Parallel Computing Research Articles

DC-Net: Divide-and-conquer for salient object detection

In this paper, to guide the model’s training process to explicitly present a progressive trend, we first introduce the concept of Divide-and-Conquer into Salient Object Detection (SOD) tasks, called DC-Net. Our DC-Net guides multiple encoders to solve different subtasks and then aggregates the feature maps with different semantic information obtained by multiple encoders into the decoder to predict the final saliency map. The decoder of DC-Net consists of newly designed two-level Residual nested-ASPP (ResASPP2) modules, which improve the sparse receptive field existing in ASPP and the disadvantage that the U-shape structure needs downsampling to obtain a large receptive field. Based on the advantage of Divide-and-Conquer’s parallel computing, we parallelize DC-Net through reparameterization, achieving competitive performance on five LR-SOD and five HR-SOD datasets under high efficiency (60 FPS and 55 FPS). Codes and results are available: https://github.com/PiggyJerry/DC-Net.

GPU-accelerated estimation of heat transfer coefficients in continuous casting under large interference by a novel multiagent-based dimensional learning Jaya algorithm

This paper addresses the challenge of estimating heat transfer coefficients (HTCs) through the measured surface temperatures under large interference in the secondary cooling zone (SCZ) of continuous casting processes. Conventional methods for calculating HTCs, typically reliant on surface temperature data, face significant challenges in achieving accuracy. These challenges primarily stem from the limitations inherent in traditional methodologies and the distortions caused by temperature outliers, which are a consequence of large interference. To address these issues more effectively, we introduced a novel algorithm–multi-agent and dimensional learning based Jaya algorithm (MADL-Jaya)–specifically designed to estimate HTCs. Furthermore, a hybrid method that integrates weighted least squares (WLS) with the MADL-Jaya was proposed to mitigate the impact of outliers. Additionally, estimating HTCs entails substantial computational demands. Consequently, a double-deck parallel algorithm employing graphics processing unit (GPU) acceleration was presented. This architecture facilitates both parallel heat transfer models and parallel MADL-Jaya calculations, thus significantly reducing the computational time required for HTC estimation. The experimental results unequivocally affirm the efficacy of the proposed algorithm in the accurate estimation of HTCs within the SCZ of continuous casting, effectively neutralizing the detrimental influence of outlier values while markedly improving computational speed.

A Hybrid Parallel Computing Architecture Based on CNN and Transformer for Music Genre Classification

Music genre classification (MGC) is the basis for the efficient organization, retrieval, and recommendation of music resources, so it has important research value. Convolutional neural networks (CNNs) have been widely used in MGC and achieved excellent results. However, CNNs cannot model global features well due to the influence of the local receptive field; these global features are crucial for classifying music signals with temporal properties. Transformers can capture long-range dependencies within an image thanks to adopting the self-attention mechanism. Nevertheless, there are still performance and computational cost gaps between Transformers and existing CNNs. In this paper, we propose a hybrid architecture (CNN-TE) based on CNN and Transformer encoder for MGC. Specifically, we convert the audio signals into mel spectrograms and feed them into a hybrid model for training. Our model employs a CNN to initially capture low-level and localized features from the spectrogram. Subsequently, these features are processed by a Transformer encoder, which models them globally to extract high-level and abstract semantic information. This refined information is then classified using a multi-layer perceptron. Our experiments demonstrate that this approach surpasses many existing CNN architectures when tested on the GTZAN and FMA datasets. Notably, it achieves these results with fewer parameters and a faster inference speed.

Improving the precision of forces in real-space pseudopotential density functional theory.

The high-order finite difference real-space pseudopotential density functional theory (DFT) approach is a valuable method for large-scale, massively parallel DFT calculations. A significant challenge in the approach is the oscillating "egg-box" error introduced by aliasing associated with a coarse grid spacing. To address this issue while minimizing computational cost, we developed a finite difference interpolation (FDI) scheme [Roller et al., J. Chem. Theory Comput. 19, 3889 (2023)] as a means of exploiting the high resolution of the pseudopotential to reduce egg-box effects systematically. Here, we show an implementation of this method in the PARSEC code and examine the practical utility of the combination of FDI with additional methods for improving force precision and/or reducing its computational cost, including orbital-based forces, compensating charges (namely, adding and subtracting a judiciously chosen charge density such that the total density is unaltered), and a modified spatial domain in which the real-space grid is defined. Using selected small molecules, as well as metallic Li, as test cases, we show that a combination of all four aspects leads to a significant reduction in computational cost while retaining a high level of precision that supports accurate structures and vibrational spectra, as well as stable and accurate molecular dynamics runs.

SNN-BERT: Training-efficient Spiking Neural Networks for energy-efficient BERT

Spiking Neural Networks (SNNs) are naturally suited to process sequence tasks such as NLP with low power, due to its brain-inspired spatio-temporal dynamics and spike-driven nature. Current SNNs employ ”repeat coding” that re-enter all input tokens at each timestep, which fails to fully exploit temporal relationships between the tokens and introduces memory overhead. In this work, we align the number of input tokens with the timestep and refer to this input coding as ”individual coding”. To cope with the increase in training time for individual encoded SNNs due to the dramatic increase in timesteps, we design a Bidirectional Parallel Spiking Neuron (BPSN) with following features: First, BPSN supports spike parallel computing and effectively avoids the issue of uninterrupted firing; Second, BPSN excels in handling adaptive sequence length tasks, which is a capability that existing work does not have; Third, the fusion of bidirectional information enhances the temporal information modeling capabilities of SNNs; To validate the effectiveness of our BPSN, we present the SNN-BERT, a deep direct training SNN architecture based on the BERT model in NLP. Compared to prior repeat 4-timestep coding baseline, our method achieves a 6.46× reduction in energy consumption and a significant 16.1% improvement, raising the performance upper bound of the SNN domain on the GLUE dataset to 74.4%. Additionally, our method achieves 3.5× training acceleration and 3.8× training memory optimization. Compared with artificial neural networks of similar architecture, we obtain comparable performance but up to 22.5× energy efficiency. We would provide the codes.

IDAD: An improved tensor train based distributed DDoS attack detection framework and its application in complex networks

With the vigorous development of Internet technology, the scale of systems in the network has increased sharply, which provides a great opportunity for potential attacks, especially the Distributed Denial of Service (DDoS) attack. In this case, detecting DDoS attacks is critical to system security. However, current detection methods exhibit limitations, leading to compromises in accuracy and efficiency. To cope with it, three key strategies are implemented in this paper: (i) Using tensors to model large-scale and heterogeneous data in complex networks; (ii) Proposing a denoising algorithm based on the improved and distributed tensor train (IDTT) decomposition, which optimizes the tensor train(TT) decomposition in terms of parallel computation and low-rank estimation; (iii) Combining (i), (ii) and Light Gradient Boosting Machine (LightGBM) classification model, an efficient DDoS attack detection framework is proposed. Datasets CIC-DDoS2019 and NSL-KDD are used to evaluate the framework, and results demonstrate that accuracy can reach 99.19% while having the characteristics of low storage consumption and well speedup ratio.

Parallel Algebraic Effect Handlers

Algebraic effect handlers support composable and structured control-flow abstraction. However, existing designs of algebraic effects often require effects to be executed sequentially. This paper studies parallel algebraic effect handlers. In particular, we formalize λ p , a lambda calculus which models two key features, effect handlers and parallelizable computations, the latter of which takes the form of a for expression, inspired by the Dex programming language. We present various interesting examples expressible in our calculus. To show that our design can be implemented in a type-safe way, we present a higher-order polymorphic lambda calculus F p that extends λ p with a lightweight value dependent type system, and prove that F p preserves the semantics of λ p and enjoys syntactic type soundness. Lastly, we provide an implementation of the language design as a Haskell library, which mirrors both λ p and F p and reveals new connections to free applicative functors. All examples presented can be encoded in the Haskell implementation. We believe this paper is the first to study the combination of user-defined effect handlers and parallel computations, and it is our hope that it provides a basis for future designs and implementations of parallel algebraic effect handlers.

Machine Learning Analysis Using the Black Oil Model and Parallel Algorithms in Oil Recovery Forecasting

The accurate forecasting of oil recovery factors is crucial for the effective management and optimization of oil production processes. This study explores the application of machine learning methods, specifically focusing on parallel algorithms, to enhance traditional reservoir simulation frameworks using black oil models. This research involves four main steps: collecting a synthetic dataset, preprocessing it, modeling and predicting the oil recovery factors with various machine learning techniques, and evaluating the model’s performance. The analysis was carried out on a synthetic dataset containing parameters such as porosity, pressure, and the viscosity of oil and gas. By utilizing parallel computing, particularly GPUs, this study demonstrates significant improvements in processing efficiency and prediction accuracy. While maintaining the value of the R2 metric in the range of 0.97, using data parallelism sped up the learning process by, at best, 10.54 times. Neural network training was accelerated almost 8 times when running on a GPU. These findings underscore the potential of parallel machine learning algorithms to revolutionize the decision-making processes in reservoir management, offering faster and more precise predictive tools. This work not only contributes to computational sciences and reservoir engineering but also opens new avenues for the integration of advanced machine learning and parallel computing methods in optimizing oil recovery.

Damage-related imbalance identification for UAV composite propeller blades based on bidirectional temporal convolutional network and a flexible sensing system

Abstract Damage to the composite propeller blades could lead to rotational imbalance, which seriously affects the operational safety of unmanned aerial vehicles (UAVs), therefore, a novel method combining the Teager energy operator (TEO) and bidirectional temporal convolutional network (BiTCN) is proposed for detecting, localizing, and quantifying the damage-related imbalance in the blades. A flexible sensing system that contains Micro electro mechanical sensor accelerometers, signal conditioning, and wireless transmission is integrated with the composite propeller for in-situ signal acquisition of the propeller blades. TEO is applied to demodulate and enhance the pulse compositions in vibration signals and singular value decomposition (SVD) is employed to suppress random noise, resulting in denoised Teager energy spectrums for model input. Temporal convolutional network (TCN) has been widely used in sequence signal modeling because the causal dilated convolution could learn the context information of sequence signals while maintaining the advantages of parallel computing. To fully extract the signal features, BiTCN models are established to learn both the forward and backward signal features. Experimental verification results show that the proposed method detects the existence of imbalance with 100% accuracy, and the accuracies of localization and quantization are 99.65% and 98.61%, respectively, which are much higher than those of the models with the original signal as input. In addition, compared with the other four different algorithms, BiTCN is superior in terms of convergence speed and prediction accuracy.

Proposal for optimizing number of servers in closed BCMP queueing network

AbstractIn this study, we use a closed BCMP queueing network model designed for multiple customer classes and servers to optimize the number of servers at each node. This optimization is achieved by setting an upper limit on the number of servers and using an objective function that combines the standard deviation of the average number of customers in the system with the server installation cost. We use a genetic algorithm with parallel computations for the optimization process. Our findings demonstrate that this approach is viable for closed BCMP network models that require extensive computational resources. The optimal server count is validated by comparing the optimization results with the maximum number of servers utilized. Node popularity is predetermined, and a gravity model is employed to generate transition probabilities, rendering the model applicable to real-world scenarios. Our optimization results indicate that both the node popularity and distance between nodes influence the server count. Furthermore, simulations were conducted to evaluate the effect of the number of servers on the optimization outcomes. Allowing variations in the node count, location, and popularity makes this study flexible and adaptable to various real-world scenarios, such as transportation systems, healthcare facilities, and commercial spaces. Moreover, by providing an efficient and scalable solution, this study serves as a cornerstone for future research in this field and offers a practical tool for facility managers aiming to minimize both congestion and operational costs.

Abstractions for C++ code optimizations in parallel high-performance applications

Many computational problems consider memory throughput a performance bottleneck, especially in the domain of parallel computing. Software needs to be attuned to hardware features like cache architectures or concurrent memory banks to reach a decent level of performance efficiency. This can be achieved by selecting the right memory layouts for data structures or changing the order of data structure traversal. In this work, we present an abstraction for traversing a set of regular data structures (e.g., multidimensional arrays) that allows the design of traversal-agnostic algorithms. Such algorithms can easily optimize for memory performance and employ semi-automated parallelization or autotuning without altering their internal code. We also add an abstraction for autotuning that allows defining tuning parameters in one place and removes boilerplate code. The proposed solution was implemented as an extension of the Noarr library that simplifies a layout-agnostic design of regular data structures. It is implemented entirely using C++ template meta-programming without any nonstandard dependencies, so it is fully compatible with existing compilers, including CUDA NVCC or Intel DPC++. We evaluate the performance and expressiveness of our approach on the Polybench-C benchmarks.

Research on image compression technology based on improved SPIHT compression algorithm for power grid data

Abstract At present, the amount of information on power grid operation and maintenance monitoring image data is increasing, and the requirements for data compression are higher and higher. Based on the improved SPIHT image compression algorithm, this study presents the research of power grid data compression. First, the basic theory of image compression, and the principle of one-dimensional wavelet transform and two-dimensional wavelet transform are introduced. The development process, characteristics, and advantages of image coding are discussed. Then, the shortcomings of the SPIHT algorithm are analyzed, and the SPIHT coding is improved by parallel computation. The parallel wavelet transform algorithm based on the block idea and the parallel SPIHT coding algorithm based on the code tree are proposed in the data parallelism of the compression algorithm. At the same time, the data dependence between tasks in the process of SPIHT image compression coding is analyzed, and the task parallelism in the compression algorithm is realized by using the relative independence of tasks in different threshold coding. Finally, the application and simulation analysis of power grid data based on the SPIHT compression algorithm, the construction of power grid data model simulation, and the composition of two-dimensional power grid data images are carried out. Secondly, the obtained 2D power grid data image is compressed by the SPIHT algorithm and improved SPIHT algorithm, respectively, and the compression effect of the two algorithms on the power grid data image is compared. When the bit rate is 0.5, the compression effect of the improved SPIHT algorithm is 13.6506. When the bit rate is 1, the compression effect of the improved SPIHT algorithm is 18.9287. The results show that the improved SPIHT algorithm can compress the grid data to obtain better grid image quality.

Ultrasound normalized cumulative residual entropy imaging: Theory, methodology, and application

Background and Objective:Ultrasound information entropy imaging is an emerging quantitative ultrasound technique for characterizing local tissue scatterer concentrations and arrangements. However, the commonly used ultrasound Shannon entropy imaging based on histogram-derived discrete probability estimation suffers from the drawbacks of histogram settings dependence and unknown estimator performance. In this paper, we introduced the information-theoretic cumulative residual entropy (CRE) defined in a continuous distribution of cumulative distribution functions as a new entropy measure of ultrasound backscatter envelope uncertainty or complexity, and proposed ultrasound CRE imaging for tissue characterization. Methods:We theoretically analyzed the CRE for Rayleigh and Nakagami distributions and proposed a normalized CRE for characterizing scatterer distribution patterns. We proposed a method based on an empirical cumulative distribution function estimator and a trapezoidal numerical integration for estimating the normalized CRE from ultrasound backscatter envelope signals. We presented an ultrasound normalized CRE imaging scheme based on the normalized CRE estimator and the parallel computation technique. We also conducted theoretical analysis of the differential entropy which is an extension of the Shannon entropy to a continuous distribution, and introduced a method for ultrasound differential entropy estimation and imaging. Monte-Carlo simulation experiments were performed to evaluate the estimation accuracy of the normalized CRE and differential entropy estimators. Phantom simulation and clinical experiments were conducted to evaluate the performance of the proposed normalized CRE imaging in characterizing scatterer concentrations and hepatic steatosis (n = 204), respectively. Results:The theoretical normalized CRE for the Rayleigh distribution was π/4, corresponding to the case where there were ≥10 randomly distributed scatterers within the resolution cell of an ultrasound transducer. The theoretical normalized CRE for the Nakagami distribution decreased as the Nakagami parameter m increased, corresponding to that the ultrasound backscattered statistics varied from pre-Rayleigh to Rayleigh and to post-Rayleigh distributions. Monte-Carlo simulation experiments showed that the proposed normalized CRE and differential entropy estimators can produce a satisfying estimation accuracy even when the size of the test samples is small. Phantom simulation experiments showed that the proposed normalized CRE and differential entropy imaging can characterize scatterer concentrations. Clinical experiments showed that the proposed ultrasound normalized CRE imaging is capable to quantitatively characterize hepatic steatosis, outperforming ultrasound differential entropy imaging and being comparable to ultrasound Shannon entropy and Nakagami imaging. Conclusion:This study sheds light on the theory and methodology of ultrasound normalized CRE. The proposed ultrasound normalized CRE can serve as a new, flexible quantitative ultrasound envelope statistics parameter. The proposed ultrasound normalized CRE imaging may find applications in quantified characterization of biological tissues. Our code will be made available publicly at https://github.com/zhouzhuhuang.

Massively parallel axisymmetric fluid model for streamer discharges

A highly parallelizable fluid plasma simulation tool based upon the first-order drift-diffusion equations is discussed. Atmospheric pressure plasmas have densities and gradients that require small element sizes in order to accurately simulate the plasm resulting in computational meshes on the order of millions to tens of millions of elements for realistic size plasma reactors. To enable simulations of this nature, parallel computing is required and must be optimized for the particular problem. Here, a finite-volume, electrostatic drift-diffusion implementation for low-temperature plasma is discussed. The implementation is built upon the Message Passing Interface (MPI) library in C++ using Object Oriented Programming. The underlying numerical method is outlined in detail and benchmarked against simple streamer formation from other streamer codes. Electron densities, electric field, and propagation speeds are compared with the reference case and show good agreement. Convergence studies are also performed showing a minimal space step of approximately 4 μm required to reduce relative error to below 1% during early streamer simulation times and even finer space steps are required for longer times. Additionally, strong and weak scaling of the implementation are studied and demonstrate the excellent performance behavior of the implementation up to 100 million elements on 1024 processors. Finally, different advection schemes are compared for the simple streamer problem to analyze the influence of numerical diffusion on the resulting quantities of interest.

GPU accelerated sparse curvelet-constrained wavefield reconstruction inversion with source estimation: Application to Chevron Benchmark 2014 blind test data set

Full-waveform inversion (FWI) is a pivotal tool in seismic exploration and seismology but encounters a persistent challenge known as “cycle-skipping,” causing it to frequently converge to local minima. Wavefield reconstruction inversion (WRI) has emerged as a potential solution to mitigate the risk of cycle skipping. Researchers have provided numerical examples indicating that WRI is less prone to being ensnared in local minima caused by the absence of low-frequency component data and the absence of a well-defined initial model compared with conventional FWI. Despite its potential, the computational demands of WRI have hindered its widespread application, especially in scenarios involving ocean towed streamer seismic acquisition and unknown sources, where the augmented systems differ from source to source. In our study, we introduce a novel approach — sparse curvelet-constrained WRI with source estimation (WRI-SE-CC) — accelerated by graphics processing unit (GPU). Real-time source function estimation is achieved through the variable projection method, and noise-related artifacts are suppressed using sparse curvelet constraints. By optimizing the utilization of hundreds of computation processors within a GPU for parallel computing of matrix-vector multiplications, we present a GPU-based grouped conjugate gradient method to accelerate the computation of WRI-SE-CC. Numerical experiments demonstrate a significant 240-fold acceleration compared with the preconditioned conjugate gradient using one CPU core for computations involving multiple sources. Inversion experiments with the overthrust model demonstrate the capability of our method in mitigating local minima and suppressing noise-related artifacts. Furthermore, we validate the framework on the Chevron 2014 blind test data set, showcasing its effectiveness in addressing practical challenges in the field.

Replicated multistage interconnection networks: QoS evaluation for parallel and distributed computing

Introduction to big data becomes very important with dealing in high-performance parallel distributed computing, especially in those systems where communication among a number of processors is required. The current paper takes into consideration different topologies for the Shuffle Exchange Network (SEN) in order to attain optimal data transfer among such scenarios. SEN topologies are a key feature in connecting several processors and implementing the data transfer among them where a single processor fails to handle the load. The study hereby reports in the performance, reliability, and cost analysis of these topologies. These topologies, some of which are advocated to have better performance in many studies, include replicated networks. Our research demystifies claims made in earlier papers that replicated networks have inflated reliability and higher costs due to their additional links. Researchers are therefore guided on accurate performance data that will lead them to make optimum choices of SEN topologies for targeted applications. These findings further highlight that the trade-offs between reliability and cost must be carefully considered during network design so as to arrive at better results for big data communication and computation in parallel and distributed systems. This work provides important insights into the correct evaluation of SEN topologies, helping to correct the misleading facts and to have better network selection for various real-time application realizations.

Understanding the Non-Volatile Resistive Switching Behavior of Lithium Titanium Oxide during Spinel to Rock Salt Conversion for Neuromorphic Computing

Lithium based synaptic device research has been a growing new field in the last few years for enabling robust neuromorphic hardware systems. It can realize parallel analogue computation which is beyond the conventional von Neumann computing architecture which is slow, energy inefficient and expensive to execute complex tasks. In this study, we showcase the remarkable potential of Lithium Titanium Oxide (Li4Ti5O12) spinel, both computationally through Density of States (DoS) analysis using Density Functional Theory (DFT), and experimentally via direct current (DC) polarization investigation of electrochemically lithiated Li4Ti5O12 thin film, examining diverse states of discharge (SoD) in both in-plane and out-of-plane configurations. The observed conductivity jump spans up to six orders of magnitude, transitioning from the Li4Ti5O12 spinel phase to the fully lithiated Li7Ti5O12 rock salt phase. Raman Spectroscopy validates the transformation from the Li4Ti5O12 spinel phase to the Li7Ti5O12 rock salt phase. Moreover, the onset of increase in conductivity during this transformation is corroborated by changes in the oxidation state of Titanium (Ti), as explained with the help of X-Ray Photoelectron Spectroscopy. To actualize these findings, we engineered a three-terminal artificial synaptic device integrating Lithium Phosphorous Oxynitride (LiPON) solid-state electrolyte as the lithium ion source and conductor, while employing Lithium Titanium Oxide (Li4Ti5O12) spinel as the channel material. In device testing, we maintained a source-drain voltage (VSD) of 0.5 V, while sweeping the gate voltage (VG) from -3 V to 3 V, resulting in switching spanning up to 4 orders of magnitude, consistently replicable across multiple cycles. Widening the gate voltage window from ±3 V to ±7V induced an increase in conductance. Furthermore, non-volatile testing affirmed the sustained retention of the conductance state across all tested gate voltage windows, showcasing the device's stability over time. These compelling results portray a promising trajectory for the development of a robust synaptic device architecture reliant on lithium-based all solid-state materials, catalyzing the evolution of high-precision analogue neuromorphic computing systems.

(Invited) Emerging Nonvolatile Memories for Analog Neuromorphic Computing

Emerging non-volatile memory (NVM) devices, such as STT-MRAM, PCM, RRAM, have been explored for embedded memory and storage applications to replace CMOS-based SRAM/DRAM and Flash devices. Recently, many of these memory devices have been utilized for new computing paradigms beyond Boolean logic and von Neumann architectures. For example, in-memory analog computing reduces data movement between computing and memory units and exploits the intrinsic parallelism in memory arrays. It finds a natural application in deep neural network (DNN) accelerators by implementing high-throughput high-efficiency multiply accumulate (MAC) operations. Here the conductance of memory devices in a crossbar array represents DNN weights and the activations are encoded in input electrical signals (e.g., pulse height or duration). The MAC operation is conducted via the Ohm’s law (multiplication between voltage and conductance) and Kirchhoff’s law (accumulate via current summation) at constant time even for very large networks. DNN has surpassed human performance in various AI applications, e.g., image classification, natural language processing, etc. While general-purpose CPU/GPU and special-purpose digital accelerators provide current and near-term DNN hardware, there are longer-term opportunities for analog DNN accelerators based on emerging memory devices to achieve significantly higher performance and energy-efficiency. At the same time, analog accelerators impose new requirements on these devices beyond traditional memory applications, e.g., analog tunability, gradual and symmetric weight modulation, high precision, etc. Memory devices with analog nature in their physical mechanisms (e.g., filament growth in RRAM) may be optimized to meet these requirements, while some abrupt and asymmetric characteristics (e.g., filament rupture) present challenges. Increasingly large neural network models have been demonstrated on these memory arrays designed as analog accelerators, but they are still orders of magnitude smaller than state-of-the-art DNN models. While analog accelerators enable massively parallel computation, they are also susceptible to unique challenges in analog devices and circuitry (e.g., device variability, circuit noise), which may degrade network performance (e.g., accuracy).To benefit from the massively parallel MAC operation in analog memory arrays, these arrays need to be large enough to efficiently map the layers in modern DNN models. Among emerging NVM devices, PCM has the advantages of maturity and the availability of large-scale arrays, but also face some challenges in device characteristics, e.g., conductance drift, asymmetry, and noise. PCM-based analog DNN accelerators have been demonstrated at advanced technology node with millions of devices and achieved iso-accuracy on increasingly large network models. These accelerators integrate highly efficient analog PCM tiles for MAC operations with advanced CMOS circuitry for auxiliary digital functions. While material/device engineering continues to be explored to improve the analog properties of PCM devices, design and operation innovations can also help to improve the performance of PCM-based DNN weights, e.g., multiple-device-per-weight design, close-loop tunning. In addition, circuit innovations are essential for analog accelerator performance. Fig. 1 shows a 14nm PCM-based DNN inference accelerator, which incorporate design techniques such as 4-PCM weight units, 2D mesh for tile-to-tile communication, pulse-duration-based coding, etc. On top of technology and design innovations, some DNN models can also be modified to be more resilient against hardware imperfection and noise. PCM-based analog accelerators have achieved iso-accuracy on large DNN models with millions of weights. This talk will discuss the progress that we have achieved on PCM-based analog DNN inference accelerators, the challenges of PCM materials and devices, and promising solutions in technology and design. Figure 1

Atmospheric Aircraft Conceptual Design Based on Multidisciplinary Optimization with Differential Evolution Algorithm and Neural Networks

A methodology for selecting rational parameters of atmospheric aircraft during the initial design stages using a differential evolutionary optimization algorithm and numerical mathematical modeling of aerodynamics problems is proposed. The technique involves implementing weight and aerodynamic balance in the main flight modes, considering atmospheric aircraft with one or two lifting surfaces, applying parallel calculations, and auto-generating a three-dimensional geometric model of the aircraft’s appearance based on the optimization results. A method for accelerating the process of optimizing aircraft parameters in terms of takeoff weight by more than three times by introducing an objective function into the set of design variables is proposed and demonstrated. The reliability of mathematical models used in aerodynamics and the accuracy of the objective function calculation considering various constraints are explored. A comprehensive test of the performance and efficiency of the methodology is conducted by solving demonstration problems to optimize more than ten main design parameters for the appearance of two existing heavy-class unmanned aerial vehicles with known characteristics from open sources.

Exchange-correlation kernel for perturbation dependent auxiliary functions in auxiliary density perturbation theory

ContextAnalytic exchange-correlation kernel formulations are of the outermost importance for density functional theory (DFT) perturbation calculations. In this paper, the working equation for the exchange-correlation kernel of the generalized gradient approximation (GGA) for perturbation dependent auxiliary functions is derived and discussed in the framework of auxiliary density functional theory (ADFT). The presented new formulation is extended to the unrestricted approach, too. A comprehensive discussion of the implementation of the GGA ADFT kernel, using either the native exchange-correlation functional implementations in deMon2k or the ones from the LibXC library, is given. Calculations with analytic exchange-correlation kernels are compared to their finite difference counterparts. The obtained results are in quantitative agreement. Nevertheless, analytic GGA ADFT kernel implementations show substantial improvement in the computational performance. Similar results are reported for analytic second derivatives of effective core potential (ECP) and model core potential (MCP) matrix elements when compared to their finite difference counterparts in molecular frequency analyses.MethodAll calculations are performed in the framework of ADFT as implemented in deMon2k. In the ADFT analytic frequency calculations, auxiliary density perturbation theory was used. The underlying two-center exchange-correlation kernel matrix elements are calculated by numerical integration either with analytic or finite difference kernel expressions. Validation calculations are performed with the VWN and PBE functionals employing DFT-optimized DZVP basis sets in conjunction with automatically generated GEN-A2 auxiliary density function sets. In the (Pt3Cu)n cluster benchmark calculations, the RPBE functional was used. For Pt atoms, the quasi-relativistic LANL2DZ effective core potential with the corresponding valence basis set was employed, whereas for Cu atoms, the all-electron DFT-optimized TZVP basis was applied. The auxiliary density was expanded by the automatically generated GEN-A2* auxiliary function set. We run all benchmark calculations in parallel on 24 cores.