SCAR: Scheduling Multi-Model AI Workloads on Heterogeneous Multi-Chiplet Module Accelerators

TL;DR

This paper introduces SCAR, a scheduler for heterogeneous multi-chiplet AI accelerators, optimizing multi-model workloads to significantly reduce energy-delay product in datacenter and AR/VR applications.

Contribution

The paper presents a novel heuristic-based scheduler, SCAR, for heterogeneous multi-chiplet AI accelerators, improving workload adaptivity and energy efficiency over homogeneous architectures.

Findings

Achieves 27.6% reduction in energy-delay product for datacenter workloads.

Achieves 29.6% reduction in energy-delay product for AR/VR workloads.

Effectively manages the complex scheduling space with heuristics and pipelining techniques.

Abstract

Emerging multi-model workloads with heavy models like recent large language models significantly increased the compute and memory demands on hardware. To address such increasing demands, designing a scalable hardware architecture became a key problem. Among recent solutions, the 2.5D silicon interposer multi-chip module (MCM)-based AI accelerator has been actively explored as a promising scalable solution due to their significant benefits in the low engineering cost and composability. However, previous MCM accelerators are based on homogeneous architectures with fixed dataflow, which encounter major challenges from highly heterogeneous multi-model workloads due to their limited workload adaptivity. Therefore, in this work, we explore the opportunity in the heterogeneous dataflow MCM AI accelerators. We identify the scheduling of multi-model workload on heterogeneous dataflow MCM AI accelerator is an important and challenging problem due to its significance and scale, which reaches O(10^56) even for a two-model workload on 6x6 chiplets. We develop a set of heuristics to navigate the huge scheduling space and codify them into a scheduler, SCAR, with advanced techniques such as inter-chiplet pipelining. Our evaluation on ten multi-model workload scenarios for datacenter multitenancy and AR/VR use-cases has shown the efficacy of our approach, achieving on average 27.6% and 29.6% less energy-delay product (EDP) for the respective applications settings compared to homogeneous baselines.