Comparative Analysis of Video Decoding Algorithms: CUDA vs.Vulkan Compute Shaders on Embedded Platforms : Assessing GPU Compute Frameworks for Embedded Video Decoding

GPU acceleration is essential for meeting the performance demands of modern video processing, particularly on embedded systems. CUDA is a widely adopted framework offering mature tooling and high performance on NVIDIA hardware, but its proprietary nature restricts portability. In contrast, Vulkan provides a cross-platform, low-level API with explicit control over GPU resources, making it a promising candidate for portable, high-performance computing. To evaluate their relative strengths, both in terms of performance and code complexity, a JPEG-based video decoder originally implemented in CUDA was ported to Vulkan and executed on the NVIDIA Jetson AGX Orin platform. Performance measurements were based on two computationally intensive segments of the decoding pipeline, each composed of multiple algorithmic stages. Each segment was benchmarked across four resolutions: 480p, 720p, 1080p, and 1440p. Code complexity was evaluated using established software metrics, including Lines of Code, Cyclomatic Complexity, Halstead Volume, and the Maintainability Index. Vulkan consistently outperformed CUDA in computation time across all tested resolutions, with the performance advantage increasing at higher resolutions and in more memory-intensive stages. In addition to faster execution, Vulkan exhibited more stable timing behaviour, indicating greater runtime consistency. However, this performance and predictability came at the cost of slightly increased code verbosity, reduced maintainability, and a steeper development curve, owing to Vulkan’s explicit design philosophy and less mature tooling ecosystem. These findings highlight a clear trade-off between performance and development effort, offering guidance for developers choosing between CUDA and Vulkan for GPU-accelerated video decoding on embedded platforms.