GPU-accelerated cloud computing services and performance evaluation

This paper explores the feasibility of replacing traditional CPU-based cloud computing with Graphic Processing Unit GPU-accelerated services. Using NVIDIA’s CUDA GPU-accelerated C/C++ and Python libraries, we benchmark the performance of GPU computing against multithreaded CPU computing across several domains, including machine learning and large-scale image processing. A novel contribution of this work is an intelligent autoscaling system that maximizes single-GPU resource utilization before scaling to additional GPUs, improving efficiency in cloud-based deployments. Our simulation experiments demonstrate significant performance gains for GPU-accelerated computing and highlight the impact of optimized resource allocation in cloud environments. For example, in a machine learning experiment, using a dataset with 8.790 entries, the execution of a GeForce 3060 ti GPU is 3.42 times faster than a 16-thread CPU computer. Compared with the same 16-thread CPU, Tesla K80 GPU is 4.17 times faster. Furthermore, we provide an analysis of GPU performance optimization strategies, including memory management, concurrency techniques, and workload distribution methodologies, offering insights into the long-term scalability and cost-effectiveness of GPU-accelerated cloud infrastructure.