GPU Features Research Articles

Edge and cloud computing approaches in the early diagnosis of skin cancer with attention-based vision transformer through hyperspectral imaging

Hyperspectral imaging is applied in the medical field for automated diagnosis of diseases, especially cancer. Among the various classification algorithms, the most suitable ones are machine and deep learning techniques. In particular, Vision Transformers represent an innovative deep architecture to classify skin cancers through hyperspectral images. However, such methodologies are computationally intensive, requiring parallel solutions to ensure fast classification. In this paper, a parallel Vision Transformer is evaluated exploiting technologies in the context of Edge and Cloud Computing, envisioning portable instruments’ development through the analysis of significant parameters, like processing times, power consumption and communication latency, where applicable. A low-power GPU, different models of desktop GPUs and a GPU for scientific computing were used. Cloud solutions show lower processing times, while Edge boards based on GPU feature the lowest energy consumption, thus resulting as the optimal choice regarding portable instrumentation with no compelling time constraints.

An interactive triangle-based ODF glyph rendering for high angular resolution diffusion imaging

Diffusion magnetic resonance imaging quantifies the diffusion of water molecules. It is unique in showing white matter fiber structures and connectivity in-vivo, relying on the fact that the water molecules’ displacements are more significant along with the fiber orientation than in other directions due to biological barriers. High angular resolution diffusion imaging (HARDI) acquisition enables the differentiation of multiple fiber signals passing through a voxel. Advanced methods for processing diffusion data were developed to synthesize the HARDI data of each voxel into an orientation distribution function (ODF) that relates each direction sample to its probability of water diffusion. Visualizing ODFs through glyphs helps gain insight into the local structure of the brain white matter. Because of the high dimensionality of the data, interactive glyph exploration remains a challenge for a volume containing millions of voxels. This work presents a 3D triangle-based multi-resolution rendering scheme for the ODF glyph. We explored data properties and GPU features to make the triangle-based rendering interactive. Linear performance relative to some fixed-resolution rendered glyphs was assessed using NVIDIA® Nsight Visual StudioTM. A control volume, available in the Retest dataset of the WU-Minn Human Connectome Project, was used to assess the effectiveness of interactive 3D visualization over a multiplanar and curvilinear reformatted volume of ODF glyphs in voxels with multiple fibers, namely the triple crossing at the centrum semiovale, the fanning at the corona radiata, and the crossing of the cingulum and corpus callosum fiber bundles.

PaScaL_TDMA 2.0: A multi-GPU-based library for solving massive tridiagonal systems

PaScaL_TDMA 2.0: A multi-GPU-based library for solving massive tridiagonal systems

Experimental implementation of a neural network optical channel equalizer in restricted hardware using pruning and quantization

The deployment of artificial neural networks-based optical channel equalizers on edge-computing devices is critically important for the next generation of optical communication systems. However, this is still a highly challenging problem, mainly due to the computational complexity of the artificial neural networks (NNs) required for the efficient equalization of nonlinear optical channels with large dispersion-induced memory. To implement the NN-based optical channel equalizer in hardware, a substantial complexity reduction is needed, while we have to keep an acceptable performance level of the simplified NN model. In this work, we address the complexity reduction problem by applying pruning and quantization techniques to an NN-based optical channel equalizer. We use an exemplary NN architecture, the multi-layer perceptron (MLP), to mitigate the impairments for 30 GBd 1000 km transmission over a standard single-mode fiber, and demonstrate that it is feasible to reduce the equalizer’s memory by up to 87.12%, and its complexity by up to 78.34%, without noticeable performance degradation. In addition to this, we accurately define the computational complexity of a compressed NN-based equalizer in the digital signal processing (DSP) sense. Further, we examine the impact of using hardware with different CPU and GPU features on the power consumption and latency for the compressed equalizer. We also verify the developed technique experimentally, by implementing the reduced NN equalizer on two standard edge-computing hardware units: Raspberry Pi 4 and Nvidia Jetson Nano, which are used to process the data generated via simulating the signal’s propagation down the optical-fiber system.

Accelerating Smith-Waterman Algorithm for Faster Sequence Alignment using Graphical Processing Unit

In bioinformatics, sequence alignment is very important task to compare and find similarity between biological sequences. Smith Waterman algorithm is most widely used for alignment process but it has quadratic time complexity. This algorithm is using sequential approach so if the no. of biological sequences is increasing then it takes too much time to align sequences. In this paper, parallel approach of Smith Waterman algorithm is proposed and implemented according to the architecture of graphic processing unit using CUDA in which features of GPU is combined with CPU in such a way that alignment process is three times faster than sequential implementation of Smith Waterman algorithm and helps in accelerating the performance of sequence alignment using GPU. This paper describes the parallel implementation of sequence alignment using GPU and this intra-task parallelization strategy reduces the execution time. The results show significant runtime savings on GPU.

Large-scale terrain-adaptive LOD control based on GPU tessellation

The real-time rendering of large-scale terrain scenes has been a challenge and is a popular research topic in the fields of 3D games and virtual reality. To fully utilize GPU features, this paper proposes a real-time rendering algorithm for large-scale regular grid terrain based on GPU Tessellation. The CPU processes a coarse-grained quadtree to generate terrain patches and then passes the patches to the GPU after view culling. Then, the GPU refines the terrain by using GPU Tessellation and displacement mapping. The most important advantage of our method is that during the tessellation stage, a new rule that the distance, screen space projection error and variance of the height, as the standard of terrain roughness, is designed. During the tessellation stage, the distance, screen space projection error and variance of the height, as the standard of terrain roughness, are considered. The experimental results show that our algorithm can greatly reduce the CPU processing time and main memory space requirement, and fully utilize the new features of the GPU; In conclusion, the proposed approach can render large-scale terrain with high-precision in real-time.

Performance analysis of CUDA, OpenACC and OpenMP programming models on TESLA V100 GPU

Graphics processors are widely utilized in modern supercomputers as accelerators. Ability to perform efficient parallelization and low-level allow scientists to greatly boost performance of their codes. Modern Nvidia GPUs feature low-level approaches, such as CUDA, along with high-level approaches: OpenACC and OpenMP. While the low-level approach aims to explore all possible abilities of SIMT GPU architecture by writing low-level C/C++ code, it takes significant effort from programmer. OpenACC and OpenMP programming models are opposite to CUDA. Using these models the programmer only have to identify the blocks of code to be parallelized using pragmas. We compare the performance of CUDA, OpenMP and OpenACC on state-of-the-art Nvidia Tesla V100 GPU in various typical scenarios that arise in scientific programming, such as matrix multiplication, regular memory access patterns and evaluate performance of physical simulation codes implemented using these programming models. Moreover, we study the performance matrix multiplication implemented in vendor-optimized BLAS libraries for Nvidia Tesla V100 GPU and modern Intel Xeon processor.

Side-Channel Attacks With Multi-Thread Mixed Leakage

Side-channel attacks are one of the greatest practical threats to security-related applications, because they are capable of breaking ciphers that are assumed to be mathematically secure. Lots of studies have been devoted to power or electro-magnetic (EM) analysis against desktop CPUs, mobile CPUs (including ARM, MSP, AVR, etc) and FPGAs, but rarely targeted modern GPUs. Modern GPUs feature their special and specific single instruction multiple threads (SIMT) execution fashion, which makes their power/EM leakage more sophisticated in practical scenarios. In this article, we study side-channel attacks with leakage from SIMT systems, and propose leakage models suited to any SIMT systems and specifically to CUDA-enabled GPUs. Afterwards, we instantiate the models with a GPU AES implementation, which is also used for performance evaluations. In addition to the models, we provide optimizations on the attacks that are based on the models. To evaluate the models and optimizations, we run the GPU AES implementation on a CUDA-enabled GPU and, at the same time, collect its EM leakage. The experimental results show that the proposed models are more efficient and the optimizations are effective as well. Our study suggests that GPU-based cryptographic implementations may be much vulnerable to microarchitecture-based side-channel attacks. Therefore, GPU-specific countermeasures should be considered for GPU-based cryptographic implementations in practical applications.

G 3

This paper demonstrates G 3 , a framework for <u>G</u>raph Neural Network (GNN) training, tailored from <u>G</u>raph processing systems on <u>G</u>raphics processing units (GPUs). G 3 aims at improving the efficiency of GNN training by supporting graph-structured operations using parallel graph processing systems. G 3 enables users to leverage the massive parallelism and other architectural features of GPUs in the following two ways: building GNN layers by writing sequential C/C++ code with a set of flexible APIs (Application Programming Interfaces); creating GNN models with essential GNN operations and layers provided in G 3 . The runtime system of G 3 automatically executes the user-defined GNNs on the GPU, with a series of graph-centric optimizations enabled. We demonstrate the steps of developing some popular GNN models with G 3 , and the superior performance of G 3 against existing GNN training systems, i.e., PyTorch and TensorFlow.

A parallel sliding-window belief propagation algorithm for Q-ary LDPC codes accelerated by GPU

In this paper, a parallel Sliding-Window Belief Propagation algorithm to decode Q-ary Low-Density-Parity-Codes is proposed. This algorithm is accelerated by taking advantage of high parallel features of GPU, and applied to video compression under distributed video coding framework. The experiment results show that our parallel algorithm achieves 2.3× to 30.3× speedup ratio under 256 to 2048 codeword length and 69.21× to 78.31× speedup ratio under 16,384 codeword length than sequential algorithm.

CD-Xbar: A Converge-Diverge Crossbar Network for High-Performance GPUs

Modern GPUs feature an increasing number of streaming multiprocessors (SMs) to boost system throughput. How to construct an efficient and scalable network-on-chip (NoC) for future high-performance GPUs is particularly critical. Although a mesh network is a widely used NoC topology in manycore CPUs for scalability and simplicity reasons, it is ill-suited to GPUs because of the many-to-few-to-many traffic pattern observed in GPU-compute workloads. Although a crossbar NoC is a natural fit, it does not scale to large SM counts while operating at high frequency. In this paper, we propose the converge-diverge crossbar (CD-Xbar) network with round-robin routing and topology-aware concurrent thread array (CTA) scheduling. CD-Xbar consists of two types of crossbars, a local crossbar and a global crossbar. A local crossbar converges input ports from the SMs into so-called converged ports; the global crossbar diverges these converged ports to the last-level cache (LLC) slices and memory controllers. CD-Xbar provides routing path diversity through the converged ports. Round-robin routing and topology-aware CTA scheduling balance network traffic among the converged ports within a local crossbar and across crossbars, respectively. Compared to a mesh with the same bisection bandwidth, CD-Xbar reduces NoC active silicon area and power consumption by 52.5 and 48.5 percent, respectively, while at the same time improving performance by 13.9 percent on average. CD-Xbar performs within 2.9 percent of an idealized fully-connected crossbar. We further demonstrate CD-Xbar's scalability, flexibility and improved performance per Watt (by 17.1 percent) over state-of-the-art GPU NoCs which are highly customized and non-scalable.

Visually Lossless Content and Motion Adaptive Shading in Games

We present a technique that adaptively adjusts the shading rate during rendering based on the scene content and motion. Our goal is to improve performance with no loss in perceived quality. We determine per-screen-tile shading rate by testing an error estimate against a perceptually-corrected just-noticeable difference threshold. Our design features an effective and efficient error estimate using spatial and frequency analysis of half and quarter rate shading. We also study the effect of motion in reducing perceived error, a consequence of display-persistence and/or motion blur effects. Our implementation uses the computed per-tile shading rate with variable rate shading (a recent GPU feature) to lower shading cost. We demonstrate the quality and performance of our technique on two high-end game engines and shipped games.

Memory system optimization for graph processing: a survey

Containing a variety of information, a graph is a complex data structure comprising vertices and edges. Graph processing or graph computing is the abstraction of the relation and properties of graph structures in real-life situations and performing some complex computations. Owing to the bottlenecks in the performance of the central processing unit, many coprocessors and domain-specific accelerators have been designed to improve running speed and to save energy. Considering the strong data dependence of graphs, improving the efficiency of memory access is a critical issue to improve system performance. In particular, memory management and optimization have become extremely important due to the expansion of graph data scale and the acceleration of various graph processing. This study aims to propose a memory architecture and management methods in graph processing on heterogeneous architecture. We describe the graph data layout that can improve the efficiency of memory access, analyze recent work on memory optimization and features of GPU, FPGA, ASIC, PIM, among others. Furthermore, we summarize relevant research progresses in recent years in China, and conclude the opportunities and challenges of graph processing in memory.

Comparative analysis of 3D RTM Implementation Strategies for an Efficient Use of Memory in a Single GPU:

Reverse-Time Migration (RTM) is a two-way wave-equation based method used to generate images of the Earth’s subsurface. RTM has been successfully used in seismic imaging as it allows defining complex structural areas. However, RTM is a highly computational expensive algorithm that requires the computation of both the source and the receiver wavefields for each shot. Fortunately, numerical methods that compute the wave propagation using the wave equation are highly parallelizable, so they can take leverage on GPU features. However, the main problem of a GPU-RTM implementation is memory management. To take advantage of the GPU computing capabilities, the transfers to host RAM memory storage, or more expensive hard disk storage must be avoided. We present the analysis of three different strategies to implement RTM using only the memory available on a single GPU: (1) Stored wavefield checkpointing (2) Backpropagation of source wavefield using stored boundaries, and (3) Backpropagation of source wavefield using the two last snapshots and random boundaries, showing that the large amount of memory required in the first two strategies becomes a restriction over the model size. The last method (using random boundary conditions) is shown as a suggested solution to the memory problem of using a single GPU.

Enhanced multicore–manycore interaction in high-performance video encoding

This paper presents an efficient cooperative interaction between multicore (CPU) and manycore (GPU) resources in the design of a high-performance video encoder. The proposed technique, applied to the well-established and highly optimized VP8 encoding format, can achieve a significant speed-up with respect to the mostly optimized software encoder (up to $$\times$$ 6), with minimum degradation of the visual quality and low processing latency. This result has been obtained through a highly optimized CPU–GPU interaction, the exploitation of specific GPU features, and a modified search algorithm specifically adapted to the GPU execution model. Several experimental results are reported and discussed, confirming the effectiveness of the proposed technique. The presented approach, though implemented for the VP8 standard, is of general interest, as it could be applied to any other video encoding scheme.

Heterogeneous and unconventional cluster architectures and applications

Heterogeneous and unconventional cluster architectures and applications

Design considerations for GPU‐aware collective communications in MPI

SummaryGPU accelerators have established themselves in the state‐of‐the‐art clusters by offering high performance and energy efficiency. In such systems, efficient inter‐process GPU communication is of paramount importance to application performance. This paper investigates various algorithms in conjunction with the latest GPU features to improve GPU collective operations. First, we propose a GPU Shared Buffer‐aware (GSB) algorithm and a Binomial Tree Based (BTB) algorithm for GPU collectives on single‐GPU nodes. We then propose a hierarchical framework for clusters with multi‐GPU nodes. By studying various combinations of algorithms, we highlight the importance of choosing the right algorithm within each level. The evaluation of our framework on MPI_Allreduce shows promising performance results for large message sizes.To address the shortcoming for small and medium messages, we present the benefit of using the Hyper‐Q feature and the MPS service in jointly using CUDA IPC and host‐staged copy types to perform multiple inter‐process communications. However, we argue that efficient designs are still required to further harness this potential. Accordingly, we propose a static and a dynamic algorithm for MPI_Allgather and MPI_Allreduce and present their effectiveness on various message sizes. Our profiling results indicate that the achieved performance is mainly rooted in overlapping different copy types.

A multi-factor monitoring fault tolerance model based on a GPU cluster for big data processing

High-performance computing clusters are widely used in large-scale data mining applications, and have higher requirements for persistence, stability and real-time use and sre therefore computationally intensive. To support large-scale data processing, we design a multi-factor real-time monitoring fault tolerance (MRMFT) model based on a GPU cluster. However, the higher clock frequency of GPU chips results in excessively high energy consumption in computing systems. Moreover, the ability to support a long-lasting high temperature operation varies greatly between different GPUs owing to the individual differences between the chips. In this paper, we design a GPU cluster energy consumption monitoring system based on wireless sensor networks (WSNs) and propose an energy consumption aware checkpointing (ECAC) for high energy consumption problems with the following two advantages: the system sets checkpoints according to actual energy consumption and the device temperature to improve the utilization of checkpoints and reduce time cost; and it exploits the parallel computing features of CPU and GPU to hide the CPU detection overhead in GPU parallel computation, and further reduce the time and energy consumption overhead in the fault tolerance phase. Using ECAC as the constraint and aiming for a persistent and reliable operation, the dynamic task migration mechanism is designed, and the reliability of the cluster is greatly improved. The theoretical analysis and experiment results show that the model improves the persistence and stability of the computing system while reducing checkpoint overhead.

Optimizing Data Placement on GPU Memory: A Portable Approach

Modern GPUs feature complex memory system designs. One GPU may contain many types of memory of different properties. The best way to place data in memory is sensitive to many factors (e.g., program inputs, architectures), making portable optimizations of GPU data placement a difficult challenge. PORPLE is a recently proposed method that overcomes the difficulties by enabling online optimizations of data placement through a three-way synergy: a specification language for memory system description, a compiler framework for data access analysis and code staging, and a runtime library for efficiently finding and materializing data placement on the fly. This article provides a comprehensive description of this method, and presents several extensions that significantly improve the scalability of PORPLE, which include a novel algorithm design for efficiently searching for the best data placements, the use of active profiling for reducing the online-profiling overhead, and a systematic examination of a path-based performance model. By automatically tailoring data placements for each execution of a GPU program, the enhanced PORPLE brings significant speedups (1.72X on average) to many GPU kernels across GPU architectures and program inputs.

Fast 3D elastic micro-seismic source location using new GPU features

In this paper, we describe new GPU features and their applications in passive seismic – micro-seismic location. Locating micro-seismic events is quite important in seismic exploration, especially when searching for unconventional oil and gas resources. Different from the traditional ray-based methods, the wave equation method, such as the method we use in our paper, has a remarkable advantage in adapting to low signal-to-noise ratio conditions and does not need a person to select the data. However, because it has a conspicuous deficiency due to its computation cost, these methods are not widely used in industrial fields. To make the method useful, we implement imaging-like wave equation micro-seismic location in a 3D elastic media and use GPU to accelerate our algorithm. We also introduce some new GPU features into the implementation to solve the data transfer and GPU utilization problems. Numerical and field data experiments show that our method can achieve a more than 30% performance improvement in GPU implementation just by using these new features.