Communication Characterization of AI Workloads for Large-scale Multi-chiplet Accelerators

TL;DR

This paper analyzes communication bottlenecks in multi-chiplet AI accelerators, highlighting multicast traffic as a key scalability challenge and proposing flexible interconnect solutions to enhance performance.

Contribution

It provides a detailed analysis of communication patterns in multi-chiplet AI workloads and suggests architectural improvements for better scalability.

Findings

Multicast traffic can significantly limit AI workload scalability.

Communication bottlenecks are workload-dependent and critical for performance.

Flexible interconnect architectures can mitigate communication bottlenecks.

Abstract

Next-generation artificial intelligence (AI) workloads are posing challenges of scalability and robustness in terms of execution time due to their intrinsic evolving data-intensive characteristics. In this paper, we aim to analyse the potential bottlenecks caused due to data movement characteristics of AI workloads on scale-out accelerator architectures composed of multiple chiplets. Our methodology captures the unicast and multicast communication traffic of a set of AI workloads and assesses aspects such as the time spent in such communications and the amount of multicast messages as a function of the number of employed chiplets. Our studies reveal that some AI workloads are potentially vulnerable to the dominant effects of communication, especially multicast traffic, which can become a performance bottleneck and limit their scalability. Workload profiling insights suggest to architect a flexible interconnect solution at chiplet level in order to improve the performance, efficiency and scalability of next-generation AI accelerators.