Search NASA⌕ Search

SEARCH · Search NASA

Results for “graphics processing unit”

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 595 records · Page 33

Comparing the Performance of Julia on CPUs versus GPUs and Julia-MPI versus Fortran-MPI: a case study with MPAS-Ocean (Version 7.1)

Abstract. Some programming languages are easy to develop at the cost of slow execution, while others are fast at runtime but much more difficult to write. Julia is a programming language that aims to be the best of both worlds – a development and production language at the same time. To test Julia's utility in scientific high-performance computing (HPC), we built an unstructured-mesh shallow water model in Julia and compared it against an established Fortran-MPI ocean model, the Model for Prediction Across Scales–Ocean (MPAS-Ocean), as well as a Python shallow water code. Three versions of the Julia shallow water code were created: for single-core CPU, graphics processing unit (GPU), and Message Passing Interface (MPI) CPU clusters. Comparing identical simulations revealed that our first version of the Julia model was 13 times faster than Python using NumPy, where both used an unthreaded single-core CPU. Further Julia optimizations, including static typing and removing implicit memory allocations, provided an additional 10–20× speed-up of the single-core CPU Julia model. The GPU-accelerated Julia code was almost identical in terms of performance to the MPI parallelized code on 64 processes, an unexpected result for such different architectures. Parallelized Julia-MPI performance was identical to Fortran-MPI MPAS-Ocean for low processor counts and ranges from 2× faster to 2× slower for higher processor counts. Our experience is that Julia development is fast and convenient for prototyping but that Julia requires further investment and expertise to be competitive with compiled codes. We provide advice on Julia code optimization for HPC systems.

54 ENVIRONMENTAL SCIENCES↗

SERGHEI (SERGHEI-SWE) v1.0: a performance-portable high-performance parallel-computing shallow-water solver for hydrology and environmental hydraulics

The Simulation EnviRonment for Geomorphology, Hydrodynamics, and Ecohydrology in Integrated form (SERGHEI) is a multi-dimensional, multi-domain, and multi-physics model framework for environmental and landscape simulation, designed with an outlook towards Earth system modelling. At the core of SERGHEI's innovation is its performance-portable high-performance parallel-computing (HPC) implementation, built from scratch on the Kokkos portability layer, allowing SERGHEI to be deployed, in a performance-portable fashion, in graphics processing unit (GPU)-based heterogeneous systems. In this work, we explore combinations of MPI and Kokkos using OpenMP and CUDA backends. In this contribution, we introduce the SERGHEI model framework and present with detail its first operational module for solving shallow-water equations (SERGHEI-SWE) and its HPC implementation. This module is designed to be applicable to hydrological and environmental problems including flooding and runoff generation, with an outlook towards Earth system modelling. Its applicability is demonstrated by testing several well-known benchmarks and large-scale problems, for which SERGHEI-SWE achieves excellent results for the different types of shallow-water problems. Finally, SERGHEI-SWE scalability and performance portability is demonstrated and evaluated on several TOP500 HPC systems, with very good scaling in the range of over 20 000 CPUs and up to 256 state-of-the art GPUs.

58 GEOSCIENCES↗

The ocean model for E3SM global applications: Omega version 0.1.0 – a new high-performance computing code for exascale architectures

This paper introduces Omega, the Ocean Model for E3SM Global Applications. Omega is a new ocean model designed to run efficiently on high performance computing (HPC) platforms, including exascale heterogeneous architectures with accelerators, such as Graphics Processing Units (GPUs). Omega is written in C and uses the Kokkos performance portability library. These were chosen because they are well-supported and will help future-proof Omega for upcoming HPC architectures. Omega will eventually replace the Model for Prediction Across Scales-Ocean (MPAS-Ocean) in the US Department of Energy's (DOE's) Energy Exascale Earth System Model (E3SM). Omega runs on unstructured horizontal meshes with variable-resolution capability and implements the same horizontal discretization as MPAS-Ocean. This work documents the design and performance of Omega Version 0.1.0 (Omega-V0), which solves the shallow water equations with passive tracers and is the first step towards the full primitive equation ocean model. On Central Processing Units (CPUs), Omega-V0 is 1.4 times faster than MPAS-Ocean with the same configuration. Omega-V0 is more efficient on GPUs than CPUs on a per-watt basis – by a factor of 5.3 on Frontier and 3.6 on Aurora, two of the world's fastest exascale computers.

54 ENVIRONMENTAL SCIENCES↗

Evolution of the Antarctic Ice Sheet from 2000–2300 and beyond: model sensitivity and uncertainty analysis using MPAS-Albany Land Ice

We present a description of the Antarctic Ice Sheet model configuration submitted to the ISMIP6-Antarctica-2300 experiment using the MPAS-Albany Land Ice model, along with three new sets of simulations: (1) a set of extended simulations to 2500 for three forced experiments and to 2775 for the control experiment; (2) a sensitivity analysis of our model configuration to parameters controlling basal sliding and sub-shelf melt, and to model structural choices including the choice of the energy and stress balances; and (3) a 72-member ensemble run on graphics processing units (GPUs) and analysis of variance to determine the primary sources of uncertainty in our ice-sheet model projections. Our extended simulations predict rapid retreat beginning after 2300 for SSP1-2.6 forcing and after 2500 for present-day (control) forcing, primarily in the Amundsen Sea Embayment. We find that varying the sub-shelf melt parameter between the 5th to 95th percentile values for a mean-Antarctic calibration target results in an up to ∼ ± 40 % change in sea-level contribution relative to our baseline simulations that used the median value. Using a linear basal sliding law reduces sea-level contribution by 51 %–73 % relative to our baseline nonlinear sliding law with an exponent of 1/5. When using basal sliding law exponents of 1/3 and 1/10, the overall difference from our baseline simulations at 2300 is on the order of 10 %. The Amundsen Sea Embayment region displays a strongly non-linear dependence of mass loss on the sliding law exponent, with no discernible relationship between the sliding law exponent and the mass loss by 2300, while the sectors feeding the Ross and Filchner-Ronne ice shelves exhibit more mass loss with a more-plastic sliding law. Our model fidelity sensitivity experiments reveal a 9 %–31 % increase in sea-level contribution when using a depth-integrated stress balance approximation relative to our three-dimensional solver, while using a fixed-in-time temperature field increases sea-level contribution by 14 %–88 % relative to two thermomechanically coupled configurations. Our 72-member ensemble and analysis of variance show that the uncertainty in long-term projections is dominated by the choice of Earth system model forcing and the presence or absence of hydrofracture forcing, rather than uncertainty in sliding and sub-shelf melt parameters.

58 GEOSCIENCES↗

Optimal utilization of heterogeneous resources for biomolecular simulations.

Biomolecular simulations have traditionally benefited from increases in the processor clock speed and coarse-grain inter-node parallelism on large-scale clusters. With stagnating clock frequencies, the evolutionary path for performance of microprocessors is maintained by virtue of core multiplication. Graphical processing units (GPUs) offer revolutionary performance potential at the cost of increased programming complexity. Furthermore, it has been extremely challenging to effectively utilize heterogeneous resources (host processor and GPU cores) for scientific simulations, as underlying systems, programming models and tools are continually evolving. In this paper, we present a parametric study demonstrating approaches to exploit resources of heterogeneous systems to reduce time-to-solution of a production-level application for biological simulations. By overlapping and pipelining computation and communication, we observe up to 10-fold application acceleration in multi-core and multi-GPU environments illustrating significant performance improvements over code acceleration approaches, where the host-to-accelerator ratio is static, and is constrained by a given algorithmic implementation.

99 GENERAL AND MISCELLANEOUS↗

Failing in place for low-serviceability storage infrastructure using high-parity GPU-based RAID.

In order to provide large quantities of high-reliability disk-based storage, it has become necessary to aggregate disks into fault-tolerant groups based on the RAID methodology. Most RAID levels do provide some fault tolerance, but there are certain classes of applications that require increased levels of fault tolerance within an array. Some of these applications include embedded systems in harsh environments that have a low level of serviceability, or uninhabited data centers servicing cloud computing. When describing RAID reliability, the Mean Time To Data Loss (MTTDL) calculations will often assume that the time to replace a failed disk is relatively low, or even negligible compared to rebuild time. For platforms that are in remote areas collecting and processing data, it may be impossible to access the system to perform system maintenance for long periods. A disk may fail early in a platform's life, but not be replaceable for much longer than typical for RAID arrays. Service periods may be scheduled at intervals on the order of months, or the platform may not be serviced until the end of a mission in progress. Further, this platform may be subject to extreme conditions that can accelerate wear and tear on a disk, requiring even more protection from failures. We have created a high parity RAID implementation that uses a Graphics Processing Unit (GPU) to compute more than two blocks of parity information per stripe, allowing extra parity to eliminate or reduce the requirement for rebuilding data between service periods. While this type of controller is highly effective for RAID 6 systems, an important benefit is the ability to incorporate more parity into a RAID storage system. Such RAID levels, as yet unnamed, can tolerate the failure of three or more disks (depending on configuration) without data loss. While this RAID system certainly has applications in embedded systems running applications in the field, similar benefits can be obtained for servers that are engineered for storage density, with less regard for serviceability or maintainability. A storage brick can be designed to have a MTTDL that extends well beyond the useful lifetime of the hardware used, allowing the disk subsystem to require less service throughout the lifetime of a compute resource. This approach is similar to the Xiotech ISE. Such a design can be deliberately placed remotely (without frequent support) in order to provide colocation, or meet cost goals. For workloads where reliability is key, but conditions are sub-optimal for routine serviceability, a high-parity RAID can provide extra reliability in extraordinary situations. For example, for installations requiring very high Mean Time To Repair, the extra parity can eliminate certain problems with maintaining hot spares, increasing overall reliability. Furthermore, in situations where disk reliability is reduced because of harsh conditions, extra parity can guard against early data loss due to lowered Mean Time To Failure. If used through an iSCSI interface with a streaming workload, it is possible to gain all of these benefits without impacting performance.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

LBMD : a layer-based mesh data structure tailored for generic API infrastructures.

A new mesh data structure is introduced for the purpose of mesh processing in Application Programming Interface (API) infrastructures. This data structure utilizes a reduced mesh representation to increase its ability to handle significantly larger meshes compared to full mesh representation. In spite of the reduced representation, each mesh entity (vertex, edge, face, and region) is represented using a unique handle, with no extra storage cost, which is a crucial requirement in most API libraries. The concept of mesh layers makes the data structure more flexible for mesh generation and mesh modification operations. This flexibility can have a favorable impact in solver based queries of finite volume and multigrid methods. The capabilities of LBMD make it even more attractive for parallel implementations using Message Passing Interface (MPI) or Graphics Processing Units (GPUs). The data structure is associated with a new classification method to relate mesh entities to their corresponding geometrical entities. The classification technique stores the related information at the node level without introducing any ambiguities. Several examples are presented to illustrate the strength of this new data structure.

99 GENERAL AND MISCELLANEOUS↗

GPU-based Scalable Volumetric Reconstruction for Multi-view Stereo

We present a new scalable volumetric reconstruction algorithm for multi-view stereo using a graphics processing unit (GPU). It is an effectively parallelized GPU algorithm that simultaneously uses a large number of GPU threads, each of which performs voxel carving, in order to integrate depth maps with images from multiple views. Each depth map, triangulated from pair-wise semi-dense correspondences, represents a view-dependent surface of the scene. This algorithm also provides scalability for large-scale scene reconstruction in a high resolution voxel grid by utilizing streaming and parallel computation. The output is a photo-realistic 3D scene model in a volumetric or point-based representation. We demonstrate the effectiveness and the speed of our algorithm with a synthetic scene and real urban/outdoor scenes. Our method can also be integrated with existing multi-view stereo algorithms such as PMVS2 to fill holes or gaps in textureless regions.

97 MATHEMATICS AND COMPUTING↗

Interactive, graphical processing unitbased evaluation of evacuation scenarios at the state scale

In large-scale scenarios, transportation modeling and simulation is severely constrained by simulation time. For example, few real- time simulators scale to evacuation traffic scenarios at the level of an entire state, such as Louisiana (approximately 1 million links) or Florida (2.5 million links). New simulation approaches are needed to overcome severe computational demands of conventional (microscopic or mesoscopic) modeling techniques. Here, a new modeling and execution methodology is explored that holds the potential to provide a tradeoff among the level of behavioral detail, the scale of transportation network, and real-time execution capabilities. A novel, field-based modeling technique and its implementation on graphical processing units are presented. Although additional research with input from domain experts is needed for refining and validating the models, the techniques reported here afford interactive experience at very large scales of multi-million road segments. Illustrative experiments on a few state-scale net- works are described based on an implementation of this approach in a software system called GARFIELD. Current modeling cap- abilities and implementation limitations are described, along with possible use cases and future research.

99 GENERAL AND MISCELLANEOUS↗

High-Performance Beam Simulator for the LANSCE Linac

A high performance multiparticle tracking simulator is currently under development at Los Alamos. The heart of the simulator is based upon the beam dynamics simulation algorithms of the PARMILA code, but implemented in C++ on Graphics Processing Unit (GPU) hardware using NVIDIA's CUDA platform. Linac operating set points are provided to the simulator via the EPICS control system so that changes of the real time linac parameters are tracked and the simulation results updated automatically. This simulator will provide valuable insight into the beam dynamics along a linac in pseudo real-time, especially where direct measurements of the beam properties do not exist. Details regarding the approach, benefits and performance are presented.

43 PARTICLE ACCELERATORS↗

Singular value decomposition utilizing parallel algorithms on graphical processors

One of the current challenges in underwater acoustic array signal processing is the detection of quiet targets in the presence of noise. In order to enable robust detection, one of the key processing steps requires data and replica whitening. This, in turn, involves the eigen-decomposition of the sample spectral matrix, Cx = 1/K xKX(k)XH(k) where X(k) denotes a single frequency snapshot with an element for each element of the array. By employing the singular value decomposition (SVD) method, the eigenvectors and eigenvalues can be determined directly from the data without computing the sample covariance matrix, reducing the computational requirements for a given level of accuracy (van Trees, Optimum Array Processing). (Recall that the SVD of a complex matrix A involves determining V, , and U such that A = U VH where U and V are orthonormal and is a positive, real, diagonal matrix containing the singular values of A. U and V are the eigenvectors of AAH and AHA, respectively, while the singular values are the square roots of the eigenvalues of AAH.) Because it is desirable to be able to compute these quantities in real time, an efficient technique for computing the SVD is vital. In addition, emerging multicore processors like graphical processing units (GPUs) are bringing parallel processing capabilities to an ever increasing number of users. Since the computational tasks involved in array signal processing are well suited for parallelization, it is expected that these computations will be implemented using GPUs as soon as users have the necessary computational tools available to them. Thus, it is important to have an SVD algorithm that is suitable for these processors. This work explores the effectiveness of two different parallel SVD implementations on an NVIDIA Tesla C2050 GPU (14 multiprocessors, 32 cores per multiprocessor, 1.15 GHz clock - peed). The first algorithm is based on a two-step algorithm which bidiagonalizes the matrix using Householder transformations, and then diagonalizes the intermediate bidiagonal matrix through implicit QR shifts. This is similar to that implemented for real matrices by Lahabar and Narayanan ("Singular Value Decomposition on GPU using CUDA", IEEE International Parallel Distributed Processing Symposium 2009). The implementation is done in a hybrid manner, with the bidiagonalization stage done using the GPU while the diagonalization stage is done using the CPU, with the GPU used to update the U and V matrices. The second algorithm is based on a one-sided Jacobi scheme utilizing a sequence of pair-wise column orthogonalizations such that A is replaced by AV until the resulting matrix is sufficiently orthogonal (that is, equal to U ). V is obtained from the sequence of orthogonalizations, while can be found from the square root of the diagonal elements of AH A and, once is known, U can be found from column scaling the resulting matrix. These implementations utilize CUDA Fortran and NVIDIA's CUB LAS library. The primary goal of this study is to quantify the comparative performance of these two techniques against themselves and other standard implementations (for example, MATLAB). Considering that there is significant overhead associated with transferring data to the GPU and with synchronization between the GPU and the host CPU, it is also important to understand when it is worthwhile to use the GPU in terms of the matrix size and number of concurrent SVDs to be calculated.

99 GENERAL AND MISCELLANEOUS↗

Detection of Coherent Structures in Extreme-Scale Simulations

The analysis of coherent structures is a common problem in many scientific domains ranging from astrophysics to combustion, fusion, and materials science. The data from three-dimensional simulations are analyzed to detect the structures, extract statistics on them, and track them over time to gain insights into the phenomenon being modeled. This analysis is typically done off-line, using data that have been written out by the simulations. However, the move towards extreme scale architectures, with multi-core processors and graphical processing units, will affect how such analysis is done as it is unlikely that the systems will support the I/O bandwidth required for off-line analysis. Moving the analysis in-situ is a solution only if we know a priori what analysis will be done, as well as the algorithms used and their parameter settings. Even then, we need to ensure that this will not substantially increase the memory requirements or the data movement as the former will be limited and the latter will be expensive. In the Exa-DM project, a collaboration between Lawrence Livermore National Laboratory and University of Minnesota, we are exploring ways in which we can address the conflicting demands of coherent structure analysis of simulation data and the architecture of modern parallel systems, while enabling scientific discovery at the exascale. In this paper, we describe our work in two areas: the in situ implementation of an existing algorithm for coherent structure analysis and the use of graph-based techniques to efficiently compress the data.

79 ASTRONOMY AND ASTROPHYSICS↗

The Tradeoffs of Fused Memory Hierarchies in Heterogeneous Computing Architectures

With the rise of general purpose computing on graphics processing units (GPGPU), the influence from consumer markets can now be seen across the spectrum of computer architectures. In fact, many of the high-ranking Top500 HPC systems now include these accelerators. Traditionally, GPUs have connected to the CPU via the PCIe bus, which has proved to be a significant bottleneck for scalable scientific applications. Now, a trend toward tighter integration between CPU and GPU has removed this bottleneck and unified the memory hierarchy for both CPU and GPU cores. We examine the impact of this trend for high performance scientific computing by investigating AMD's new Fusion Accelerated Processing Unit (APU) as a testbed. In particular, we evaluate the tradeoffs in performance, power consumption, and programmability when comparing this unified memory hierarchy with similar, but discrete GPUs.

43 PARTICLE ACCELERATORS↗

Efficient GPU implementation for Particle in Cell algorithm

Particle in cell method is widely used method in the plasma physics to study the trajectories of charged particles under electromagnetic fields. The PIC algorithm is computationally intensive and its time requirements are proportional to the number of charged particles involved in the simulation. The focus of the paper is to parallelize the PIC algorithm on Graphics Processing Unit (GPU). We present several performance tradeoffs related to the small shared memory and atomic operations on the GPU to achieve high performance.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Hardware Architectures for Data-Intensive Computing Problems: A Case Study for String Matching

DNA analysis is an emerging application of high performance bioinformatic. Modern sequencing machinery are able to provide, in few hours, large input streams of data, which needs to be matched against exponentially growing databases of known fragments. The ability to recognize these patterns effectively and fastly may allow extending the scale and the reach of the investigations performed by biology scientists. Aho-Corasick is an exact, multiple pattern matching algorithm often at the base of this application. High performance systems are a promising platform to accelerate this algorithm, which is computationally intensive but also inherently parallel. Nowadays, high performance systems also include heterogeneous processing elements, such as Graphic Processing Units (GPUs), to further accelerate parallel algorithms. Unfortunately, the Aho-Corasick algorithm exhibits large performance variability, depending on the size of the input streams, on the number of patterns to search and on the number of matches, and poses significant challenges on current high performance software and hardware implementations. An adequate mapping of the algorithm on the target architecture, coping with the limit of the underlining hardware, is required to reach the desired high throughputs. In this paper, we discuss the implementation of the Aho-Corasick algorithm for GPU-accelerated high performance systems. We present an optimized implementation of Aho-Corasick for GPUs and discuss its tradeoffs on the Tesla T10 and he new Tesla T20 (codename Fermi) GPUs. We then integrate the optimized GPU code, respectively, in a MPI-based and in a pthreads-based load balancer to enable execution of the algorithm on clusters and large sharedmemory multiprocessors (SMPs) accelerated with multiple GPUs.

Tumeo, Antonino↗

Efficient Aho-Corasick String Matching on Emerging Multicore Architectures

String matching algorithms are critical to several scientific fields. Beside text processing and databases, emerging applications such as DNA protein sequence analysis, data mining, information security software, antivirus, ma- chine learning, all exploit string matching algorithms [3]. All these applica- tions usually process large quantity of textual data, require high performance and/or predictable execution times. Among all the string matching algorithms, one of the most studied, especially for text processing and security applica- tions, is the Aho-Corasick algorithm. 1 2 Book title goes here Aho-Corasick is an exact, multi-pattern string matching algorithm which performs the search in a time linearly proportional to the length of the input text independently from pattern set size. However, depending on the imple- mentation, when the number of patterns increase, the memory occupation may raise drastically. In turn, this can lead to significant variability in the performance, due to the memory access times and the caching effects. This is a significant concern for many mission critical applications and modern high performance architectures. For example, security applications such as Network Intrusion Detection Systems (NIDS), must be able to scan network traffic against very large dictionaries in real time. Modern Ethernet links reach up to 10 Gbps, and malicious threats are already well over 1 million, and expo- nentially growing [28]. When performing the search, a NIDS should not slow down the network, or let network packets pass unchecked. Nevertheless, on the current state-of-the-art cache based processors, there may be a large per- formance variability when dealing with big dictionaries and inputs that have different frequencies of matching patterns. In particular, when few patterns are matched and they are all in the cache, the procedure is fast. Instead, when they are not in the cache, often because many patterns are matched and the caches are continuously thrashed, they should be retrieved from the system memory and the procedure is slowed down by the increased latency. Efficient implementations of string matching algorithms have been the fo- cus of several works, targeting Field Programmable Gate Arrays [4, 25, 15, 5], highly multi-threaded solutions like the Cray XMT [34], multicore proces- sors [19] or heterogeneous processors like the Cell Broadband Engine [35, 22]. Recently, several researchers have also started to investigate the use Graphic Processing Units (GPUs) for string matching algorithms in security applica- tions [20, 10, 32, 33]. Most of these approaches mainly focus on reaching high peak performance, or try to optimize the memory occupation, rather than looking at performance stability. However, hardware solutions supports only small dictionary sizes due to lack of memory and are difficult to customize, while platforms such as the Cell/B.E. are very complex to program.

Tumeo, Antonino↗

The Genome Portal of the Department of Energy Joint Genome Institute

The JGI Genome Portal (http://genome.jgi.doe.gov) provides unified access to all JGI genomic databases and analytical tools. A user can search, download and explore multiple data sets available for all DOE JGI sequencing projects including their status, assemblies and annotations of sequenced genomes. Genome Portal in the past 2 years was significantly updated, with a specific emphasis on efficient handling of the rapidly growing amount of diverse genomic data accumulated in JGI. A critical aspect of handling big data in genomics is the development of visualization and analysis tools that allow scientists to derive meaning from what are otherwise terrabases of inert sequence. An interactive visualization tool developed in the group allows us to explore contigs resulting from a single metagenome assembly. Implemented with modern web technologies that take advantage of the power of the computer's graphical processing unit (gpu), the tool allows the user to easily navigate over a 100,000 data points in multiple dimensions, among many biologically meaningful parameters of a dataset such as relative abundance, contig length, and G+C content.

59 BASIC BIOLOGICAL SCIENCES↗

A Study of Complex Deep Learning Networks on High Performance, Neuromorphic, and Quantum Computers

Current Deep Learning models use highly optimized convolutional neural networks (CNN) trained on large graphical processing units (GPU)-based computers with a fairly simple layered network topology, i.e., highly connected layers, without intra-layer connections. Complex topologies have been proposed, but are intractable to train on current systems. Building the topologies of the deep learning network requires hand tuning, and implementing the network in hardware is expensive in both cost and power. In this paper, we evaluate deep learning models using three different computing architectures to address these problems: quantum computing to train complex topologies, high performance computing (HPC) to automatically determine network topology, and neuromorphic computing for a low-power hardware implementation. Due to input size limitations of current quantum computers we use the MNIST dataset for our evaluation. The results show the possibility of using the three architectures in tandem to explore complex deep learning networks that are untrainable using a von Neumann architecture. We show that a quantum computer can find high quality values of intra-layer connections and weights, while yielding a tractable time result as the complexity of the network increases; a high performance computer can find optimal layer-based topologies; and a neuromorphic computer can represent the complex topology and weights derived from the other architectures in low power memristive hardware. This represents a new capability that is not feasible with current von Neumann architecture. It potentially enables the ability to solve very complicated problems unsolvable with current computing technologies.

Potok, Thomas E↗