Compute Unified Device Architecture NVIDIA CUDA

Complex systems research has long faced the fundamental challenge of disciplinary isolation and lack of unified mathematical language. This paper starts from a classic constrained dynamics problem, whose analytical solution reveals a universal principle of system architecture decomposition: overall behavior consists of the synergy between "intrinsic growth" () and "interaction effects" (). Based on this, we propose a dual paradigm unified framework that classifies systems into two basic paradigms: efficiency-driven (polynomial growth) and innovation-driven (exponential growth), and establishes a corresponding fivedimensional parameter diagnostic system. The breakthrough of the theory lies in its quantum extension: we extend it to the field of social information dynamics, constructing a formal quantum social dynamics model that achieves cross-scale unified description from microscopic cognition to macroscopic social regulation. This framework not only provides a unified explanation for classic laws like urban scaling laws, but also demonstrates breakthrough potential in solving AI bottlenecks and social governance challenges through the hybrid cognitive AI architecture and social multi-network regulation system it inspires.

Primeiramente, como de praxe de minha parte, agradeço a Deus, seja lá como ele for, por existir. Agradeço aos meus pais e minha irmã por me conhecerem o suficiente para confiarem a mim meu trabalho e não a responsabilidade de cuidar de mim mesmo. Agradeço ao meu orientador, Prof. Dr. Ádamo Lima de Santana, por ser tão companheiro, quanto orientador, me ajudando e confiando em mim, mesmo nos momentos mais complicados, sempre acreditando que posso evoluir mais. Agradeço à meus amigos Nathalia Nascimento, Petterson Marques e Iury Batalha pelo sacrifício de me ajudar mesmo quando doentes e ocupados, dizendo que "é isso que faz um amigo". Agradeço aos meus colegas do LINC, em especial ao grupo de Computação Bioinspirada, Igor, Iago e Paulo, que me auxiliaram no desenvolvimento deste trabalho. Por último, mas não menos importante, agradeço à banca examinadora desta Dissertação de Mestrado, os Professores

A aplicacao em logistica de distribuicao e diversa, a exemplo do planejamento de transporte e entrega de mercadorias ou no roteamento de dados em redes de telecomunicacoes. Dado a amplitude e capilaridade desses problemas, trabalhos vem sendo desenvolvidos visando reduzir os gastos para o funcionamento de redes dessa magnitude, sobretudo no que tange a demanda de energia eletrica. Sendo assim, o presente trabalho apresenta uma proposta de metodo de resolucao de problemas de roteamento com alto grau de demanda. O metodo proposto e baseado em algoritmos bioinspirados, que aliados a outros metodos, garantem a integridade das solucoes obtidas, alem de sua proximidade ao otimo. Entretanto, tais algoritmos se tornam computacionalmente custosos a medida que a complexidade da aplicacao em questao aumenta e, portanto, ambientes multiprocessados, como plataformas de computacao em GPU, vem sendo largamente utilizados para aumentar a performance dos mesmos. Sendo assim, este trabalho visa reali...

Bio-inspired techniques like Genetic Algorithms have a comprehensive applicability to optimization problems. Given the ease of parallelism implementation inherent of these techniques several researches have been developed in such area making use of parallel platforms, especially the CUDA platform. However, the majority of these works are focused on strategies to improve the algorithms convergence without concern for the performance against their runtime. In this context, the present research introduces a runtime comparison of a Genetic Algorithm implementation in CUDA and traditional CPU aiming to investigate the performance difference between them. The results showed that although the Genetic Algorithm has a native parallelism, the CPU overcomes the GPU implemetation due to a hardware limitation.

Graphics Processing Units (GPUs), originally developed for computer games, now provide computational power for scientific applications. A study on the comparison of computational speed-up and efficiency of a GPU with a CPU for the Finite Pointset Method (FPM), which is a numerical tool in Computational Fluid Dynamics (CFD) is presented. As FPM is based on the point cloud, it is so expensive when the number of particles are in millions. We have demonstrated the application of the FPM using a single-GPU (Nvidia Tesla M2050) and Intel CPU (Dual Xeon). Importance of the GPU is realized by the FPM since GPU yields a computational speed-up of 70× for the Poisson equation with various boundary conditions.

This paper presents optical flow estimation technique to estimate the motion vectors in each frame of the video sequence. By thresholding and performing morphological closing on the motion vectors, we produces binary feature images. Using these binary features the cars are located. A bounding Box is drawn around the cars that pass beneath the white line. The algorithm used for this is lucas kanade. Use of the threshold to reduce the noise in small movements between frames is analyzed. Higher the threshold ,the less small movements impact the optical flow calculation. Experiments are done to find the value that best achieves our results.

The numerical solution of the ideal magnetohidrodinamica equations (MHD) is presented. The study corresponds to a non-steady problem. The technique introduced by Powell [1] is used. The eigenvectors are normalized to avoid problems during the simulation. To carry out the numeric simulation an approximate Riemann solver is used in combined form with a TVD technique to evaluate the numeric flows. The analyzed problem to verify the appropriate behavior of the computacional code is of Riemann problem for the MHD. This is an extension of the shock tube used in gas dynamic. The obtained results show the richness of the single fluid MHD solution. Also the results here presented satisfies those found by other authors.

In questo lavoro, dopo alcuni cenni storici sullo sviluppo della teoria della valutazione delle opzioni, si è messo in evidenza che la mancanza di negoziabilità delle stock options assegnate a dirigenti e dipendenti non mina le fondamenta teoriche (il dynamic hedging) su cui si basa la dimostrazione di Merton della formula Black-Scholes.
Tuttavia, se i problemi di reputazione non consentono agli assegnatari di aggirare il divieto di negoziabilità delle opzioni, la valutazione delle stock options esercitabili prima della scadenza si complica. Occorre infatti tener conto del fatto che l’impossibilità di cedere le stock options costringe gli assegnatari ad esercitarle in modo sub-ottimale. Ne segue che, ai fini valutativi, occorre far ricorso alla classica massimizzazione dell’utilità in un modello d’equilibrio e, in quest’ottica, l’approccio più avanzato è quello proposto da Jonathan Ingersoll nell’ambito del capital asset pricing model in tempo continuo.
Secondo quest’approccio, il valore oggettivo delle opzioni da riportare in bilancio va calcolato in base a parametri di mercato ma solo dopo aver determinato – da un punto di vista soggettivo – le condizioni di esercizio ottimale per gli assegnatari.
Un’esemplificazione chiarisce i diversi passi da compiere per giungere alla determinazione del fair value (valore oggettivo) delle opzioni.

This paper describes a numerical method for the parallel solution of the differential measure inclusion problem posed by mechanical multibody systems containing bilateral and unilateral frictional constraints. The method proposed has been implemented as a set of parallel algorithms leveraging NVIDIA’s Compute Unified Device Architecture (CUDA) library support for multi-core stream computing. This allows the proposed solution to run on a wide variety of GeForce and TESLA NVIDIA graphics cards for high performance computing. Although the methodology relies on the solution of cone complementarity problems known to be fine-grained in terms of data dependency, a suitable approach has been developed to exploit parallelism with low overhead in terms of memory access and thread synchronization. Additionally, a parallel collision detection algorithm has been incorporated to further exploit available parallelism. Initial numerical tests described in this paper demonstrate a speedup of one ord...

GPU based on CUDA Architecture developed by NVIDIA is a high performance computing device. Multiplication of matrices of large order can be computed in few seconds using GPU based on CUDA Architecture. A modern GPU consists of 16 highly threaded streaming multiprocessors (SMs). GPU named Fermi consists of 32 SMs. These are computing intensive devices. GPUs have been found to be the best platform for massive data parallelism. CUDA architecture is based on the heterogeneous platform comprising of both CPU and GPU that offers enormous potential to solve complex harder problems with high speed. In most applications the sequential part of a program is executed using CPU and numeric intensive part on GPU. But mere execution of numeric intensive part on GPU will not increase the performance of the computation. Since GPU consists of highly threaded multiprocessors, threads must be well organised into Grids and Grid into blocks to maximize performance of parallel computation, depending upon ...

This report details the implementation of a snapshot-driven Artificial General Intelligence (AGI) agent that integrates environmental state processing, natural language understanding, and symbolic reasoning. The agent operates in the CartPole-v1 environment from Gymnasium, a modern reinforcement learning framework, while processing a fixed text prompt to demonstrate its multimodal capabilities. Building on the provided base report, this document explains the role of the example input, summarizes the code’s functionality, addresses resolved issues, and outlines the theoretical foundations of the approach. The complete Python code is included to provide a clear reference for the
implementation.

The image segmentation consists in obtaining a region of interest within a larger area. GrabCut is a recent tech- nique of 2D segmentation which presents excellent results. This is based on representing the image as a flow network, and then apply a minimum cut algorithm on the graph and separate the background image (background) and the region of interest (foreground).

The Uintah computational framework is used for the parallel solution of partial differential equations on adaptive mesh refinement grids using modern supercomputers. Uintah is structured with an application layer and a separate runtime system. Uintah is based on a distributed directed acyclic graph (DAG) of computational tasks, with a task scheduler that efficiently schedules and executes these tasks on both CPU cores and on-node accelerators. The runtime system identifies task dependencies, creates a task graph prior to the execution of these tasks, automatically generates MPI message tags, and automatically performs halo transfers for simulation variables. Automating halo transfers in a heterogeneous environment poses significant challenges when tasks compute within a few milliseconds, as runtime overhead affects wall time execution, or when simulation variables require large halos spanning most or all of the computational domain, as task dependencies become expensive to process. These challenges are magnified at production scale when application developers require each compute node perform thousands of different halo transfers among thousands simulation variables. The principal contribution of this work is to (1) identify and address inefficiencies that arise when mapping tasks onto the GPU in the presence of automated halo transfers, (2) implement new schemes to reduce runtime system overhead, (3) minimize application developer involvement with the runtime, and (4) show overhead reduction results from these improvements.

The fast-growing advancement in AI technologies has resulted in huge loads on the data center architecture resulting in the need to create extremely resistant, and fault-tolerant AI fabrics. This paper looks at AI design principles and technologies necessitated in the construction of fault-tolerant AI infrastructures that can support complex, data-heavy workloads. The major technologies of VXLAN EVPN, RDMA and ultra-low latency interconnect like RoCEv2, NV Link and PCIe Gen5 are paramount to high availability, low latency and high throughput. This article reviews industrial best practice by observing reference architecture of industry leaders like NVIDIA DGX, Meta RSC, AWS Triennium, and reflects on practical approaches in developing a robust fabric of AI computing. The article offers an in-depth road map of how next-gen AI data centers should be designed by paying attention to failure domains, fault tolerance and optimization of convergence. Such robust AI frameworks play a pivotal role by facilitating scalable performance of AI models, inference and training.

The mining association algorithm well known as A-priori is one of the most popular data mining algorithms, The mining association algorithm requires two parameters support and confidence to derive rules, The A-priori algorithm scans the database multiple passes to discover the rules, In this research we proposed an algorithm to reduce the multiple passes to just one pass to derive rules with high confidence, which gives the same results and also reduces the time and the effort, the algorithm gives analytical ability for understanding the results because it focusing on small itemsets.

Breadth-first search (BFS) has wide applications in electronic design automation (EDA) as well as in other fields. Researchers have tried to accelerate BFS on the GPU, but the two published works are both asymptotically slower than the fastest CPU implementation. In this paper, we present a new GPU implementation of BFS that uses a hierarchical queue management technique and a three-layer kernel arrangement strategy. It guarantees the same computational complexity as the fastest sequential version and can achieve up to 10 times speedup.

lthough the scientific computing community has questioned graphics processing unit (GPU) efficiency when it comes to energy per operation, the latest Green500 list (www.green500.org) should put to rest these concerns. The Green500 sorts systems on the Top-500 list (www.top500.org) according to their power efficiency; on the latest list, released in November 2010, the number two and three systems were GPU-based. Here, we introduce the system that ranked at number three: EcoG, a 128-node GPU-accelerated cluster computer deployed at the US National Center for Supercomputing Applications (NCSA). EcoG was built through a joint effort by Nvidia, the

This paper introduces a library which supports programmers to write parallel programs on GPU architecture, especially with a system consisting of multi-GPUs. The library is designed from the idea of skeletons, which helps us to make parallel programs easily and quickly as if writing sequential programs. Skeletons usually are described by functional language which supports high-order function totally. Because of the performance and popularity of C++ language, we try to re-annotate C++ language to support high-order functions completely, hence, it is convenience for us to create general-purpose skeletons.

This study examines the integration of fundamental physical principles into three-dimensional computational algorithms and analyzes how digital systems simulate physical phenomena from the real world. Through a systematic analysis of next-generation physics engines and their mathematical foundations, this research investigates how Newtonian mechanics, collision dynamics, and force interactions are translated into real-time digital simulations. The review covers a comparative analysis of leading physics engines, an evaluation of numerical integration methods, and an assessment of performance optimization strategies within modern 3D programming environments. The overall findings highlight that effective physics engine simulations require a critical balance between computational accuracy and real-time performance constraints, with significant implications for applications in gaming, engineering simulations, and educational visualization systems.

Quantum computing emerges as a field that captures a great theoretical interest. Its simulation represents a problem with high memory and computational requirements which makes advisable the use of parallel platforms. In this work we deal with the simulation of an ideal quantum computer on the Compute Unified Device Architecture (CUDA), as such a problem can benefit from the high computational capacities of Graphics Processing Units (GPU). After all, modern GPUs are becoming very powerful computational architectures which is causing a growing interest in their application for general purpose. CUDA provides an execution model oriented towards a more general exploitation of the GPU allowing to use it as a massively parallel SIMT (Single-Instruction Multiple-Thread) multiprocessor. A simulator that takes into account memory reference locality issues is proposed, showing that the challenge of achieving a high performance depends strongly on the explicit exploitation of memory hierarchy. Several strategies have been experimentally evaluated obtaining good performance results in comparison with conventional platforms.

A utilização de Unidades de Processamento Gráfico para Propósito Geral (GPGPU) tem crescido muito nos últimos anos. Uma das arquiteturas que se utilizam desse conceito é a arquitetura CUDA da Nvidia, que consegue aumentos significativos de performance computacional repartindo o código em vários “pedaços” (threads) e o processando de forma paralela nas GPUs de uma placa de vídeo. Um de seus limitadores é a forma como o programador reparte o código em threads e como as organiza as mesmas em blocos, para aproveitar ao máximo o processamento das GPUs. Esse trabalho apresenta uma análise de desempenho de um algoritmo de esqueletização de imagens já implementado em CUDA, testando algumas formas de divisão dos pixels da imagem a ser processada por thread a fim de entender melhor o funcionamento da arquitetura CUDA e tentar encontrar um melhor desempenho para algoritmo em questão. O objetivo é gerar um gráfico de desempenho e através dele analisar a eficiência de um algoritmo de esqueletiza...

We present a formal synthesis of recently proposed concepts in recursive prime theory and harmonic analysis with classical principles of number theory and fractal geometry. Building on the frameworks introduced in "The Hidden Simplicity Between 0 and Infinity" and "Recursive Harmonics and QTPTC," we formally define new constructs including recursive primes, quantum-twisted prime triangular compression (QTPTC), and harmonic locks. We show how these constructs integrate with established results such as the Fundamental Theorem of Arithmetic (unique prime factorization) and Euclid's proof of infinitely many primes. Numerical evidence, including benchmarking of standard factorization algorithms, is presented to compare performance and motivate the need for novel approaches. We then explore bidirectional prime symmetries-reading primes forward and backward-and demonstrate how reverse-symmetric fractal patterns emerge from modular arithmetic (e.g. Pascal's triangle mod 2 yielding the Sierpiński triangle). These patterns suggest that the simplicity of primes manifests as recursive compressibility even in complex visual domains. Our results indicate that prime numbers, far from being random, underlie self-similar structures and resonant "locks" in mathematical and physical systems. We discuss theoretical implications of viewing prime factorization as a recursive geometric-harmonic transform, and outline future research directions for applying the QTPTC paradigm to computational number theory and physics. The paper is structured with sections on background, methodology, results (including runtime tables and fractal visualizations), a unifying theoretical discussion, and proposals for future work.

Se analiza la performance, y el grado de aceleración obtenido en relación a la solución secuencial. A modo de referencia, se compara con la paralelización en un cluster de multicores. Como herramientas de programación se utilizaron CUDA en GPU y MPI para la versión en cluster de multicore. Los resultados experimentales muestran una buena respuesta de las soluciones en GPU en relación a las restantes alternativas.

We present an evaluation of the heterogeneous memory system of the Intel Xeon Phi KNL architecture, using applications with di!erent characteristics. Applications that perform many data transfer operations from/to to main memory, when associated with the e"cient use of cache memory, are best candidates to obtain performance gains when mapping data structures to the high bandwidth memory unit . Resumo. Neste artigo é apresentada uma avaliação do sistema de memória heterogênea da arquitetura Intel Xeon Phi KNL, usando aplicações com diferentes características. Aplicações que realizam muitas operações de transferências de dados de e para a memória principal, quando associadas ao uso e#ciente de memória cache, são fortes candidatas a terem ganhos de desempenho ao mapearem estruturas de dados para a unidade de memória de grande largura de banda.

A computação do fecho transitivo de um grafo é um problema que foi considerado pela primeira vez em 1959. Muitos algoritmos sequenciais para solução deste problema foram propostos e algoritmos paralelos foram considerados a partir de 1990. Apresentamos um algoritmo paralelo para computação do fecho transitivo em grafos gerais (orientados ou não). O algoritmo proposto utiliza o modelo BSP/CGM, tendo como base a utilização de busca em largura em grafos (BFS). Apresentamos três implementações paralelas baseadas no algoritmo BFS: uma em MPI, uma em OpenMP e uma híbrida (MPI/OpenMP). Para avaliar a eficiência do algoritmo proposto e das implementações desenvolvidas, foram realizados experimentos com grafos de tamanhos e características variadas. A complexidade computacional do algoritmo é O( n 2 p (n + m)).

A collision involving only two bodies. -A collision scenario where more than two bodies are involved concurrently. -Concurrent collisions sharing at least one common rigid body number in their collision pair lists. -Collisions occurring at exactly the same instant in time. -A template class designed to be able to contain any type of data in a generic manner. -A measure of the similarity between solution values in two successive iterations. -Values obtained by successive iterations are tending towards specific values. -Sets of standardised programming solutions to often encountered problems. -Iteration results not settling on any specific values but rather appearing to change by ever greater values. -A cluster collision involving seven rigid bodies. -A cluster collision involving six rigid bodies. -Repetitive calculation or algorithm execution that stops when a certain termination condition is met. -A three-dimensional array of values or objects. -A two-dimensional array of values or objects. -A function providing a measure of the success or reliability of a solution to an optimisation problem when the solution is substituted into the expression. -A cluster collision involving eight rigid bodies. -A computational problem with the aim of finding an optimal or best solution for a set of requirements. -If a vector is oriented exactly perpendicular to some reference vector, it is said to be orthogonal to that reference vector and it implies that their scalar product (dot product) should be zero. -Profile drag coefficient tensor for body i. -Tangential drag coefficient tensor for body i. -Point i to facet j perpendicular projection distance. -Inter-object distance between geometrical objects i and j. -Infinitesimal volume for rigid body i. -Force vector acting upon body i. -Mass moment of inertia tensor for body i. -Mass moment of inertia tensor entry in row j, column k for body i . -Mass of body i. -Mass tensor for body i. -Surface normal vector for facet i. -General contact normal vector between object i and j (pointing to i). -Contact normal vector between object i and j (pointing to 9 at contact number nb. -Number of evenly spaced bins in Cartesian axis direction k. -Plane normal vector for any arbitrary plane to be intersected by an edge. -Orthonormal tangential vector between object i and j at contact number nb. -Any arbitrary point position vector. -Perpendicularly projected point from point i to edge j. -Perpendicularly projected point from point i to geometric entity j. -Perpendicularly projected point from facet i's vertex k to its opposite facet edge. -Point on any arbitrary plane to be intersected by an edge. -General point on and edge. -Radius of spherical body i. -Radius of spherical body i for contact number nb. -Any radial position relative to a reference point in metres ( m ) -Radial position vector from body i centre of mass ( 5 , ) to contact number nb. -Component i of a general radius relative to a fixed point ( d ) . -General radius relative to the centre of mass for any body. -Position radius of body i relative to the system centre of mass (A, ). -Tangential vector between object i and j at contact number nb. -Torque vector acting upon body i. -Total volume of rigid body i. -Centre of mass position vector for total rigid body system. -Centre of mass position vector for body i. -Position vector for first vertex of edge i. -Position vector for second vertex of edge i. -Position vector for first vertex of facet i. -Position vector for second vertex of facet i. -Position vector for third vertex of facet i. -Position vector for vertex i. -Any general linear velocity in metres per second ( m . s l ) . -Centre of mass (x_) average linear velocity for total rigid body system. -Time derivative of centre of mass position vector for body i. -Velocity of object i relative to object j. -Normal component of velocity of object i relative to object j. -Tangential component of velocity of object i relative to objectj. -Resultant velocity due to linear (&) and angular (8) velocities at given radius ( r ). a,, -Point i to-edge j perpendicular projection ratio. -Point-to-edge projection ratio from point i's perpendicular line k of facet j. -Original sorting bin division size in Cartesian axis direction k. A,,,,+, -Sorting bm division overlap in Cartesian axis direction k. -The time step adjustment to be made during a critical time search iteration process. At,,. -The time step adjustment to be made during a critical time search iteration process.

Dichos métodos consisten en procesar el espectro de las imágenes por un banco de filtros para, a partir de ahí, extraer las características que más información proporcionen para la posterior fase de clasificación. Concretamente, se compararán dos métodos alternativos; uno de ellos ha sido ya ampliamente usado en clasificación de texturas, y sus prestaciones servirán como referencia; el otro, el cual es nuestro principal objeto de estudio, ha sido aplicado satisfactoriamente en la clasificación de géneros musicales, y se pretende su extrapolación para el problema de la clasificación de texturas. Dichos métodos son, respectivamente: • Extracción mediante un banco de filtros de Gabor, el cual es fijo, y está basado en el sistema de reconocimiento del cerebro humano. • Extracción mediante filtros variables, adaptados a la base de datos, obtenidos mediante un método de aprendizaje máquina supervisado denominado POPLS. Una vez establecido el sistema de clasificación para ambos métodos se evaluarán por separado sus prestaciones: tasa de aciertos y matriz de confusión, para determinar la viabilidad del método supervisado. Capítulo 5. Conclusiones y líneas futuras .

In Science a large number of areas are being benefited by the reduction of computational time with the use of Graphics Processing Units (GPU). In the case of Epidemiology through the speeding of the simulation of scenarios with large populations where the processing time is very significant. This article introduces an epidemiological event simulation with a model based on Stochastic Cellular Automata (SCA). This provides an implementation of the main features of a large-scale infectious disease: Contact, Neighborhood, Trajectories and Transmissibility. A case study is simulated on an implementation of the SCA algorithm, for an infectious disease type SEIR (Susceptible, Exposed, Infected and Recovered). In a population with 1,000,000 individuals. This is parallelized through a process balancing algorithm implemented in C-CUDA. The result given by the GPU parallelized software is compared against a parallelized model analysis made by multi-threaded CPU. The results show that the compu...

Helicopters can experience brownout when flying close to a dusty surface. The uplifting of dust in the air can remarkably restrict the pilot’s visibility area. Consequently, a brownout can disorient the pilot and lead to the helicopter collision against the ground. Given its risks, brownout has become a high-priority problem for civil and military operations. Proper helicopter design is thus critical, as it has a strong influence over the shape and density of the cloud of dust that forms when brownout occurs. A way forward to improve aircraft design against brownout is the use of particle simulations. For simulations to be accurate and comparable to the real phenomenon, billions of particles are required. However, using a large number of particles, serial simulations can be slow and too computationally expensive to be performed. In this work, we investigate an message passing interface (MPI) + graphics processing unit (multi-GPU) approach to simulate brownout. In specific, we use a ...

The paper presents an overview of recent research on the Particle Swarm Optimization (PSO) algorithm parallelization on the Graphics Processing Unit for general-purpose computations (GPGPU). This survey attempts to collect, organize, and present reports in the area published since 2007 in a unified way. In order to organize the literature a classification by objective functions and PSO variants is proposed. The paper also compares experimental results taking into account the most popular factor, the calculating acceleration ratio called speedup. Results of the survey are given in a very compact and comprehensive way and could be used as a guide in this area. As a summary, conclusions from categorization, a comparability problem, and possible research areas are discussed.

In the rapidly evolving landscape of artificial intelligence, code generation has emerged as a critical frontier for automating software development, optimizing workflows, and democratizing programming expertise. Among the pioneers in this domain, DeepSeek-a series of large language models (LLMs) with a focus on Mixture-of-Experts (MoE) architecture-has gained global attention for its unparalleled speed, precision, and adaptability. This survey dissects the technological innovations behind DeepSeek's success, exploring how its modular architecture and specialized "experts" enable lightning-fast code generation while maintaining state-of-the-art performance. We contextualize DeepSeek's contributions within the broader AI ecosystem, analyze its benchmarks against competitors, and discuss its implications for the future of software engineering and AI-driven creativity.

The current research paper aims to outline the capabilities of GP-GPU accelerated parallel computing for database operations and compare it with Central Processing Units (CPU)-based methodologies. The study evaluates the performance, scalability, and cost-effectiveness of both processing architectures through the execution of three database operations: None external critical success factors were identified for this system. We found out that GPUs take situational advantages by virtue of their parallelism nature for different computations and also show limitations arising from data transfer overhead, memory availability. The results are further elaborated in detail and recommendations for future work are also provided in order to add to the extant knowledge on GPU-accelerated databases.

Quantum computing emerges as a field that captures a great theoretical interest. Its simulation represents a problem with high memory and computational requirements which makes advisable the use of parallel platforms. In this work we deal with the simulation of an ideal quantum computer on the Compute Unified Device Architecture (CUDA), as such a problem can benefit from the high computational capacities of Graphics Processing Units (GPU). After all, modern GPUs are becoming very powerful computational architectures which is causing a growing interest in their application for general purpose. CUDA provides an execution model oriented towards a more general exploitation of the GPU allowing to use it as a massively parallel SIMT (Single-Instruction Multiple-Thread) multiprocessor. A simulator that takes into account memory reference locality issues is proposed, showing that the challenge of achieving a high performance depends strongly on the explicit exploitation of memory hierarchy. Several strategies have been experimentally evaluated obtaining good performance results in comparison with conventional platforms.

We propose a novel approach for runtime verification on computers with a large number of computation cores, without any hardware extension to mainstream PC environment. The goal of the approach is making use of all hardware resources to decouple the computational overhead of traditional race checkers via parallelizing the runtime verification. We distinguish between two kinds of computational overhead:(i) overhead caused by monitoring, and (ii) overhead due to the verification algorithm (s). So far, ...

This thesis examines the parallelization techniques that can be used to speed up econometric computations. Specifically, Monte Carlo simulations implemented using the R programming language. Parallelized Monte Carlo simulations can use multiple processing elements simultaneously, thereby providing answers to computational problems quicker. Three Monte Carlo experiments were examined; a Monte Carlo estimation of the performance of a Bootstrap confidence interval, a Monte Carlo estimation of a Maximum Likelihood Estimation, and the Metropolis-Hastings Independence Chain algorithm. Each experiment was implemented using multiple parallelization techniques, which were tested using multiple parallelization distribution methods. Analysis of the parallel programs shows that parallel Monte Carlo simulations with coarsegrained subproblems, i.e., relatively large subproblems, perform better compared to programs with fine-grained subproblems. The impact of the granularity of the subproblems also appears to increase as the delay between processing elements increases.

Am zweiten Oktober-Wochenende des Jahres 1913 versammelten sich weit über zweitausend junge Männer und Frauen auf dem Hohen Meißner in Hessen. Das Treffen fand auf Einladung einer losen Vereinigung von Jugendbünden statt und sollte den patriotischen Übersteigerungen des Deutschen Reiches etwas entgegensetzen. Denn dieser Oktober war zugleich Schauplatz offizieller Festakte zum hundertjährigen Jubiläum der Völkerschlacht bei Leipzig, in der die napoleonischen Truppen ihre entscheidende Niederlage erlitten hatten. Russland, Preußen, Österreich und Schweden hatten Frankreich besiegt und dem bislang erfolgreichen Eroberungszug seiner revolutionären Armee unter Napoleon Bonaparte damit ein Ende gesetzt. Anlässlich des feierlichen Rückblicks auf dieses historische Ereignis sollte durch den Deutschen Kaiser und die Bundesfürsten ein monumentales Denkmal eingeweiht werden, zur ewigen Erinnerung an die große Befreiungsschlacht des europäischen Nordens gegen den europäischen Süden, zumal eine Auseinandersetzung feindlicher Lager jederzeit erneut bevorstehen konnte. Denn die Zeremonie der Einweihung stand bereits im Zeichen eines Krieges, der als erster von zwei Weltkriegen in die Geschichte eingehen sollte. Zu den jugendlichen Gegnern der Reichspolitik gehörten die Vereinigungen der Wandervögel und Lebensreformer, die sich für einen fundamentalen politischen Wandel einsetzten. Ihr Ziel war es, die Geschlossenheit einer Jugendbewegung zu demonstrieren, die sich von den alten Feindschaften aus der Gründerzeit losgesagt und diese überwunden hatte. Die Alternativveranstaltung zur kaiserlichen Jubiläumsfeier war daher als ein "Fest der Jugend" gedacht, mit Verweis auf den politischen Aufbruch der frühen Wartburgfeste im neunzehnten Jahrhundert, die ebenfalls Protestkundgebungen einer emanzipierten, gegen reaktionäre Politik und Kleinstaaterei aufbegehrenden Jugend gewesen waren. Zwar fand auch das "Fest der Jugend" anlässlich des Gedenkens an den militärischen Erfolg von 1813 statt, aber der Sieg wurde von den Jugendbünden nicht nur als Befreiung von den französischen Besatzern begriffen, sondern auch als Befreiung von der deutschen Kleinstaaterei und den alten Autoritäten. Es bezeichnet den demokratischen Auftakt einer jungen Nation, die sich erst noch zu finden hatte. Mit den beiden konkurrierenden Veranstaltungen im Oktober 1913 stand die deutsche Nation sich in ihren entscheidenden Grundzügen selbst gegenüber.

The on-line event reconstruction in ALICE is performed by the High Level Trigger, which should process up to 2000 events per second in proton-proton collisions and up to 300 central events per second in heavy-ion collisions, corresponding to an input data stream of 30 GB/s. In order to fulfill the time requirements, a fast on-line tracker has been developed. The algorithm combines a Cellular Automaton method being used for a fast pattern recognition and the Kalman Filter method for fitting of found trajectories and for the final track selection. The tracker was adapted to run on Graphics Processing Units (GPU) using the NVIDIA Compute Unified Device Architecture (CUDA) framework. The implementation of the algorithm had to be adjusted at many points to allow for an efficient usage of the graphics cards. In particular, achieving a good overall workload for many processor cores, efficient transfer to and from the GPU, as well as optimized utilization of the different memories the GPU offers turned out to be critical. To cope with these problems a dynamic scheduler was introduced, which redistributes the workload among the processor cores. Additionally a pipeline was implemented so that the tracking on the GPU, the initialization and the output processed by the CPU, as well as the DMA transfer

The polyhedral model is powerful for analyzing and transforming static control programs, hence its intensive use for the optimization of data locality and automatic parallelization. Affine transformations excel at modeling control flow, to promote data reuse and to expose parallelism. The approach has also successfully been applied to the optimization of memory accesses (array expansion and contraction), although the available tools in the area are not as mature. Yet data locality also depends on other parameters such as data layout and data placement relatively to the memory hierarchy; these include spatial locality in cache lines and scalability on NUMA systems. This paper presents Ivie, a parallel intermediate language which complements affine transformations implemented in state-of-the-art polyhedral compilers and supports spatial and NUMA-aware data locality optimizations. We validate the design of the intermediate language on representative benchmarks.

In this age where data is growing at an astronomical rate, with unfettered access to digital information, complexities have been introduced to scientific computations, analysis, and inferences. This is because such data could not be easily processed with traditional approaches. However, with innovative designs brought to the fore by NVIDIA and other market players in recent times, there have been productions of state-of-the-art GPUs such as NVIDIA A100 Tensor Core GPU, Tesla V100, and NVIDIA H100 that seamlessly handle complex mathematical simulations and computations, artificial intelligence, machine learning, and high-performance computing, producing highly improved speed and effi
ciency, with room for scalability. These innovations have made it possible to efficiently deploy many parallel programming models like shared memory, distributed memory, data parallelization, and Partitioned Global Address Space (PGAS) with high-performance metrics. In this work, we analyzed the parquet-formatted New York City yellow taxi dataset
on a RAPIDS and DASK supported distributed data-parallel training platform using a high-performance cluster of 7 NVIDIA TITAN RTX GPUs (24GB GDDR6 each) running CUDA 12.4. The dataset was used to train Extreme Gradient Boosting (XGBoost), RandomForest Regressor, and Elastic Net models for trip fare predictions. Our models achieved notable performance metrics. The XGBoost achieved a mean squared error (MSE) of 10.87, R2 of 96.9%, and a training time of 21.1 seconds despite the huge size of the training dataset, showing how computationally efficient the system was. RandomForest achieved MSE of 27.46, R2 of 92.2% and a training time of 25.9 seconds.  In the bid to show the scalability and versatility of our experimental design to different machine learning domains, our multi-GPU accelerated training was extended to image classification tasks by using MobileNet-V3-Large pre-trained architecture on a CIFAR-100 dataset. The following parallelization results were achieved: a low Karp-Flatt metric of 0.013, indicating minimal serialization, 98.7% parallel fraction, demonstrating excellent parallelization, and only 7.1% communication overhead relative to computation time.  For the model performance, we achieved a ROC AUC of over 95% for the implementation. This work advances the state-of-the-art in parallel computing through implementation of RAPIDS and DASK frameworks on a distributed data-parallel training platform making use of NVIDIA multi-GPUs. The work is built on a well established theoretical framework using Amdahl and Gustafson’s laws on parallel computation. By integrating RAPIDS and DASK, we contribute to advancing parallel computing capabilities, offering potential applications in smart city development and the field of logistics and transportation management services where rapid fare predictions are very important. The contribution could also be xtended
to the field of image classification, vision systems, object detection and embedded systems for mobile applications.

Due to the limited capacity of GPU memory, the majority of prior work on graph applications on GPUs has been restricted to graphs of modest sizes that fit in memory. Recent hardware and software advances make it possible to address much larger host memory transparently as a part of a feature known as unified virtual memory. While accessing host memory over an interconnect is understandably slower, the problem space has not been sufficiently explored in the context of a challenging workload with low computational intensity and an irregular data access pattern such as graph traversal. We analyse the performance of breadth first search (BFS) for several large graphs in the context of unified memory and identify the key factors that contribute to slowdowns. Next, we propose a lightweight offline graph reordering algorithm, HALO (Harmonic Locality Ordering), that can be used as a pre-processing step for static graphs. HALO yields speedups of 1.5x-1.9x over baseline in subsequent traversa...

Due to the limited capacity of GPU memory, the majority of prior work on graph applications on GPUs has been restricted to graphs of modest sizes that fit in memory. Recent hardware and software advances make it possible to address much larger host memory transparently as a part of a feature known as unified virtual memory. While accessing host memory over an interconnect is understandably slower, the problem space has not been sufficiently explored in the context of a challenging workload with low computational intensity and an irregular data access pattern such as graph traversal. We analyse the performance of breadth first search (BFS) for several large graphs in the context of unified memory and identify the key factors that contribute to slowdowns. Next, we propose a lightweight offline graph reordering algorithm, HALO (Harmonic Locality Ordering), that can be used as a pre-processing step for static graphs. HALO yields speedups of 1.5x-1.9x over baseline in subsequent traversa...

In this paper, we develop a novel kind of energy-dissipation splitting finite-difference timedomain scheme for solving two-dimensional Maxwell equations with perfectly matched layers. The discrete energy dissipation law, convergence, dispersion relation, and stability are investigated for the scheme. Theoretical analysis shows that the scheme is unconditionally stable, and of second order accuracy both in time and space. Numerical experiments confirm the theoretical results.

Engineering is many times described as making practical application of the knowledge of pure sciences in the solution of a problem or the application of scientific and mathematical principles to develop economical solutions to technical problems, creating products, facilities, and structures that are useful to people. What if the opposite occurs? There is some product that may be a solution to some problem but the inner workings of the solution or even the problem it addresses may be unknown. Reverse engineering is the process of analyzing and understanding a product which functioning and purpose are unknown. In Computer Science in particular, reverse engineering may be defined as the process of analyzing a system's code, documentation, and behavior to identify its current components and their dependencies to extract and create system abstractions and design information. The subject system is not altered; however, additional knowledge about the system is produced. The definition...

The paper discusses digital filter bank implementation using the weighted overlap-add (WOLA) algorithm modification introduced in the context of multichannel signal processing. The suggested modification is applied to hydroacoustic monitoring tasks. High-performance softwarehardware implementations on the basis of CPU (Central Processing Unit) and CUDA (Compute Unified Device Architecture) are compared and CUDA is shown to provide advantages for the filter banks studied. Streszczenie. W artykule przedstawiono implementację banku filtrów cyfrowych opartych na algorytmie Weighted Overlap-Add (WOLA), zmodyfikowanego w kontekście przetwarzania sygnału wielokanałowego. Zaproponowane modyfikacje zostały zastosowane w hydroakustycznych zadaniach monitorowania. Dokonano porównania wysokowydajnych programowo-sprzętowych implementacji opartych na CPU (Central Processing Unit) oraz CUDA (Compute i Unified Device Architecture). Badania wykonane dla zdefiniowanego banku filtrów pokazały, że najbardziej korzystne jest zastosowanie CUDA. (Implementacja banku filtrów w hydroakustycznych zadaniach monitorowania).

Real-time depth extraction from stereo images is an important process in computer vision. This paper proposes a new implementation of the dynamic programming algorithm to calculate dense depth maps using the CUDA architecture achieving real-time performance with consumer graphics cards. We compare the running time of the algorithm against CPU implementation and demonstrate the scalability property of the algorithm by testing it on different graphics cards.

Este trabajo tiene como objetivo presentar las principales características del diseño e implementación de algoritmos paralelos para el procesamiento de señales de radar de apertura sintétitca (SAR). Se analizan las razones por las que el problema es paralelizable y se presentan implementaciones de algoritmos clásicos para la obtención de imágenes. También se presentan adaptaciones de estos algoritmos para ser utilizados sobre arquitecturas paralelas de tipo GPU de propósito general. Estas modificaciones son propuestas con el objetivo de obtener algoritmos altamente eficientes y escalables en este tipo de arquitecturas.