Methods of consumer gpu virtualization for cloud services: comparative analysis of performance and latency

This paper investigates methods of consumer graphics processor (GPU) virtualization for cloud services applications. A comparative analysis of GPU passthrough, SR-IOV, MIG, and time-sliced vGPU technologies is conducted. A methodology for evaluating virtualized GPU performance based on computational efficiency and latency metrics is presented. An experimental comparison of the consumer NVIDIA RTX 2060 12GB graphics card (Turing architecture, TU106 chip) with professional Quadro RTX solutions on identical silicon is performed. It is shown that virtualization overhead amounts to 5–10 % for graphics workloads and reaches 6–8 % for neural network training. The impact of virtualization on end-to-end latency in cloud gaming, VR/AR rendering, and machine learning inference scenarios is investigated. The economic efficiency of using consumer GPUs with modified software in laboratory and educational environments is substantiated. The research results can be used in designing virtual desktop infrastructure and cloud computing clusters.