3D models are widely used in computer graphics, computer vision, and robotics applications. Multiple hardware accelerators are used for running 3D model related applications, since the computations required for models in 3D space are an order of magnitude higher than the computations in 2D space. Due to the high computation intensity of 3D model workloads, using large 3D model datasets for performance characterization is not a feasible choice during accelerator design. Representative subsets are widely used to save the execution or simulation time, e.g, ModelNet10, a subset of ModelNet40, is widely used in the machine learning (ML) domain to save training and inference time. However, this subset is picked by programmers from a software and application perspective.
In this paper, we deploy statistical analysis based methodologies to guide the identification of hardware-aware representative subsets, which can maintain higher performance accuracy and achieve larger execution time savings with respect to subsets picked by software programmers. We believe that the methodology proposed in this paper can help hardware architects and engineers design efficient graphics or ML accelerators rapidly.