Fleet: Hierarchical Task-Based Abstraction for Multi-Die GPUs
Fleet is a new task model that reduces memory traffic in multi-die GPUs, enhancing performance.
Modern GPUs utilize chiplet-based designs with multiple private cache hierarchies, yet current programming models fail to express chiplet-level locality and synchronization. Fleet introduces a multi-level task model that maps computation to memory scopes. This innovation aims to enhance performance in memory-bound workloads by reducing redundant memory traffic and improving cache utilization.