Search NASA⌕ Search

SEARCH · Search NASA

Results for “memory efficiency”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,765 records · Page 98

Efficient Support for Matrix Computations on Heterogeneous Multi-core and Multi-GPU Architectures

We present a new methodology for utilizing all CPU cores and all GPUs on a heterogeneous multicore and multi-GPU system to support matrix computations e ciently. Our approach is able to achieve the objectives of a high degree of parallelism, minimized synchronization, minimized communication, and load balancing. Our main idea is to treat the heterogeneous system as a distributed-memory machine, and to use a heterogeneous 1-D block cyclic distribution to allocate data to the host system and GPUs to minimize communication. We have designed heterogeneous algorithms with two di erent tile sizes (one for CPU cores and the other for GPUs) to cope with processor heterogeneity. We propose an auto-tuning method to determine the best tile sizes to attain both high performance and load balancing. We have also implemented a new runtime system and applied it to the Cholesky and QR factorizations. Our experiments on a compute node with two Intel Westmere hexa-core CPUs and three Nvidia Fermi GPUs demonstrate good weak scalability, strong scalability, load balance, and e ciency of our approach.

97 MATHEMATICS AND COMPUTING↗

Sylvester-preconditioned adaptive-rank implicit time integrators for advection-diffusion equations with variable coefficients

Here, we consider the adaptive-rank integration of multi-dimensional time-dependent advection-diffusion partial differential equations (PDEs) with variable coefficients. We employ a standard finite-difference method for spatial discretization coupled with high-order diagonally implicit Runge-Kutta temporal schemes. The discrete equation is a generalized Sylvester equation (GSE), which we solve with a projection-based adaptive-rank algorithm structured around two key strategies: (i) constructing dimension-wise subspaces using a novel atypical extended Krylov strategy, and (ii) efficiently solving the basis coefficient matrix with a preconditioned GMRES solver. The low-rank decomposition is performed in 2D using SVD and with high-order SVD (HOSVD) in 3D to represent the tensor in a compressed Tucker format. For d-dimensional problems (here, d = 2 or 3), the computational complexity and memory storage of the approach are found numerically to scale as and $\mathscr{O}(Nr^2) + \mathscr{O} (r^{d+1})$ and $\mathscr{O}(Nr) + \mathscr{O} (r^{d})$, respectively, with the one-dimensional resolution and the maximal rank during the Krylov iteration (which we find to be largely independent of on our numerical examples). We present numerical examples that illustrate the advertised properties of the algorithm.

97 MATHEMATICS AND COMPUTING↗

Precisely computing phonons via irreducible derivatives

Computing phonons from first principles is typically considered a solved problem, yet inadequacies in existing techniques continue to yield deficient results in systems with sensitive phonons. Here, in this study, we circumvent this issue using the lone irreducible derivative (LID) and bundled irreducible derivative (BID) approaches to computing phonons via finite displacements, where the former optimizes precision via energy derivatives and the latter provides the most efficient algorithm using force derivatives. A condition number optimized basis for BID is derived which guarantees the minimum amplification of error. Additionally, a hybrid LID-BID approach is formulated, in which select irreducible derivatives computed using LID replace BID results. We illustrate our approach on two prototypical systems with sensitive phonons: the shape memory alloy AuZn and metallic lithium. Comparing our resulting phonons in the aforementioned crystals to calculations in the literature reveals nontrivial inaccuracies. Our approaches can be fully automated, making them well suited for both niche systems of interest and high-throughput approaches.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗

A Length Adaptive Algorithm-Hardware Co-design of Transformer on FPGA Through Sparse Attention and Dynamic Pipelining

Transformers are considered one of the most important deep learning models since 2018, in part because it establishes state-of-the-art (SOTA) records and could potentially replace existing Deep Neural Networks (DNNs). Despite the remarkable triumphs, the prolonged turnaround time of Transformer models is a widely recognized roadblock. The variety of sequence lengths imposes additional computing overhead where inputs need to be zero-padded to the maximum sentence length in the batch to accommodate the parallel computing platforms. This paper targets the field-programmable gate array (FPGA) and proposes a coherent sequence length adaptive algorithm–hardware co-design for Transformer acceleration. Particularly, we develop a hardware-friendly sparse attention operator and a length-aware hardware resource scheduling algorithm. The proposed sparse attention operator brings the complexity of attention-based models down to linear complexity and alleviates the off-chip memory traffic. The proposed length-aware resource hardware scheduling algorithm dynamically allocates the hardware resources to fill up the pipeline slots and eliminates bubbles for NLP tasks. Experiments show that our design has very small accuracy loss and has 80.2 × and 2.6 × speedup compared to CPU and GPU implementation, and 4 × higher energy efficiency than state-of-the-art GPU accelerator optimized via CUBLAS GEMM.

Peng, Hongwu↗

Phase I Final Technical Report on Energy-Efficient Reconfigurable Universal Accelerator Interconnect

It is well known that application specific computing systems, optimally designed and configured for a given workload, offer much higher energy-efficiency and throughput than general purpose systems. In modern computing systems, heterogeneous computing systems have emerged that exploit the energy and performance benefits of combining various different domain-specific processor architectures. Application domains such as high-performance computing and machine learning now process terabyte-sized data sets, requiring enormous processing and memory resources. These applications have very high-power consumption due to bottlenecks in the electrical interconnection between processing units. This project aims to reduce both communication energy and latency by integrating universally available accelerators with silicon photonics. It also aims to increase system throughput by exploiting emerging technologies in silicon photonic reconfigurable interconnects; this will allow the system to balance itself in real time to accommodate changes in workloads and data flows.

2.5D/3D integration↗

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↗

A domain wall-magnetic tunnel junction artificial synapse with notched geometry for accurate and efficient training of deep neural networks

Inspired by the parallelism and efficiency of the brain, several candidates for artificial synapse devices have been developed for neuromorphic computing, yet a nonlinear and asymmetric synaptic response curve precludes their use for backpropagation, the foundation of modern supervised learning. Spintronic devices—which benefit from high endurance, low power consumption, low latency, and CMOS compatibility—are a promising technology for memory, and domain-wall magnetic tunnel junction (DW-MTJ) devices have been shown to implement synaptic functions such as long-term potentiation and spike-timing dependent plasticity. In this work, we propose a notched DW-MTJ synapse as a candidate for supervised learning. Using micromagnetic simulations at room temperature, we show that notched synapses ensure the non-volatility of the synaptic weight and allow for highly linear, symmetric, and reproducible weight updates using either spin transfer torque (STT) or spin–orbit torque (SOT) mechanisms of DW propagation. We use lookup tables constructed from micromagnetics simulations to model the training of neural networks built with DW-MTJ synapses on both the MNIST and Fashion-MNIST image classification tasks. Accounting for thermal noise and realistic process variations, the DW-MTJ devices achieve classification accuracy close to ideal floating-point updates using both STT and SOT devices at room temperature and at 400 K. Our work establishes the basis for a magnetic artificial synapse that can eventually lead to hardware neural networks with fully spintronic matrix operations implementing machine learning.

42 ENGINEERING↗

Vaccine-elicited antibody that neutralizes H5N1 influenza and variants binds the receptor site and polymorphic sites

Antigenic drift of circulating seasonal influenza viruses necessitates an international vaccine effort to reduce the impact on human health. A critical feature of the seasonal vaccine is that it stimulates an already primed immune system to diversify memory B cells to recognize closely related, but antigenically distinct, influenza glycoproteins (hemagglutinins). Influenza pandemics arise when hemagglutinins to which no preexisting adaptive immunity exists acquire the capacity to infect humans. Hemagglutinin 5 is one subtype to which little preexisting immunity exists and is only a few acquired mutations away from the ability to transmit efficiently between ferrets, and possibly humans. In this paper, we describe the structure and molecular mechanism of neutralization by H5.3, a vaccine-elicited antibody that neutralizes hemagglutinin 5 viruses and variants with expanded host range. H5.3 binds in the receptor-binding site, forming contacts that recapitulate many of the sialic acid interactions, as well as multiple peripheral interactions, yet is not sensitive to mutations that alter sialic acid binding. H5.3 is highly specific for a subset of H5 strains, and this specificity arises from interactions to the periphery of the receptor-binding site. Finally, H5.3 is also extremely potent, despite retaining germ line-like conformational flexibility.

60 APPLIED LIFE SCIENCES↗

Potential of minicomputer/array-processor system for nonlinear finite-element analysis

The potential of using a minicomputer/array-processor system for the efficient solution of large-scale, nonlinear, finite-element problems is studied. A Prime 750 is used as the host computer, and a software simulator residing on the Prime is employed to assess the performance of the Floating Point Systems AP-120B array processor. Major hardware characteristics of the system such as virtual memory and parallel and pipeline processing are reviewed, and the interplay between various hardware components is examined. Effective use of the minicomputer/array-processor system for nonlinear analysis requires the following: (1) proper selection of the computational procedure and the capability to vectorize the numerical algorithms; (2) reduction of input-output operations; and (3) overlapping host and array-processor operations. A detailed discussion is given of techniques to accomplish each of these tasks. Two benchmark problems with 1715 and 3230 degrees of freedom, respectively, are selected to measure the anticipated gain in speed obtained by using the proposed algorithms on the array processor.

Strohkorb, G. A.↗

Machine vision for real time orbital operations

Machine vision for automation and robotic operation of Space Station era systems has the potential for increasing the efficiency of orbital servicing, repair, assembly and docking tasks. A machine vision research project is described in which a TV camera is used for inputing visual data to a computer so that image processing may be achieved for real time control of these orbital operations. A technique has resulted from this research which reduces computer memory requirements and greatly increases typical computational speed such that it has the potential for development into a real time orbital machine vision system. This technique is called AI BOSS (Analysis of Images by Box Scan and Syntax).

Vinz, Frank L.↗

A parallel algorithm for multi-level logic synthesis using the transduction method

The Transduction Method has been shown to be a powerful tool in the optimization of multilevel networks. Many tools such as the SYLON synthesis system (X90), (CM89), (LM90) have been developed based on this method. A parallel implementation is presented of SYLON-XTRANS (XM89) on an eight processor Encore Multimax shared memory multiprocessor. It minimizes multilevel networks consisting of simple gates through parallel pruning, gate substitution, gate merging, generalized gate substitution, and gate input reduction. This implementation, called Parallel TRANSduction (PTRANS), also uses partitioning to break large circuits up and performs inter- and intra-partition dynamic load balancing. With this, good speedups and high processor efficiencies are achievable without sacrificing the resulting circuit quality.

Lim, Chieng-Fai↗

Neural Network Based Sensory Fusion for Landmark Detection

NASA is planning to send numerous unmanned planetary missions to explore the space. This requires autonomous robotic vehicles which can navigate in an unstructured, unknown, and uncertain environment. Landmark based navigation is a new area of research which differs from the traditional goal-oriented navigation, where a mobile robot starts from an initial point and reaches a destination in accordance with a pre-planned path. The landmark based navigation has the advantage of allowing the robot to find its way without communication with the mission control station and without exact knowledge of its coordinates. Current algorithms based on landmark navigation however pose several constraints. First, they require large memories to store the images. Second, the task of comparing the images using traditional methods is computationally intensive and consequently real-time implementation is difficult. The method proposed here consists of three stages, First stage utilizes a heuristic-based algorithm to identify significant objects. The second stage utilizes a neural network (NN) to efficiently classify images of the identified objects. The third stage combines distance information with the classification results of neural networks for efficient and intelligent navigation.

Kumbla, Kishan -K.↗

xSDK: Building an ecosystem of highly efficient math libraries for exascale

Current efforts to build increasingly powerful computer architectures are opening up new avenues for more complex and higher fidelity simulations coupled with data analytics and learning, leading to new scientific insights and deeper understanding. At one extreme, exascale computers will be much faster than previous computer generations (performing 10 18 operations per second—that is, 1,000 times faster than petascale). To achieve these performance improvements, computer architectures are becoming increasingly complex, with deep memory hierarchies, very high node and core counts, and heterogeneous features such as graphics processing units (GPUs). Such architectural changes impact the full breadth of computing scales, as heterogeneity pervades even current-generation laptops, workstations, and moderate-sized clusters. While emerging advanced architectures provide unprecedented opportunities, they also present significant challenges for developers of scientific applications, such as multiphysics and multiscale codes, who must adapt their software to handle disruptive changes in architectures and new programming models that have not yet stabilized. Developers must consider increasing concurrency while reducing communication and synchronization, and other complexities such as the potential for using mixed precision to leverage the compute power available in low-precision tensor cores. On one hand, developers must implement new scientific capabilities, which in turn increase code complexity. On the other hand, the codes must be ported to new architectures, requiring the inclusion of new programming models and the restructuring of code to achieve good performance. Addressing these issues is beyond the capability of any single person or team—leading to the need for collaboration among many teams, who encapsulate their expertise in reusable software and work together to create sustainable software ecosystems.

97 MATHEMATICS AND COMPUTING↗

Lithium indium diselenide: A new scintillator for neutron imaging

Lithium indium diselenide, 6 LiInSe 2 or LISe, is a newly developed neutron detection material that shows both semiconducting and scintillating properties. The 24% atomic density of 6 Li yields a thermal neutron mean free path of only 920 μm. This paper reports on the performance of LISe crystals in scintillation mode for its potential use as a converter screen for thermal/cold neutron imaging. The spatial resolution of LISe, determined using a 10% value of the Modulation Transfer Function (MTF), was found to not scale linearly with thickness. Crystals having a thickness of 450 μm or larger resulted in an average spatial resolution of 67 μm, and the thinner crystals exhibited an increase in spatial resolution down to the Nyquist frequency of the CCD. The highest measured spatial resolution of 198 μm thick LISe (27 μm) outperforms a commercial 50 μm thick ZnS(Cu): 6 LiF scintillation screen (100 μm) by more than a factor of three. For the thicknesses considered in this study, it has been found that the light yield of LISe did not scale with its thickness, suggesting the need for optimizing the synthesis to enhance the scintillation mechanism. Absorption measurements indicate that the 6 Li concentration is uniform throughout the samples and its absorption efficiency as a function of thickness follows general nuclear theory, indicating that the variation in apparent brightness is likely due to a combination of particle escape, light transport, and activation of the scintillation mechanisms. As a result, the presence of 115 In and its long-lived 116 In activation product did not result in ghosting (memory of past neutron exposure), demonstrating potential for using LISe for imaging transient systems.

38 RADIATION CHEMISTRY, RADIOCHEMISTRY, AND NUCLEA↗

FPDeep: Scalable Acceleration of CNN Training on Deeply-Pipelined FPGA Clusters

In this paper, we propose a framework called FPDeep, which uses a hybrid of model and layer paral- lelism to configure distributed reconfigurable clusters to train DNNs. This approach has numerous benefits. First, the design does not suffer from batch size growth. Second, novel workload and weight partitioning leads to balanced loads of both among nodes. And third, the entire system is fine-grained pipeline. This leads to high parallelism and utilization and also minimizes the time features need to be cached while waiting for back-propagation.

Wang, Tianqi↗

MPI-FAUN: An MPI-Based Framework for Alternating-Updating Nonnegative Matrix Factorization

Non-negative matrix factorization (NMF) is the problem of determining two non-negative low rank factors W and H, for the given input matrix A, such that A≈WH. NMF is a useful tool for many applications in different domains such as topic modeling in text mining, background separation in video analysis, and community detection in social networks. Despite its popularity in the data mining community, there is a lack of efficient parallel algorithms to solve the problem for big data sets. The main contribution of this work is a new, high-performance parallel computational framework for a broad class of NMF algorithms that iteratively solves alternating non-negative least squares (NLS) subproblems for W and H. It maintains the data and factor matrices in memory (distributed across processors), uses MPI for interprocessor communication, and, in the dense case, provably minimizes communication costs (under mild assumptions). The framework is flexible and able to leverage a variety of NMF and NLS algorithms, including Multiplicative Update, Hierarchical Alternating Least Squares, and Block Principal Pivoting. Our implementation allows us to benchmark and compare different algorithms on massive dense and sparse data matrices of size that spans from few hundreds of millions to billions. We demonstrate the scalability of our algorithm and compare it with baseline implementations, showing significant performance improvements. The code and the datasets used for conducting the experiments are available online.

2D↗

Scalable Computation of Topological Abstractions for Scalar Data

Topological data analysis has become an important tool for large scale scalar data analysis and visualization, efficiently extracting the inherent structure and features of interest of the data. However, with growing dataset sizes and complexity, it is increasingly becoming infeasible to compute topological abstractions of interest in serial and on single machines. This paper presents the state of the art in the scalable computation of topological abstractions on scalar data, in shared memory parallel on single machines, and in distributed memory parallel on multiple machines. We highlight results for set‐based, graph‐based and complex‐based abstractions and organize the state of the art based on this taxonomy. The paper identifies parallelization and distribution techniques common in topological algorithms and highlights further areas of interest with underdeveloped efforts.

97 MATHEMATICS AND COMPUTING↗

Efficient three-dimensional Laplace-Fourier domain acoustic-wave simulations using discontinuous finite-difference meshes with embedded boundaries

© 2018 SEG. We develop embedded boundary and discontinuous mesh methods to handle arbirarily shaped topography and accurately simulate acoustic seismic wave propagation in Laplace-Fourier domain. The purpose of the embedded boundary method is to enhance accurate wave simulation near the surface and the discontinuous mesh method is used to achieve considerable savings in both computation time and memory savings relative to fixed mesh schemes.

AlSalem, Hussain↗