While sparse AI accelerators promise significant performance gains, the impact of unpredictable activation sparsity on the performance variation and system uncertainty remains largely overlooked and so far has never been quantitatively analyzed. This gap is particularly problematic for time-sensitive applications where predictable performance is essential.
This seminar presents SunPar, an open-source modeling framework and the first analytical simulator designed to evaluate performance variation in sparse AI accelerators. SunPar introduces a new taxonomy that categorizes sparsity uncertainty into two types: tile-level intra-tensor uncertainty (IntraU) and image-level inter-tensor uncertainty (InterU). Using this classification, the framework derives probability density functions that characterize hardware performance under varying sparsity patterns. Based on the framework, two case studies were conducted to show how SunPar identifies Pareto-optimal hardware configurations for real-world applications, providing practical insights for designing time-sensitive AI systems.
24/10/2025 11:00 - 12:00
ESAT Aula L
