Optimized Network Architectures for Large Language Model Training with
  Billions of Parameters

Weiyang Wang; Manya Ghobadi; Kayvon Shakeri; Ying Zhang; Naader Hasani

返回文献探索

Paper · arXiv 2307.12169

Optimized Network Architectures for Large Language Model Training with Billions of Parameters

Weiyang Wang, Manya Ghobadi, Kayvon Shakeri, Ying Zhang, Naader Hasani

11 upvotesJuly 22, 2023arXiv 预印本

AI 摘要

A new network architecture optimizes GPU communication for Large Language Model training by partitioning GPUs into high-bandwidth domains, reducing network costs.

Large Language Models (LLMs)any-to-any networksGPUshigh-bandwidth communicationHB domainsrail-only connectionClos networks

Abstract

This paper challenges the well-established paradigm for building any-to-any networks for training Large Language Models (LLMs). We show that LLMs exhibit a unique communication pattern where only small groups of GPUs require high-bandwidth any-to-any communication within them, to achieve near-optimal training performance. Across these groups of GPUs, the communication is insignificant, sparse, and homogeneous. We propose a new network architecture that closely resembles the communication requirement of LLMs. Our architecture partitions the cluster into sets of GPUs interconnected with non-blocking any-to-any high-bandwidth interconnects that we call HB domains. Across the HB domains, the network only connects GPUs with communication demands. We call this network a "rail-only" connection, and show that our proposed architecture reduces the network cost by up to 75% compared to the state-of-the-art any-to-any Clos networks without compromising the performance of LLM training.