Fleet: Hierarchical Task-Based Abstraction for Megakernels on Multi-Die GPUs

Hardware Architecture(arxiv.org)view on HackerNews
GPUmulti-dietask-basedabstractioncachememoryhierarchychiplet

Author: matt_d

Date: 7/16/2026

Article Summary:
The paper presents Fleet, a hierarchical task model that maps computation to memory scopes and introduces a new abstraction called Chiplet-tasks to improve cache utilization and reduce memory traffic in multi-die GPU designs.