Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “INTERCONNECTIONS”

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 829 records · Page 46
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Keystone Solar Future Project (KSFP)

The Keystone Solar Future Project (KSFP, also referred to as “Keystone Project” in the report) implements a cost-effective, secure, reliable, and safe technology platform that paves the way for future interconnected distributed energy resources (DER), creates a framework that enables high penetration of DER, and transforms the DER interconnection process. The project is important and necessary in a few aspects. First, the industry definition of Distributed Energy Resource Management System (DERMS) is not yet validated against field deployments and mostly implemented in small pilots relying on third party integration. Keystone project implements a direct plug and play communication to inverter based DER systems. Secondly, most interconnection application portal is not yet fully integrated with back-end system, thus lacks the capability to drive a complete and accurate modeling of DER assets into the planning and operational systems. Keystone project rolled out a fully integrated modeling process of DERs from interconnection to operational management. Lastly, customers were given an opportunity to learn about the advanced functions of the inverters and take full advantages of their benefits through participation in the pilot program. During the project, the technical effectiveness and economic feasibility of DER management were investigated and demonstrated in real life events and simulation studies. The project highlighted the importance of coordinated DER interconnection, installation and management with the rest of the distribution system.

14 SOLAR ENERGY↗
DOE OSTI2018

FY2018 Discovery Through Situational Awareness: Anomalies, Oscillations, and Classification

This technical report is a summary of work performed under the DOE GMLC program under the task titled “Discovery through Situational Awareness”. The focus of this work was to apply big data, statistical techniques to PMU data to do the following: 1) baseline PMU data and find anomalies with a focus on phase angle pairs, 2) identify the frequencies of dominant modal oscillations between different parts of a power system, and 3) identify and classify frequency events in PMU data. This work was performed on PMU data from the Western Interconnect. Many of these algorithms will be applied to wide area phase angle pair data on the Eastern Interconnect in a tool called ESAMS (Eastern Interconnect Situational Awareness Monitoring System). ESAMS was built by EPG (Electric Power Group) and PNNL and is currently being installed on the Eastern Interconnect. The results discussed in this report were made from 14 PMUs from data ranging from Oct 2016 to May 2018. The PMU data were measured 60 times each second. Fourteen phase angle pairs were included, with reactive and active power being calculated at 14 sites. All results have been de-identified for this report. The methodologies applied in the ESAMS tool are described in Section 2. A baseline of PMU data is created using the last 120 days of data. This baseline describes typical behavior as understood from the data. The baseline is then applied to the next day’s data and anomalies from that baseline are identified. Plots are created to help describe what is anomalous during that specific moment in time. These algorithms are currently being installed on the Eastern Interconnect. FY19 work will include discussing the results and findings from the ESAMS tool. A new method of identifying the frequencies of dominant modal oscillations between different parts of a power system is described in Section 3. A list of 199 frequency events provided by Bonneville Power Administration were used to create algorithms to detect frequency events and to create classification rules that can be used to identify future frequency events. Section 4 discusses the results the performance of 7 different event detection, clustering-based methods. Accuracy percentages ranged from 93.59% to 99.16%, with most methods being very fast. GBM (Gradient Boosting Machine) performed well and quickly. These 7 different methods were also tested to determine how well each built classification rules on training data and then applied those rules to identify future frequency events. Accuracy percentages ranged from 89.09% to 99.9%, with all methods performing quickly. GLM (General Linear Model) was viewed as the top performer when considering accuracy and processing time. Because the list of events consisted of 80% of them as active power events and only 20% as faults, all methods struggled to identify events as faults. Small faults were especially difficult to identify. A separate algorithm was created to help with this. Additional work will be done in FY19 on event detection using more data and more events and refinements to the methods.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2022

Big Data Synchrophasor Monitoring and Analytics for Resiliency Tracking (BDSMART)

This report contains key findings from a project titled Big Data Synchrophasor Monitoring and Analytics for Resiliency Tracking (BDSMART), which was carried out through a collaborative effort of a team of researchers from Texas A&M Engineering Experiment Station, Temple University, and Quanta Technology, LLC. The in-kind support came from OSIsoft (acquired by AVEVA), which provided their PI Historian software to demonstrate the use case of streaming PMU data. The first section of the report describes the project goals and objectives related to the development of Machine Learning (ML) models capable of detecting and classifying events by processing phasor measurements captured in the field by Phasor Measurement Units (PMUs). The data for this study was contributed by the utilities/ISOs from the Western and Eastern interconnects and ERCOT, further referred to as Interconnect B (IC B), Interconnect A (IC A), and Interconnect C (IC C), respectively. The approach that the BDSMART Research Team proposed and the key research tasks defined by the team are outlined in this section. The next section describes the technical approach. We first discuss the data constraints related to the PMU measurements and data interpretation constraints imposed by the data contributors. They provided neither the topological information of the grid nor PMU placement locations and captured recorded data at very few locations in the system with the reporting rate of either 30 or 60 fps. The recordings are mostly positive sequence voltage, frequency, and ROCOF, and in some limited cases, three-phase voltages and currents. We then reflect on the bad data issues that stem from poor recording practices and vague definitions of the PMU status bits to supposedly be used for bad data identification. Finally, the data discovery points to imprecise time stamps with incomplete event start/end time, as well as inconsistent and incomplete event labeling, which combined make the implementation of the data models using supervising learning quite challenging. Following the data discovery study, we hypothesize that because the IC B data has the most complete labels, we should focus our model development on that data and then test it on data from other interconnects. We also define the common metrics used to evaluate the results from the ML algorithm tests. We concluded this section by summarizing the common ML models we used and explaining how we implemented and tested them. The issues from this section are expanded in the Training Dataset Report from this project. The final section of this report deals with the accomplishments and conclusions. As the accomplishments, we formulate the problem we are solving and what is achieved by solving the problem. We then reflect on each of the analytics tools we developed and point out the performance of each tool when applied to solving the mentioned problems. We reference this work for further details to the papers we published on each tool. In the conclusions, we give recommendations on how to improve future PMU recording practices to facilitate the ML algorithm implementation and guidance for the future standardization work aimed at clarifying the ambiguities associated with the PMU status bits. We finally list future tasks that can bring about further improvements in the proposed algorithms. The issues from this section are expanded in the Training, and Test Dataset Report filed at the project completion date.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

Innovative, Versatile and Cost-Effective Solid Oxide Fuel Cell Stack Concept

This report summarizes the work performed for the U. S. Department of Energy/National Energy Technology Laboratory (DOE/NETL) under Cooperative Agreement DE-FE0026211 entitled “Innovative, Versatile and Cost-Effective Solid Oxide Fuel Cell Stack Concept”. The overall objective of this project was to develop and evaluate a solid oxide fuel cell (SOFC) stack configuration based on a novel prime-surface interconnect design for a broad range of power generation applications. The prime-surface is a one-piece metallic interconnect that incorporates both fuel and oxidant flow fields with peaks on one side of the interconnect serving as flow channels (valleys) on the other side. (One example is the egg-carton shaped interconnect design.) This stack technology has several attractive features including reduced stack weight and volume, decreased performance losses in stacking, improved sealing, versatility in incorporation of different types of cell construction (thin-film cells or sintered cells) and flexibility in gas flow configuration (cross-flow, co-flow or counter-flow). These features lead to lower cost, better performance, and enhanced reliability for the stack. The R&D work of this project focused on demonstrating fabrication, operability and cost effectiveness of the stack design and was conducted by University of California, San Diego (UCSD) in conjunction with OxEon Energy (OxEon). The key technical activities of this project involved prime-surface interconnect design and fabrication development; supported thin-film cell structure development; stack development; stack operation demonstration; and stack cost assessment. The results of these activities are summarized in this report.

30 DIRECT ENERGY CONVERSION↗
NASA NTRS1990

Solar array arcing in plasmas

Solar cells in space plasma conditions are known to arc into the plasma when the interconnects are at a negative potential of a few hundred volts, relative to plasma potential. For cells with silver-coated interconnects, a threshold voltage for arcing exists at about -230 V, as found in both ground and LEO experiments. The arc rate beyond the threshold voltage depends nearly linearly on plasma density, but has a strong power-law dependence on voltage, such that for small increments in operating voltage there is a large increment in arc rate. The arcs generate broadband radio interference and visible light. In ground tests, interconnects have been damaged by arcs in cells having insufficient isolation from a source of high current. Models for the arcs are highly dependent on the choice of interconnect conductor material exposed to the plasma and possibly on the geometry and choice of adjacent insulator material. Finally, new technology solar cells use copper for the cell interconnects, a material which may have a lower arcing threshold voltage than silver. It is expected, from ground tests of simulated solar cells, that any junction of conductor and insulator exposed to space plasma conditions will arc into the plasma at a few hundred volts negative potential, relative to the local plasma.

Ferguson, Dale C.↗
NASA NTRS1997

Novel Highly Parallel and Systolic Architectures Using Quantum Dot-Based Hardware

VLSI technology has made possible the integration of massive number of components (processors, memory, etc.) into a single chip. In VLSI design, memory and processing power are relatively cheap and the main emphasis of the design is on reducing the overall interconnection complexity since data routing costs dominate the power, time, and area required to implement a computation. Communication is costly because wires occupy the most space on a circuit and it can also degrade clock time. In fact, much of the complexity (and hence the cost) of VLSI design results from minimization of data routing. The main difficulty in VLSI routing is due to the fact that crossing of the lines carrying data, instruction, control, etc. is not possible in a plane. Thus, in order to meet this constraint, the VLSI design aims at keeping the architecture highly regular with local and short interconnection. As a result, while the high level of integration has opened the way for massively parallel computation, practical and full exploitation of such a capability in many applications of interest has been hindered by the constraints on interconnection pattern. More precisely. the use of only localized communication significantly simplifies the design of interconnection architecture but at the expense of somewhat restricted class of applications. For example, there are currently commercially available products integrating; hundreds of simple processor elements within a single chip. However, the lack of adequate interconnection pattern among these processing elements make them inefficient for exploiting a large degree of parallelism in many applications.

Fijany, Amir↗
NASA NTRS1997

Thermophotovoltaic Cell Technology Transferred to the Department of Energy Laboratory and a Commercial Manufacturer

Researchers in the NASA Lewis Research Center's Photovoltaic Branch have developed novel photovoltaic device, called a Monolithically Interconnected Module (MIM), for use in thermophotovoltaic (TPV) power systems. TPV power systems function by heating an emitter to produce light. This light is then converted into electricity by a photovoltaic device or a solar cell. Possible heat sources for the system include concentrated solar energy, the combustion of various fuels, and nuclear decay. NASA has an interest in TPV systems for deep space (nuclear-powered) and near-Sun (solar-powered) missions. There also are many commercial and military applications for TPV, given its potential for high efficiency, low noise, and reliable power. The Monolithically Interconnected Module consists of many small solar cells that are series-interconnected on a common substrate. The cells are fabricated from indium gallium arsenide (InGaAs), which can convert the near-infrared portion of the emitter output spectrum into electricity. The InGaAs devices are deposited on an indium phosphide (InP) substrate that provides electrical isolation. On the bottom the InP substrate is an infrared reflector that returns all the photons that are not converted by the InGaAs device back to the emitter where they are absorbed. This process helps maintain the emitter temperature and dramatically improves the system efficiency. Monolithically Interconnected Module InGaAs thermophotovoltaic cell developed by Lewis. Compared with conventional TPV cells, this TPV device has higher output voltages and lower resistive losses, higher output power density, simplified thermal management, improved reliability, and higher efficiency. The Monolithically Interconnected Module was initially developed under an internally funded effort (Director's Discretionary Fund). Development is now being funded by another government agency, and prototype devices are being produced by a commercial solar cell manufacturer.

Wilt, David M.↗
NASA NTRS2004

Progress in High Power Density SOFC Material Development for Aerospace Applications

Solid oxide fuel cell (SOFC) systems for aircraft applications require order of magnitude increase in specific power density and long life under aircraft operating conditions. Advanced SOFC materials and fabrication processes are being developed at NASA GRC to increase specific power density and durability of SOFC cell and stack. Initial research efforts for increasing specific power density are directed toward increasing the operating temperature for the SOFC system and reducing the weight of the stack. While significant research is underway to develop anode supported SOFC system operating at temperatures in the range of 650 - 850 C for ground power generation applications, such temperatures may not yield the power densities required for aircraft applications. For electrode-supported cells, SOFC stacks with power densities greater than 1.0 W/sq cm are favorable at temperatures in excess of 900 C. The performance of various commercial and developmental anode supported cells is currently being evaluated in the temperature range of 900 to 1000 C to assess the performance gains and materials reliability. The results from these studies will be presented. Since metal interconnects developed for lower temperature operation are not practical at these high temperatures, advanced perovskite based ceramic interconnects with high electronic conductivity and lower sintering temperatures are being developed. Another option for increasing specific power density of SOFC stacks is to decrease the stack weight. Since the interconnect contributes to a significant portion of the stack weight, considerable weight benefits can be derived by decreasing its thickness. Eliminating the gas channels in the interconnect by engineering the pore structure in both anode and cathode can offer significant reduction in thickness of the ceramic interconnect material. New solid oxide fuel cells are being developed with porous engineered electrode supported structures with a 10 - 20 micron thin electrolyte. The performance data for advanced SOFC cells with engineered porosity in both electrodes will be presented.

Cable, Thomas L.↗
NASA NTRS2003

Packaging Technology Designed, Fabricated, and Assembled for High-Temperature SiC Microsystems

A series of ceramic substrates and thick-film metalization-based prototype microsystem packages designed for silicon carbide (SiC) high-temperature microsystems have been developed for operation in 500 C harsh environments. These prototype packages were designed, fabricated, and assembled at the NASA Glenn Research Center. Both the electrical interconnection system and the die-attach scheme for this packaging system have been tested extensively at high temperatures. Printed circuit boards used to interconnect these chip-level packages and passive components also are being fabricated and tested. NASA space and aeronautical missions need harsh-environment, especially high-temperature, operable microsystems for probing the inner solar planets and for in situ monitoring and control of next-generation aeronautical engines. Various SiC high-temperature-operable microelectromechanical system (MEMS) sensors, actuators, and electronics have been demonstrated at temperatures as high as 600 C, but most of these devices were demonstrated only in the laboratory environment partially because systematic packaging technology for supporting these devices at temperatures of 500 C and beyond was not available. Thus, the development of a systematic high-temperature packaging technology is essential for both in situ testing and the commercialization of high-temperature SiC MEMS. Researchers at Glenn developed new prototype packages for high-temperature microsystems using ceramic substrates (aluminum nitride and 96- and 90-wt% aluminum oxides) and gold (Au) thick-film metalization. Packaging components, which include a thick-film metalization-based wirebond interconnection system and a low-electrical-resistance SiC die-attachment scheme, have been tested at temperatures up to 500 C. The interconnection system composed of Au thick-film printed wire and 1-mil Au wire bond was tested in 500 C oxidizing air with and without 50-mA direct current for over 5000 hr. The Au thick-film metalization-based wirebond electrical interconnection system was also tested in an extremely dynamic thermal environment to assess thermal reliability. The I-V curve1 of a SiC high-temperature diode was measured in oxidizing air at 500 C for 1000 hr to electrically test the Au thick-film material-based die-attach assembly.

Chen, Liang-Yu↗
NASA NTRS2007

A Symmetrical, Planar SOFC Design for NASA's High Specific Power Density Requirements

Solid oxide fuel cell (SOFC) systems for aircraft applications require an order of magnitude increase in specific power density (1.0 kW/kg) and long life. While significant research is underway to develop anode supported cells which operate at temperatures in the range of 650-800 C, concerns about Cr-contamination from the metal interconnect may drive the operating temperature down further, to 750 C and lower. Higher temperatures, 900-1000 C, are more favorable for SOFC stacks to achieve specific power densities of 1.0 kW/kg. Since metal interconnects are not practical at these high temperatures and can account for up to 75% of the weight of the stack, NASA is pursuing a design that uses a thin, LaCrO3-based ceramic interconnect that incorporates gas channels into the electrodes. The bi-electrode supported cell (BSC) uses porous YSZ scaffolds, on either side of a 10-20 microns electrolyte. The porous support regions are fabricated with graded porosity using the freeze-tape casting process which can be tailored for fuel and air flow. Removing gas channels from the interconnect simplifies the stack design and allows the ceramic interconnect to be kept thin, on the order of 50 -100 microns. The YSZ electrode scaffolds are infiltrated with active electrode materials following the high temperature sintering step. The NASA-BSC is symmetrical and CTE matched, providing balanced stresses and favorable mechanical properties for vibration and thermal cycling.

Cable, Thomas L.↗
NASA NTRS1992

Robot Serviced Space Facility

A robot serviced space facility includes multiple modules which are identical in physical structure, but selectively differing in function. and purpose. Each module includes multiple like attachment points which are identically placed on each module so as to permit interconnection with immediately adjacent modules. Connection is made through like outwardly extending flange assemblies having identical male and female configurations for interconnecting to and locking to a complementary side of another flange. Multiple rows of interconnected modules permit force, fluid, data and power transfer to be accomplished by redundant circuit paths. Redundant modules of critical subsystems are included. Redundancy of modules and of interconnections results in a space complex with any module being removable upon demand, either for module replacement or facility reconfiguration. without eliminating any vital functions of the complex. Module replacement and facility assembly or reconfiguration are accomplished by a computer controlled articulated walker type robotic manipulator arm assembly having two identical end-effectors in the form of male configurations which are identical to those on module flanges and which interconnect to female configurations on other flanges. The robotic arm assembly moves along a connected set or modules by successively disconnecting, moving and reconnecting alternate ends of itself to a succession of flanges in a walking type maneuver. To transport a module, the robot keeps the transported module attached to one of its end-effectors and uses another flange male configuration of the attached module as a substitute end-effector during walking.

Purves, Lloyd R.↗
DOE OSTI2021

Instructions for performing multi-line memory accesses

A system is described that performs memory access operations. The system includes a processor in a first node, a memory in a second node, a communication interconnect coupled to the processor and the memory, and an interconnect controller in the first node coupled between the processor and the communication interconnect. Upon executing a multi-line memory access instruction, the processor prepares a memory access operation for accessing, in the memory, a block of data including at least some of each of at least two lines of data. The processor then causes the interconnect controller to use a single remote direct memory access memory transfer to perform the memory access operation for the block of data via the communication interconnect.

Roberts, David A.↗
DOE OSTI2016

Capacity value assessments of wind power

This article describes some of the recent research into the capacity value of wind power. With the worldwide increase in wind power during the past several years, there is increasing interest and significance regarding its capacity value because this has a direct influence on the amount of other (nonwind) capacity that is needed. We build on previous reviews from IEEE and IEA W ind T ask 25 a and examine recent work that evaluates the impact of multiple‐year data sets and the impact of interconnected systems on resource adequacy. We also provide examples that explore the use of alternative reliability metrics for wind capacity value calculations. We show how multiple‐year data sets significantly increase the robustness of results compared to single‐year assessments. Assumptions regarding the transmission interconnections play a significant role. To date, results regarding which reliability metric to use for probabilistic capacity valuation show little sensitivity to the metric. WIREs Energy Environ 2017, 6:e226. doi: 10.1002/wene.226 This article is categorized under: Wind Power > Economics and Policy Wind Power > Systems and Infrastructure Energy Infrastructure > Economics and Policy

Milligan, Michael↗
DOE OSTI2020

A Station-hybrid HVDC System Structure and Control Strategies for Cross-seam Power Transmission

Cross-seam high voltage DC (HVDC) transmission in the United States can provide an efficient and flexible bulk power delivery highway between the interconnections. Recent studies have shown significant economic benefits of cross-seam HVDC transmission systems by taking the advantages of load diversity, frequency response, renewable energy diversity, and energy arbitrage in the different interconnections. This paper introduces a station-hybrid system scheme for cross-seam interconnections. Here, the proposed station-hybrid system combines the advantages of line commutated converter (LCC) and voltage source converter (VSC) technologies, thus enabling reliable and flexible bidirectional power flows across the interconnections. Control strategies of the station-hybrid system are investigated with a focus on power flow reversal control under normal operation conditions and grid support control under emergency conditions. The feasibility and effectiveness of the proposed station-hybrid system and corresponding control strategies are verified by simulations in PSCAD/EMTDC.

Station-hybrid system↗
DOE OSTI2017

Wind Generation Participation in Power System Frequency Response: Preprint

The electrical frequency of an interconnected power system must be maintained close its nominal level at all times. Excessive under- and overfrequency excursions can lead to load shedding, instability, machine damage, and even blackouts. There is a rising concern in the electric power industry in recent years about the declining amount of inertia and primary frequency response (PFR) in many interconnections. This decline may continue due to increasing penetrations of inverter-coupled generation and the planned retirements of conventional thermal plants. Inverter-coupled variable wind generation is capable of contributing to PFR and inertia with a response that is different from that of conventional generation. It is not yet entirely understood how such a response will affect the system at different wind power penetration levels. The modeling work presented in this paper evaluates the impact of wind generation's provision of these active power control strategies on a large, synchronous interconnection. All simulations were conducted on the U.S. Western Interconnection with different levels of instantaneous wind power penetrations (up to 80%). The ability of wind power plants to provide PFR - and a combination of synthetic inertial response and PFR - significantly improved the frequency response performance of the system.

17 WIND ENERGY↗
DOE OSTI2021

Persistence of crystal orientations across sub-micron-scale "super-grains" in self-organized Cu-W nanocomposites

We use precession electron diffraction to investigate the crystallographic character of copper (Cu)-tungsten (W) nanocomposites fabricated via physical vapor co-deposition at 400 oC. We observe sub-micron-scale regions, where apparently disconnected Cu and W grains have near identical crystallographic orientations. This persistence of grain orientations suggests Cu and W grains within these regions are interconnected in 3-D when they first form and may be considered as intercalated, sub-micron-scale "super-grains". Indeed, atom probe tomography provides direct evidence of 3-D interconnectivity of W domains. Our findings shed light on the structure and self organization mechanisms of nanocomposites formed by spontaneous phase separation of co-deposited metals.

Yadav, Digvijay↗
DOE OSTI2018

A generative graph model for electrical infrastructure networks

We propose a generative graph model for electrical infrastructure networks that accounts for heterogeneity in both node and edge type. To inform the design of this model, we analyze the properties of power grid graphs derived from the U.S. Eastern Interconnection, Texas Interconnection, and Poland transmission system power grids. Across these datasets, we find subgraphs induced by nodes of the same voltage level exhibit shared structural properties atypical to small-world networks, including low local clustering, large diameter and large average distance. On the other hand, we find subgraphs induced by transformer edges linking nodes of different voltage types contain a more limited structure, consisting mainly of small, disjoint star graphs. The goal of our proposed model is to match both these inter and intra-network properties by proceeding in two phases: we first generate subgraphs for each voltage level and then generate transformer edges that connect these subgraphs. The first phase of the model adapts the Chung-Lu random graph model, taking desired vertex degrees and desired diameter as inputs, while the second phase of the model is based on a simpler random star graph generation process. We test the performance of our model by comparing its output across many runs to the aforementioned real data. In nearly all categories tested, we find our model is more accurate in reproducing the unusual mixture of properties apparent in the data than the Chung-Lu model. We also include graph visualization comparisons, as well as a brief analysis of edge-deletion resiliency. Ultimately, our model may be used to generate synthetic graph data, test hypotheses and algorithms at different scales, and serve as a baseline model on top of which further electrical network properties, device models, and interdependencies to other networks, may be appended.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2017

Scaling deep learning workloads: NVIDIA DGX-1/Pascal and Intel Knights Landing

Deep Learning (DL) algorithms have become ubiquitous in data analytics. As a result, major computing vendors --- including NVIDIA, Intel, AMD, and IBM --- have architectural road-maps influenced by DL workloads. Furthermore, several vendors have recently advertised new computing products as accelerating large DL workloads. Unfortunately, it is difficult for data scientists to quantify the potential of these different products. This paper provides a performance and power analysis of important DL workloads on two major parallel architectures: NVIDIA DGX-1 (eight Pascal P100 GPUs interconnected with NVLink) and Intel Knights Landing (KNL) CPUs interconnected with Intel Omni-Path or Cray Aries. Our evaluation consists of a cross section of convolutional neural net workloads: CifarNet, AlexNet, GoogLeNet, and ResNet50 topologies using the Cifar10 and ImageNet datasets. The workloads are vendor-optimized for each architecture. Our analysis indicates that although GPUs provide the highest overall performance, the gap can close for some convolutional networks; and the KNL can be competitive in performance/watt. We find that NVLink facilitates scaling efficiency on GPUs. However, its importance is heavily dependent on neural network architecture. Furthermore, for weak-scaling --- sometimes encouraged by restricted GPU memory --- NVLink is less important.

Gawande, Nitin A.↗