in one model are difficult to transfer to other models, resulting in significant duplication of effort across the research community.We have developed an extensible software framework for pseudospectral methods called SpF that is intended to enable extreme scalability and optimal performance. High-level abstractions provided by SpF unburden applications of the responsibility of managing domain decomposition and load balance while reducing the changes in code required to adapt to new computing architectures. The key design concept in SpF is that each phase of the numerical calculation is partitioned into disjoint numerical kernels that can be performed entirely in-processor. The granularity of domain-decomposition provided by SpF is only constrained by the data-locality requirements of these kernels. SpF builds on top of optimized vendor libraries for common numerical operations such as transforms, matrix solvers, etc., but can also be configured to use open source alternatives for portability. SpF includes several alternative schemes for global data redistribution and is expected to serve as an ideal testbed for further research into optimal approaches for different network architectures.In this presentation, we will describe the basic architecture of SpF as well as preliminary performance data and experience with adapting legacy dynamo codes. We will conclude with a discussion of planned extensions to SpF that will provide pseudospectral applications with additional flexibility with regard to time integration, linear solvers, and discretization in the radial direction.

Pseudospectral (PS)↗

Using SpF to Achieve Petascale for Legacy Pseudospectral Applications

Pseudospectral (PS) methods possess a number of characteristics (e.g., efficiency, accuracy, natural boundary conditions) that are extremely desirable for dynamo models. Unfortunately, dynamo models based upon PS methods face a number of daunting challenges, which include exposing additional parallelism, leveraging hardware accelerators, exploiting hybrid parallelism, and improving the scalability of global memory transposes. Although these issues are a concern for most models, solutions for PS methods tend to require far more pervasive changes to underlying data and control structures. Further, improvements in performance in one model are difficult to transfer to other models, resulting in significant duplication of effort across the research community. We have developed an extensible software framework for pseudospectral methods called SpF that is intended to enable extreme scalability and optimal performance. Highlevel abstractions provided by SpF unburden applications of the responsibility of managing domain decomposition and load balance while reducing the changes in code required to adapt to new computing architectures. The key design concept in SpF is that each phase of the numerical calculation is partitioned into disjoint numerical kernels that can be performed entirely inprocessor. The granularity of domain decomposition provided by SpF is only constrained by the datalocality requirements of these kernels. SpF builds on top of optimized vendor libraries for common numerical operations such as transforms, matrix solvers, etc., but can also be configured to use open source alternatives for portability. SpF includes several alternative schemes for global data redistribution and is expected to serve as an ideal testbed for further research into optimal approaches for different network architectures. In this presentation, we will describe our experience in porting legacy pseudospectral models, MoSST and DYNAMO, to use SpF as well as present preliminary performance results provided by the improved scalability.

DYNAMO↗

Rising Above the Cloud - Toward High-Rate Delay-Tolerant Networking in Low-Earth Orbit

The High Data Rate Architecture (HiDRA) project is implementing a High-rate Delay Tolerant Networking (HDTN) capability that can support Low Earth Orbit (LEO) applications and environments. The present state of the effort, future work, and other elements of the work to date are described in this paper. This implementation is intended to support applications that run at 1+ Gbps, per the requirements of modern optical and high-frequency RF links. Uniquely, this implementation is also tuned to support relay and data trunking applications, which might require support for large numbers of small bundles per second. The design for this platform is based entirely on commercial-off-the-shelf (COTS) components, and possesses buffering capabilities in the 5 TB range. This document takes results from previous individual tests and integrates them to demonstrate results in the presence of a coherent use-case: consider a network aboard the ISS which intends to utilize an upcoming optical communications capability. For this use-case, orbital analysis software is used to analyze orbital dynamics, from which a list of access times are generated that might take in to account weather, schedule competition, etc. A variant of Contact Graph Routing (CGR) is applied to these windows to determine an optimal schedule. This schedule is then loaded into the HDTN prototype and, in conjunction with various measurement tools, a complete end-to-end analysis of HDTN's performance is conducted. Various bottlenecks (including storage) are identified: these bottlenecks are expected to help us focus our future work on the elements of the system that are most likely to present issues moving forward. Finally, we discuss possible paths for evolution beyond the present rates supported by the system, including (but not limited to) hardware acceleration.

Hylton, Alan↗

QuAIL Tools for Benchmarking, Analysis and Quantum Algorithm Development

HybridQ and PySA are open-source tools developed by NASA to support benchmarking, analysis and quantum algorithm development in areas such as simulation, optimization and machine learning. These tools leverage classical hardware acceleration via high-performance computing CPU and GPU architectures and support high-performance computing. HybridQ is a highly extensible platform designed to provide a common framework to integrate multiple state-of-the-art techniques to simulate large scale quantum circuits. PySA is an extensible platform to optimize a classical cost function. We provide an outline of each of these open-source tools and highlight projects using each of these tools in contexts of simulation, optimization and machine learning.

Quantum Computing↗

Space Networking Implementation for Lunar Operations

The High-Rate Delay Tolerant Networking (HDTN) project at NASA has developed a performance optimized and open-source Delay Tolerant Networking (DTN) implementation. The primary goal is to create a scalable networking solution to increase the scientific data return rate of space missions. To reach this goal, HDTN must span multiple edge cases in space networking by including tools and configurations to accommodate a wide range of space systems. Typically, HDTN evaluations are conducted on a laboratory emulation test bed, made up of hardware accelerated x86 based systems capable of data rates over 10 Gbps. HDTN must have an effective implementation process on a wide range of systems to increase the sustainability of the design. One important implementation option is with low-level embedded systems which could be used on small robotic missions. This paper details the implementation process, benchmark testing, and performance results of HDTN in multiple configurations on Raspberry Pi 4 devices. By implementing HDTN on a Raspberry Pi 4, a process for building HDTN onto ARM processors was developed and utilized to conduct benchmark tests in multiple network configurations, achieving a data rate performance exceeding 600 Mbps. Based on these results, HDTN proved to run on small ARM based systems with slight modifications to the build procedure. These results were then extended to evaluating an implementation of the HDTN software parsed across several Raspberry Pi 4 nodes. To test this capability, HDTN was configured in a simplified cut-through setup and distributed among multiple Raspberry Pi 4 processors. This distributed architecture was benchmark tested in a similar fashion to the testing of a singular HDTN implementation. The results from the benchmark testing are used to examine how these implementation options and capabilities can expand the use cases for DTN, and particularly with small robotic missions.

Space Networking↗

FatMan vs. LittleBoy: Scaling up Linear Algebraic Operations in Scale-out Data Platforms

Linear algebraic operations such as matrix manipulations form the kernel of many machine learning and other crucial algorithms. Scaling up as well as scaling out such algorithms are highly desirable to enable efficient processing over millions of data points. To this end, we present a matrix manipulation approach to effectively scale-up each node in a scale-out data parallel platform such as Apache Spark. Specifically, we enable hardware acceleration for matrix multiplications in a distributed Spark setup without user intervention. Our approach supports both dense and sparse distributed matrices, and provides flexible control of acceleration by matrix density. We demonstrate the benefit of our approach for generalized matrix multiplication operations over large matrices with up to four billion elements. To connect the effectiveness of our approach with machine learning applications, we performed Gramian matrix computation via generalized matrix multiplications. Our experiments show that our approach achieves more than 2x performance speed-up, and up to 96.1% computation improvement, compared to a state of the art Spark MLlib for dense matrices.

Lim, Seung-Hwan↗

SODA: a New Synthesis Infrastructure for Agile Hardware Design of Machine Learning Accelerators

Next generation systems, such as edge devices, will have to provide efficient processing of machine learning (ML) algorithms along several metrics, including energy, performance, area, and latency. However, the quickly evolving field of ML makes it extremely difficult to generate accelerators able to support a wide variety of algorithms. At the same time, designing accelerators in hardware description languages (HDLs) by hand is hard and time consuming, and does not allow quick exploration of the design space. This paper discusses the SODA synthesizer, an automated open source high-level ML framework-to-Verilog compiler targeting ML Application-Specific Integrated Circuits (ASICs) chiplets based on the LLVM infrastructure. The SODA synthesizers will allow implementing optimal designs by combining templated and fully tunable IPs and macros, and fully custom components generated through high-level synthesis. All these components will be provided through an extendable resource library, characterized with both commercial and open source logic design flows. Through a closed loop design space exploration engine, developers will be able to quickly explore their hardware designs along different dimension

Minutoli, Marco↗

Synchronization between processes in a coordination namespace

A system and method of supporting point-to-point synchronization among processes/nodes implementing different hardware barriers in a tuple space/coordinated namespace (CNS) extended memory storage architecture. The system-wide CNS provides an efficient means for storing data, communications, and coordination within applications and workflows implementing barriers in a multi-tier, multi-nodal tree hierarchy. The system provides a hardware accelerated mechanism to support barriers between the participating processes. Also architected is a tree structure for a barrier processing method where processes are mapped to nodes of a tree, e.g., a tree of degree k to provide an efficient way of scaling the number of processes in a tuple space/coordination namespace.

Jacob, Philip↗

Pele: An Exascale-Ready Suite of Combustion Codes

High fidelity simulations of realistic combustion devices are extremely demanding computationally because of the requirements to capture complex fuel chemical decomposition, its intricate interactions with turbulent, often multiphase, flows, and the wide separation of space and time scales between the thin flame and the device boundaries. Software required to carry out such computations tends to be extremely complex, particularly when designed to exploit hardware accelerators, and can be difficult to port and maintain. We present Pele, a performance portable suite of tools for the simulation of combustion systems, including codes to evolve reactive multiphase configurations in the low Mach number and compressible flow regimes, along with a set of inter-compatible post processing and in situ analysis tools. The Pele suite of tools is built on top of the AMReX framework for block-structured adaptive mesh refinement, which provides efficient data structures and algorithms that enable the development of a wide variety of efficient mesh and particle based PDE integration schemes. A hierarchical MPI+X parallelism scheme supports CPU-only and accelerated architectures, where X can be OpenMP, CUDA, and HIP based approaches for intra-node computational work distribution. The algorithms and data structures underlying the Pele simulation and analysis tools are highly scalable and performant across a wide variety of high-performance computing platforms, including DOEs newest exascale-class machines, Frontier and Aurora. The simulation and analysis tools are fully documented and freely distributed as open source via GitHub. We present key algorithmic and software challenges, solution strategies, performance and resulting set of capabilities.

AMReX↗

FPGA-based computing system for processing data in size, weight, and power constrained environments

Technologies that are well-suited for use in size, weight, and power (SWAP)-constrained environments are described herein. A host controller dispatches data processing instructions to hardware acceleration engines (HAEs) of one or more field programmable gate arrays (FPGAs) and further dispatches data transfer instructions to a memory controller, such that the HAEs perform processing operations on data stored in local memory devices of the HAEs in parallel with other data being transferred from external memory devices coupled to the FPGA(s) to the local memory devices.

Napier, Matthew↗

Synchronization between processes in a coordination namespace

A system and method of supporting point-to-point synchronization among processes/nodes implementing different hardware barriers in a tuple space/coordinated namespace (CNS) extended memory storage architecture. The system-wide CNS provides an efficient means for storing data, communications, and coordination within applications and workflows implementing barriers in a multi-tier, multi-nodal tree hierarchy. The system provides a hardware accelerated mechanism to support barriers between the participating processes. Also architected is a tree structure for a barrier processing method where processes are mapped to nodes of a tree, e.g., a tree of degree k, to provide an efficient way of scaling the number of processes in a tuple space/coordination namespace.

Jacob, Philip↗

Development of ARCHER towards clinical use: modeling and simulation of Varian Linac for radiation therapy dose calculations - 25449

Monte Carlo particle transport methods are well recognized as the most accurate approach to dose calculations for radiation therapy treatment planning. However, the lengthy calculation time has prevented Monte Carlo methods from being applied to dose calculations in the planning of radiation therapy. Recently, the emerging parallel computing hardware accelerators, such as general-purpose Graphics Processing Unit (GPU), have significantly boosted the performance of Monte Carlo codes. The accuracy of a clinical dose engine depends greatly on the accuracy of source term modeling and the methodology of dose calculations. Hence, in this study, an efficient and accurate source model of the Varian TrueBeam{sup R} Linac is developed and integrated with a fast photon-electron transport module of the recently developed high-performance Monte Carlo code, ARCHER. The validity of the source model and the particle transport physics are verified with established dose calculation codes such as EGSnrc and validated with experimental measurements. To demonstrate the capability of the platform in clinical applications, three clinical treatment plans (the lung, breast, and prostate) are simulated and analyzed. From the perspective of performance, the entire simulation of each clinical plan, starting from the source modeling to the end of particle transport within the patient phantom, can be finished in less than one minute using a single Titan X GPU card with 1% statistical error. In comparison, the same cases take the EGSnrc code hundreds of CPU hours to finish. The results suggest that this new version of ARCHER code has been successfully developed and tested as a clinically deployable dose engine for external photon beam IMRT plans in the clinical setting. (authors)

61 RADIATION PROTECTION AND DOSIMETRY↗

Preliminary performance evaluation of P++ single event proton scattering algorithms on GPUS - 25688

A need has been realized for fast, accurate charged particle transport in a wide variety of domains. Utilizing hardware acceleration through GPUs now common in user desktops, speedups in simulation comparable to those found with HPC systems can be realized for a fraction of the price. Preliminary results on the P++ proton transport code developed by the Rensselaer Radiation Measurement and Dosimetry Group (RRMDG) are presented. The code has been validated for the 1 GeV to 1 MeV energy regime and is targeted to benefit application areas such as proton radiography, space radiation research, and clinical accelerator therapy. The code is notable for employing a single event, 'on the fly' cross section based transport algorithm rather than the traditional condensed history or step based method seen in many production Monte Carlo codes. Performance improvements for the P++ code have been introduced through MPI parallelization as well as a CUDA version of the transport algorithm capable of running on NVidia GPUs. (authors)

61 RADIATION PROTECTION AND DOSIMETRY↗

MoE-Inference-Bench: Performance Evaluation of Mixture of Expert Large Language and Vision Models

Mixture of Experts (MoE) models have enabled the scaling of Large Language Models (LLMs) and Vision Language Models (VLMs) by achieving massive parameter counts while maintaining computational efficiency. However, MoEs introduce several inference-time challenges, including load imbalance across experts and the additional routing computational overhead. To address these challenges and fully harness the benefits of MoE, a systematic evaluation of hardware acceleration techniques is essential. We present MoE-Inference-Bench, a comprehensive study to evaluate MoE performance across diverse scenarios. We analyze the impact of batch size, sequence length, and critical MoE hyperparameters such as FFN dimensions and number of experts on throughput. We evaluate several optimization techniques on Nvidia H100 GPUs, including pruning, Fused MoE operations, speculative decoding, quantization, and various parallelization strategies. Our evaluation includes MoEs from the Mixtral, DeepSeek, OLMoE and Qwen families. The results reveal performance differences across configurations and provide insights for the efficient deployment of MoEs.

Chitty-Venkata, Krishna Teja↗

High Performance Computing Management: A Sustainable System Software Approach

The demand for high performance computing (HPC) resources continues to grow, driven by the increasing complexity of modeling and simulation, artificial intelligence (AI), and machine learning (ML) workloads [Porter]. The growing energy consumption demand of these HPC systems is a significant concern, both in terms of operational costs and environmental impact. AI hardware accelerators are expected to reach 1.5% of the world’s power consumption by 2029 [Shah].

97 - MATHEMATICS AND COMPUTING↗

Data Processing Unit Services Module

The Data Processing Services Module (DPUSM) provides the ability to perform pluggable compression, erasure coding, checksuming and other important file system operations within the Linux kernel. The pluggable provider interface allows for the use of hardware acceleration of those services. In-kernel file systems are then able to use these functions to use these accelerators to perform operations that are normally run on the processor, resulting in improved file system performance. Third parties will register "providers" with the DPUSM to communicate with their respective accelerators. Providers will implement functions with DPUSM API signatures so that the DPUSM can translate the data inputted by users of the DPUSM into data that providers recognize.

Lee, Jason↗

hls4ml: A Flexible, Open-Source Platform for Deep Learning Acceleration on Reconfigurable Hardware

We present hls4ml, a free and open-source platform that translates machine learning (ML) models from modern deep learning frameworks into high-level synthesis (HLS) code that can be integrated into full designs for field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs). With its flexible and modular design, hls4ml supports a large number of deep learning frameworks and can target HLS compilers from several vendors, including Vitis HLS, Intel oneAPI and Catapult HLS. Together with a wider eco-system for software-hardware co-design, hls4ml has enabled the acceleration of ML inference in a wide range of commercial and scientific applications where low latency, resource usage, and power consumption are critical. In this paper, we describe the structure and functionality of the hls4ml platform. The overarching design considerations for the generated HLS code are discussed, together with selected performance results.

FOS: Computer and information sciences↗

hls4ml: A Flexible, Open-Source Platform for Deep Learning Acceleration on Reconfigurable Hardware

We present hls4ml, a free and open-source platform that translates machine learning (ML) models from modern deep learning frameworks into high-level synthesis (HLS) code that can be integrated into full designs for field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs). With its flexible and modular design, hls4ml supports a large number of deep learning frameworks and can target HLS compilers from several vendors, including Vitis HLS, Intel oneAPI and Catapult HLS. Together with a wider eco-system for software-hardware co-design, hls4ml has enabled the acceleration of ML inference in a wide range of commercial and scientific applications where low latency, resource usage, and power consumption are critical. In this paper, we describe the structure and functionality of the hls4ml platform. The overarching design considerations for the generated HLS code are discussed, together with selected performance results.

Schulte, Jan-Frederik [Purdue U.] (ORCID:000000034↗