Session

DescriptionIn this session, many aspects of accelerating applications with complex data structures and online scheduling of multi-tenant deep learning inference will be discussed. GNNs and deep learning on point clouds require efficient handling data with high sparsity. Furthermore, 3D volume imaging suffers from the poor data locality which can be solved by redesigning the dataflow. For deep learning servers, effectively handling multi-tenancy with online scheduling is highly important that help satisfy the service-level-agreement. Last but not least, a novel Fibonacci quantization method will be presented that improves the energy efficiency with multiplier-free accelerator design.