Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Computational pipelines”

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 523 records · Page 29
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Wide Bandgap Generation (WBGen): Developing the Future Wide Bandgap Power Electronics Engineering Workforce

This Final Technical Report (FTR) summarizes the work conducted at the Center for Power Electronics Systems (CPES) under the Wide-Bandgap Generation (WBGen) fellowship and traineeship program established by the Advanced Manufacturing and Technologies Office (AMMTO) of the U.S. Department of Energy Office of Energy Efficiency and Renewable Energy (EERE) at Virginia Tech, which had as main objective to train the next generation of U.S. citizen power engineers with wide-bandgap (WBG) power semiconductor expertise, with the intent to aid in fulfilling the future workforce needs in this field. The latter was deemed of strategic importance given the fast-paced growth observed—and predicted—in the demand of this technical expertise, whose practitioners have become the enablers and executioners of the electrification transformation process that not just the U.S., but the whole world, is currently undergoing as it seeks for more effective and efficient ways to use energy. As such, the WBGen program set forth to achieve its educational goals, which in addition sought to broaden the range of WBG-based power electronics by conducting research and development on high-efficiency grid apparatus and high-efficiency electrical power systems, and to also enhance the power engineering curriculum by formalizing WBG-oriented design procedures replacing existent yet now obsolete design procedures developed for Silicon (Si) based power electronics. This effort led CPES to spearhead the development of a new major within The Bradley Electrical and Computer Engineering (ECE) Department at Virginia Tech, namely Electronic Power and Energy Systems (EPES), which coalesced power electronics and power systems courses to provide undergraduate students with a strong formation in the power engineering field, while creating a pipeline of graduate students that could pursue the WBG-based curriculum and conduct research at CPES. In all, in what is considered a true success, eight of the twenty WBGen fellows that graduated program were recruited from the ECE department undergraduate cohort. The traineeship emphasized as well, from its beginning, the partnership with industry and national laboratories, which took advantage of the successful industry consortium at CPES that has historically been formed by 80–90 power and energy companies working in close collaboration with the center. This gave WBGen fellows the accessibility and possibility to conduct internships at partner facilities during the summer months, focused solely on the evaluation, testing, and adoption of WBG devices, which were many times tightly related to their respective research work and plans. In addition, WBGen fellows conducted their main research work within the confines of research programs at CPES conducted with these industry partners, providing them with a unique opportunity to develop not just their technical expertise—while advancing their knowledge, but to also learn and practice a slew of skills needed for their professional growth. As such, the fellows tackled a variety of WBG-related research topics, from device reliability and capability aspects as well as packaging and integration, encompassing the use of advanced materials and new structures, current sharing challenges, and insulation systems, to advanced gate-drivers with integrated sensors and protection mechanisms and active current- and voltage-based control, to optimized layouts seeking to maximize the switching and power processing performance of these devices, to power processing solutions adopting Gallium-Nitride (GaN) and Silicon-Carbide (SiC) power semiconductors for a variety of applications; including radiation-hardened converters for space dc distribution systems, direct three-phase ac-to-ac power converters for aerospace systems, dc-ac inverters for automotive traction drives, high-frequency isolated dc-dc battery chargers also for heavy transportation systems, and medium-voltage dc-dc and dc-ac converters for distribution systems and future power grids. The WBGen program ultimately graduated a total of 20 power engineers, all experts on WBG-based power electronics, awarding 18 M.S. and 2 PhD degrees in the process. These fellows, all U.S. citizens, allowed CPES to increase the number of citizens students to 33 % at the peak of the program, as the traineeship made possible the recruitment of talent with more attractive graduate research assistantship (GRA) contracts. Unfortunately, the present U.S. citizen enrollment at CPES has declined back to historic levels—approximately 10 %, as the regular GRA rates are not competitive enough when compared with entry-level industry jobs. In all, WBGen fellows published a total of 7 peer-reviewed journal articles, 44 papers at international technical conferences, made 42 presentations at international technical conferences, and filed 4 invention disclosures and patent applications, which have since then been granted by the U.S. Patent and Trademark Office (USPTO). Their contribution to CPES, Virginia Tech, the United States, and the world, has been significant, and continues to yield results thanks to the exemplary career that the fellows have initiated at many of the partners of the program, which include Wolfspeed, Raytheon Technologies, Infineon, Lockheed Martin, Dominion Energy, Northrop Grumman, Rivian, Aerospace Corporation, Sandia National Laboratory, National Renewable Energy Laboratory, John Hopkins University, and Virginia Tech.

14 SOLAR ENERGY↗
NASA NTRS1990

High performance pipelined multiplier with fast carry-save adder

A high-performance pipelined multiplier is described. Its high performance results from the fast carry-save adder basic cell which has a simple structure and is suitable for the Gate Forest semi-custom environment. The carry-save adder computes the sum and carry within two gate delay. Results show that the proposed adder can operate at 200 MHz for a 2-micron CMOS process; better performance is expected in a Gate Forest realization.

Wu, Angus↗
DOE OSTI2026

Expanding the Domain of Applicability of Machine Learning Models with Limited Data for Drug Property Prediction

Accurate machine learning models for predicting small molecule interactions with biological targets are essential for therapeutic discovery, biothreat response, and computational drug design, but their performance is often limited for understudied targets with sparse experimental data. To address this challenge, we developed and evaluated methods to improve molecular property prediction under low-data conditions, using the NimA-related kinase (NEK) family as a proof-of-concept. This work focused on two complementary goals within the ATOM Modeling PipeLine (AMPL) and the Generative Molecular Design (GMD) loop: expanding model applicability through transfer learning, representation learning, feature scaling, sampling strategies, and active-learning-inspired compound selection; and enabling efficient virtual screening to prioritize compounds that balance predicted activity, design objectives, and synthetic accessibility.

organic↗
DOE OSTI2022

The Status and Ambitions of the US Heavy Element Program

The aim of this whitepaper is to highlight the current capabilities and priorities of the US Heavy Element community and to provide the framework for a coordinated advancement of nuclear science from these studies. This is an organized effort to reflect on what has been achieved in the field given the recommendations and initiatives of the 2015 NSAC Long Range Plan, and on what can be realized in the next decade given current and possibly expanded investments. Current investments have positioned the US community to be among the world leaders in studies of the nuclear and chemical properties of the heaviest elements. These include studies of reaction mechanisms, moving us ever closer to the “island of stability”, in spectroscopy, allowing us to better understand nuclear structure at these extreme proton numbers, in chemical behavior, looking to determine how these elements should be placed on the Periodic Table, in performing the first measurements where isotopes are directly identified by their mass numbers, and in laying the foundation for a potential US-led new element discovery experiment. At present, the highest priority of the US Heavy Element community is to capitalize on the current investments by supporting the operations of US facilities at optimal values. These facilities include the Argonne Tandem Linac Accelerator System at Argonne National Laboratory and other Department of Energy facilities such as the 88-Inch Cyclotron at Lawrence Berkeley National Laboratory, which has a dedicated superheavy element program, as well as university laboratories, including Texas A&M University. The High Flux Isotope Reactor at Oak Ridge National Laboratory is crucial to providing the radioactive isotopes required for heavy element science targets. This facility should be supported to provide the actinide materials that are essential for US-based science. Production of stable, rare isotopes for beam material, including 48 Ca, 50 Ti, 54 Cr and 58 Fe, at the Stable Isotope Production and Research Center is critical to continued research in heavy element science and should be a priority. The continued development of targets for heavy element science and retaining US-based expertise is critical for the heavy element community. This is an area that is currently under pressure. For example, the target laboratory at Argonne National Laboratory serves a broad community and is currently under threat due to loss of critical personnel. The skills needed to make targets for nuclear science and develop new targetry methods need to be supported long term at Argonne, Oak Ridge and Lawrence Livermore National Laboratories as well as maintaining the programs at Oregon State University and San José State University as vital pipelines for training students. Advances in theory are the foundation to understand how nuclei behave and to predict those behaviors in new circumstances. Progress in these studies will necessitate continued and new investment and access to high-performance computing. The future health of the heavy element field is dependent on the continuous support of talented early-career professionals at all levels. It is critical that opportunities continue to be created for the next generation to become established in heavy element research so that we can ensure the field is attracting and retaining the best minds for continued success. It is also clear that to ensure diversity of ideas, perspectives and techniques, we need to recruit diverse personnel that are trained at the best facilities. The heavy element community is in support of continued investment to programs with initiatives in diversity, equity, and inclusion. Looking to the next decade of research, support needs to maintain and grow US leadership in heavy element science. Specifically, new investments in state-of-the-art instrumentation will be essential to scientific development of the field and in expanding scientific knowledge. Advances in the next generation of electron cyclotron resonance ion sources, multi-reflection time-of-flight devices, laser spectroscopy, trapping methods and next generation alpha and gamma spectroscopy systems should be prioritized.

38 RADIATION CHEMISTRY, RADIOCHEMISTRY, AND NUCLEA↗
DOE OSTI2026

Intelligent Experiments through Real-Time AI: Fast Data Processing and Autonomous Detector Control for High-Energy Nuclear Experiments

The aim of this project is to develop software and hardware for fast real-time data processing and autonomous detector control and calibration for the sPHENIX and the future EIC experiments. Below summarizes Georgia Tech team efforts in the past year: 1. We developed a real-time clustering algorithm and FPGA-based pipeline architecture for processing fired pixel data from ALPIDE sensors in sPHENIX experiments. Our Columnar Clustering Co-Design introduces a hardware-aware, stream-friendly approach that segments pixel data by column pairs using a Column Pair Clustering (CPC) strategy, followed by Cluster Stitching to merge adjacent subclusters. Implemented in Vitis HLS, the pipeline comprises five stages—read-in, subclustering, stitching, analysis, and write-out—connected by tagged HLS streams with custom end-of-event signaling for robust synchronization. We designed a pipelined dataflow model optimized for throughput, low latency, and minimal buffering, enabling scalable clustering across events of arbitrary size. Our system maintains spatial precision via center-of-mass and shape key extraction and efficiently handles edge cases such as fragmented or nested clusters. Compared against DBSCAN in both software and hardware, our approach demonstrates competitive performance under FPGA constraints. 2. We also conducted a comprehensive algorithm-to-hardware co-design of connected component analysis tailored for sPHENIX experiments, focusing on real-time, low-latency processing using FPGAs and High-Level Synthesis (HLS). Starting from a Python-based particle tracking pipeline, the team translated the core logic—graph traversal via DFS and Union-Find—into an HLS-compatible C++ model, replacing dynamic memory and recursion with static arrays and pipelined control flow. The final design includes a fully streamed and dataflow-compatible Union-Find kernel optimized across five iterations, incorporating loop pipelining, array partitioning, AXI/FIFO interface tuning, and function flattening. Experimental results show up to 14.8× speedup over the CPU baseline, reducing per-graph latency to 1.58 μs and demonstrating strong resource efficiency with only ~7k LUTs and zero BRAM usage. The design maintains functional correctness against the Python reference using a Python-based C-simulation framework and Mean Squared Error metrics. This work validates the potential of HLS-driven FPGA designs for edge-level HEP data acquisition, laying a scalable foundation for future integration with real-time detector pipelines and multi-graph processing systems.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2023

Wildfire Smoke PM2.5 per Zipcode

This dataset contains daily aggregated measurements of daily PM2.5 from ambient wildfire smoke in the contiguous United States, spanning from 2006 to 2016. The data is sourced from a study by Childs et al. (2022), titled "Daily Local-Level Estimates of Ambient Wildfire Smoke PM2.5 for the Contiguous US" published in Environmental Science & Technology. To compute the standardized weight across different zip codes, we computed the weight calculation on a 10km grids. To compute the weight, we obtain area-weights that add up to 1 in each polygon of zipcodes. This enabled us to calculate the smoke values per zipcode for the aforementioned period. Those interested in replicating our data processing pipeline can access it at https://github.com/NSAPH-Data-Processing/smoke_aggregation.

PM2.5↗
DOE OSTI2020

MARTINI-Compatible Coarse-Grained Model for the Mesoscale Simulation of Peptoids

Peptoids (poly-N-substituted glycines) are a class of synthetic polymers that are regioisomers of peptides (poly-C-substituted glycines), in which the point of side-chain connectivity is shifted from the backbone C to the N atom. Peptoids have found diverse applications as peptidomimetic drugs, protein mimetic polymers, surfactants, and catalysts. Computational modeling is valuable in the understanding and design of peptoid-based nanomaterials. In this work, we report the bottom-up parameterization of coarse-grained peptoid force fields based on the MARTINI peptide force field against all-atom peptoid simulation data. Our parameterization pipeline iteratively refits coarse-grained bonded interactions using iterative Boltzmann inversion and nonbonded interactions by matching the potential of mean force for chain extension. We assure good sampling of the amide bond cis/trans isomerizations in the all-atom simulation data using parallel bias metadynamics. We develop coarse-grained models for two representative peptoids—polysarcosine (poly(N-methyl glycine)) and poly(N-((4-bromophenyl)ethyl)glycine)—and show their structural and thermodynamic properties to be in excellent accord with all-atom calculations but up to 25-fold more efficient and compatible with MARTINI force fields. Here, this work establishes a new rigorously parameterized coarse-grained peptoid force field for the understanding and design of peptoid nanomaterials at length and time scales inaccessible to all-atom calculations.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2025

Combining computational modeling and experimental library screening to affinity-mature VEEV-neutralizing antibody F5

Engineered monoclonal antibodies have proven to be highly effective therapeutics in recent viral outbreaks. However, despite technical advancements, an ability to rapidly adapt or increase antibody affinity and by extension, therapeutic efficacy, has yet to be fully realized. We endeavored to stand-up such a pipeline using molecular modeling combined with experimental library screening to increase the affinity of F5, a monoclonal antibody with potent neutralizing activity against Venezuelan Equine Encephalitis Virus (VEEV), to recombinant VEEV (IAB) E1E2 antigen. We modeled the F5/E1E2 binding interface and generated predictions for mutations to improve binding using a Rosetta-based approach and dTERMen, an informatics approach. The modeling was complicated by the fact that a high-resolution structure of F5 is not available and the H3 loop of F5 exceeds the length for which current modeling approaches can determine a unique structure. A subset of the predicted mutations from both methods were incorporated into a phage display library of scFvs. This library and a library generated by error-prone PCR were screened for binding affinity to the recombinant antigen. Results from the screens identified favorable mutations which were incorporated into 12 human-IgG1 variants. The best variant, containing eight mutations, improved KD from 0.63 nM (parental) to 0.01 nM. While this did not improve neutralization or therapeutic potency of F5 against IAB, it did increase cross-reactivity to other closely related VEEV epizootic and enzootic strains, demonstrating the potential of this method to rapidly adapt existing therapeutics to emerging viral strains.

affinity-maturation↗
DOE OSTI2026

Data-Driven Performance Optimization of Gamma Spectrometers With Many Channels

In gamma spectrometers with variable spectroscopic performance across many channels (e.g., many pixels or voxels), a tradeoff exists between including data from successively worse-performing readout channels and increasing efficiency. Brute-force calculation of the optimal set of included channels is exponentially infeasible as the number of channels grows, and approximate methods are required. In this work, we present a data-driven framework for attempting to find near-optimal sets of included detector channels. The framework leverages non-negative matrix factorization (NMF) to learn the behavior of gamma spectra across the detector and clusters similarly-performing detector channels together. Performance comparisons are then made between spectra with channel clusters removed, which is more feasible than brute force. The framework is general and can be applied to arbitrary, user-defined performance metrics depending on the application. We apply this framework to optimizing gamma spectra measured by H3D M400 CdZnTe (CZT) spectrometers, which exhibit variable performance across their crystal volumes. In particular, we show several examples optimizing various performance metrics for uranium and plutonium gamma spectra in non-destructive assay (NDA) for nuclear safeguards, and explore trends in performance versus parameters such as clustering algorithm type. We also compare the NMF + clustering pipeline to several non-machine-learning (ML) algorithms, including several greedy algorithms. Although, we find that the NMF + clustering pipeline tends to find the best-performing set of detector voxels, significantly improving over the unoptimized spectra, but that a greedy accumulation of spectra segmented by detector depth can, in some cases, give similar performance improvements in much less computation time.

Energy resolution↗
DOE OSTI2026

lumicap v0.1

Automated HDR luminance imaging system designed for daylighting research and building science. It controls a fisheye-lens camera to capture time-lapse bracket sequences, merges them into calibrated HDR images, and runs a full post-processing pipeline — all unattended. Features: - Scheduled LDR bracket capture via gphoto2 - HDR merging with vignetting, ND filter, and fisheye projection corrections - Illuminance and luminance meter integration (Konica Minolta T-10A, LS-100/150) - Daylight glare probability (DGP) and solar position computation - Automated false-color rendering, JPEG thumbnails, and daily time-lapse video - CSV data logging per timestep Uses: - Long-term monitoring of daylight conditions in buildings - Glare analysis for occupant comfort research - Solar irradiance and sky luminance studies Advantages: - End-to-end automation — capture, calibration, analysis, and archiving run without manual intervention - Built on the proven Radiance toolchain, ensuring photometrically accurate HDR output - Hardware-agnostic meter support via serial auto-detection - Lightweight — no GUI overhead, deployable on a headless Raspberry Pi or similar embedded system

Wang, Taoning [Lawrence Berkeley National Laborato↗
DOE OSTI2026

Detector Interface for Streaming, Control, and Open-source integration (DISCO) v1.0.0

This suite consists of a multi-package ecosystem featuring detector emulators, EPICS areaDetector drivers, and remote server frameworks designed for the Advanced Light Source (ALS). Engineered for high-bandwidth devices—including VFCCD, Timepix3, Timepix4, and related pixel detectors—the software simulates hardware, wraps vendor SDKs into remote-callable servers, and integrates with open-source control systems. Key Capabilities: Distributed SDK Architecture: Server packages wrap hardware-specific SDKs, allowing areaDetector drivers to execute remote framework calls. This isolates proprietary libraries from the EPICS IOC, enhancing stability and enabling distributed computing across beamline networks. Device Support: Custom drivers for VFCCD, the Timepix family, and similar sensors optimize the data path from hardware control to high-speed transport. Full-Stack Emulation: Sophisticated emulator packages allow end-to-end pipeline testing and software development without requiring physical hardware or beam time. Integrated Workflows: Supports high-bandwidth streaming for real-time analysis and robust, metadata-rich file-based workflows (e.g., HDF5/NeXus). By standardizing interfaces across heterogeneous hardware, this suite reduces technical debt. It provides the ALS with a scalable, open-source solution to manage massive data rates within a unified control environment.

Mahl, Johannes [Lawrence Berkeley National Laborat↗
NASA NTRS1990

Feasibility study, software design, layout and simulation of a two-dimensional fast Fourier transform machine for use in optical array interferometry

The NASA-Cornell Univ.-Worcester Polytechnic Institute Fast Fourier Transform (FFT) chip based on the architecture of the systolic FFT computation as presented by Boriakoff is implemented into an operating device design. The kernel of the system, a systolic inner product floating point processor, was designed to be assembled into a systolic network that would take incoming data streams in pipeline fashion and provide an FFT output at the same rate, word by word. It was thoroughly simulated for proper operation, and it has passed a comprehensive set of tests showing no operational errors. The black box specifications of the chip, which conform to the initial requirements of the design as specified by NASA, are given. The five subcells are described and their high level function description, logic diagrams, and simulation results are presented. Some modification of the Read Only Memory (ROM) design were made, since some errors were found in it. Because a four stage pipeline structure was used, simulating such a structure is more difficult than an ordinary structure. Simulation methods are discussed. Chip signal protocols and chip pinout are explained.

Boriakoff, Valentin↗
DOE OSTI2026

Selection of a Pair of Experiments to Optimally Reduce Uncertainty in Targeted Nuclear Data

We propose a novel process to select a pair of differential and integral experiments that best reduce uncertainties in targeted 239 ⁢Pu nuclear data while compressing the current nuclear data pipeline from 20 to 3 years. 239⁢ Pu nuclear data are poorly understood for neutrons in the intermediate energy range due to sparsity and uncertainty in historical experiments. New experiments targeting this range will enable better understanding of these nuclear data, but choosing the ideal experiments to conduct is challenging. Beginning with a prior distribution represented by samples of nuclear data generated from theory, generalized least squares adjustments are made to incorporate data from historical experiments. To quantify potential uncertainty reduction obtainable from a pair of candidate experiments, we compute the D-optimality criterion of the posterior covariance of intermediate energy range nuclear data compared to the equivalent covariance after additional adjustment to the pair of candidate experiments. Repeating the process for each of many candidate pairs facilitates the final selection. Results support 63⁢ Cu total cross section measurements for differential experiments and alumina and alumina/graphite configurations for integral experiments. This analysis enables choosing differential and integral experiments to be executed concurrently while shortening decision times relative to the current nuclear data pipeline.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2025

An Open Benchmark of One Million High-Fidelity Cislunar Trajectories

Cislunar space spans from geosynchronous altitudes to beyond the Moon and will underpin future exploration, science, and security operations. We describe and release an open dataset of one million numerically propagated cislunar trajectories generated with the open-source Space Situational Awareness Python package (SSAPy). The model includes high-degree Earth/Moon gravity, solar gravity, and Earth/Sun radiation pressure; other planetary gravities are omitted by design for computational efficiency. Initial conditions uniformly sample commonly used osculating-element ranges, and each trajectory is propagated for up to six years under a single, fixed start epoch. The dataset is intended as a reusable benchmark for method development (e.g., space domain awareness, navigation, and machine-learning pipelines), a reference library for statistical studies of orbit families, and a starting point for community-driven extensions (e.g., alternative epochs). We report empirically observed stability trends (e.g., a band near ~5 GEO and persistence of some co-orbital classes including L4/L5 librators) as dataset descriptors rather than new dynamical results. The chief contribution is the scale, fidelity, organization (CSV/HDF5 with full state time series and metadata), and open availability, which together lower the barrier to comparative and data-driven studies in the cislunar regime.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2024

Complete and Correct Transfer of Information (CACTI)

Many distributed systems, file transfer mechanisms, and message passing systems offer reliability mechanisms such as acknowledgements, retries, and durability. While these tools may be “good enough” for their typical use cases, they may not offer sufficient coverage for the wide range of faults that impact data transfers and communication. A gap in the reliability measures may lead to some small amount of data loss. Some high-consequence systems cannot tolerate the loss or corruption of even a single record. We present seven principles that will counter a wide range of faults and protect against data loss and corruption. These principles bring together lessons learned from a wide range of technologies and can inform appropriate system design and application usage. These principles will help readers reason on how prevent data loss in a multi-hop pipeline and how to properly use tools that may have a deficiency in reliability.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Benchmark Tracking System for Performance Monitoring

Benchmarking is essential for high-performance software development, particularly for monitoring performance across code iterations. This project focused on enhancing the benchmarking process for Lamellar, an asynchronous runtime for High-Performance Computing (HPC) systems developed at Pacific Northwest National Laboratory. Prior to this work, benchmark results were difficult to track and compare across code versions, presenting significant challenges in identifying performance regressions and long-term trends. The primary objective was to establish a systematic, reproducible approach for measuring performance and detecting regressions following code commits. Our methodology involved three key components: standardizing benchmark outputs, implementing data versioning, and developing analysis tools. We standardized the benchmark output format to JSON Line records containing specific fields (execution time, hardware specifications, and environmental variables). To address data management challenges, we evaluated several options and eventually chose a git repository dedicated to benchmark data. We developed a suite of Python tools that processed benchmark results, enriched them with metadata, and facilitated search in the repository. The resulting system enables more efficient filtering and comparison of performance metrics across commit histories, hardware configurations, and benchmark variants through a unified query interface. Our implementation reduces computational overhead by first checking for existing results through configuration matching before initiating new benchmark runs, thereby conserving resources. The system has been validated by Lamellar developers. It organizes results by benchmark type and build configurations for efficient retrieval. Future developments include a planned Large Language Model interface for predicting benchmark performance, incorporating the criterion package for statistical analysis, which will enable automated detection of statistically significant performance changes, and integration with continuous integration pipelines. Despite these enhancements being reserved for future work, this project has successfully provided the Lamellar development team with a framework for maintaining consistent performance standards and identifying optimization opportunities across workloads and hardware environments.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

FLASH: FPGA-Accelerated Smart Switches with GCN Case Study

Some communication switches, e.g., the Mellanox SHArP and those in the IBM BlueGene clusters, are augmented to process packets at the application level with fixed-function collectives. This approach, however, lacks flexibility, which limits their applicability in diverse and dynamic workloads. Recently, a new type of programmable packet processor, which uses high-level languages, e.g., P4, has emerged as possible candidates. P4-based switches, however, fall short in certain applications, including machine learning, where capabilities not currently supported by P4 are needed. These include more complex calculation, such as sparse computation and fused multiply-accumulate, data-intensive floating point operations, data reuse, and significant memory. The problem addressed here is that such a switch augmentation needs to support: a large amount of state, significant flexible compute capability, and ease of programming, all while maintaining full functionality, including ensuring high throughput, and demonstrating utility. In this work, we propose a programmable look-aside-type accelerator that can be embedded into, or attached to, existing communication switch pipelines and that is capable of processing packets at line-rate. The proposed in-switch accelerator is based on mixing an ISA (subset of RISC-V instructions) with dataflow graphs (found in CGRAs). To augment performance, vector instructions are also supported. To facilitate usability, we have developed a complete toolchain to compile user-provided C/C++ codes to appropriate back-end instructions for configuring the accelerator. While this approach is flexible enough to support various workloads, in this paper, we consider Graph Convolutional Networks (GCNs) as a case study. Experimental results show that this approach considerably improves the performance of distributed GCN applications.

Haghi, Pouya↗