AutoSparse: Automatic Search for Efficient Activation Sparsity-aware CNN Accelerator
DescriptionRecently, Hardware Architecture Search (HAS) has been actively researched to find optimal architectures for the given models with various constraints. However, none of them consider sparsity, to the best of our knowledge. In this paper, we propose the activation sparsity-aware architecture that can flexibly support various dataflows. We then propose the performance model of the architecture, which is utilized in Evolutionary Algorithm-based design space exploration. With the proposed framework, the searched dense accelerator for VGG-16 is 1.18x faster than the manually optimized dense accelerator, and the searched sparse accelerator is 1.16x further faster than the searched dense accelerator.
TimeTuesday, July 11th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby