A Comparative Study of CUDA and Vulkan for Real-Time Motion Analysis : Benchmarking GPU execution time and scaling in a motion-analysis pipeline

Sammanfattning

This Master’s thesis compares CUDA and Vulkan compute shaders for a three-stage real-time motion-detection pipeline and includes a qualitative MoQ integration case study in a real-world streaming setting. The motion detection pipeline consists of ViBe background subtraction, binary morphology, and connected component labeling. Two implementations were developed with identical algorithms, deterministic random updates, and shared input data to ensure fairness. Performance is evaluated on an NVIDIA RTX A4000 using per-stage GPU timing, total measured GPU stage time distributions, and tail timing behavior, supplemented by a roofline-light proxy based on estimated bytes moved per stage. Results show that Vulkan is faster at low resolution, but CUDA scales better as resolution increases. The analysis indicates that performance differences grow most in stages dominated by memory pressure and synchronization/atomic behavior, suggesting that Vulkan’s overheads and synchronization sensitivity become increasingly visible at higher resolutions. In addition, a MoQ integration was implemented to study practical tradeoffs in system integration in a live-streaming setting. This part of the study was qualitative rather than a second performance benchmark. The main trade-offs were reusing an existing mediapipeline and the complexity of working in a large codebase, keeping frame data close to the GPU, and the additional effort required to access the existing CUDA context, and smoothing the visualization of the motion metadata with the video playback at the subscriber. These observations show that integrating motion detection processing into a live MoQ workflow is achievable, but that practical deployment introduces synchronization, interoperability, and metadata-rendering challenges beyond the isolated GPU benchmark.