Comparative Performance Analysis of Single Precision Parallel Processing for Biomedical Images on Multi-Generation Gpu Architectures

Introduction. This paper evaluates the performance of single-precision parallel computing for biomedical MRI image slice processing across multi-generation GPU architectures within the MATLAB Parallel Computing Toolbox environment. Materials and Methods. Comparative benchmark tests were conducted using host CPU execution (Intel Core i7-3770) alongside entry-level NVIDIA GPU architectures representing different hardware generations, specifically Kepler (GeForce GT 730) and Pascal (GeForce GT 1030). Evaluation scenarios included spatial-domain volumetric batching across varying dataset sizes and isolated two-dimensional Fast Fourier Transforms. Results. The empirical results reveal a pronounced CUDA context initialization overhead for small datasets (N < 25 slices), temporarily reducing GPU performance relative to the CPU. However, spatial-domain volumetric batching mitigates driver setup latencies, achieving a maximum throughput of 275.86 frames per second (3.30x speedup) on the Pascal GT 1030 and 154.92 frames per second (1.85x speedup) on the Kepler GT 730 over host CPU execution (83.65 frames per second). Direct hardware comparison confirms a stable 1.78x performance advantage for Pascal over Kepler. Conclusions. Volumetric batching successfully mitigates memory transfer overheads to achieve meaningful GPU acceleration on low-power hardware setups. Based on the measured execution throughputs, an asymmetric load-balancing model (1.78:1 work-distribution ratio) is formulated to optimize heterogeneous multi-GPU execution on budget-tier hardware.