Omni-Path and Ultra Ethernet: A Convergence Solution for HPC and AI Networking

The data requirements of High Performance Computing (HPC) and Artificial Intelligence (AI) are stressing traditional data center networks. HPC workloads generate diverse communication patterns, while AI utilizes high-bandwidth collectives benefitting from short tail latency. The HPC and AI workloads are strongly supported by specialized fabrics like Omni-Path, but present challenges for Ethernet networks not designed for modern communication patterns. Efficiency in HPC and AI can be improved through various methods. These include optimizing algorithms, enhancing parallelism, and improving data locality. Co-development of advanced hardware features and lightweight, scalable middleware is crucial. Omni-Path, a dedicated high-performance network technology, has demonstrated efficacy in managing both HPC and AI. Ultra Ethernet has emerged as a promising solution, aiming to reconcile the widespread use of Ethernet with the robust HPC/AI performance provided by Omni-Path. By offering higher data rates, reduced latency, and improved congestion control, Ultra Ethernet could efficiently support both HPC and AI workloads within a mixed Ethernet infrastructure. In the realm of high-performance networking, several advancements in high speed fixed function logic and signaling, programmable offloads, and CXL composability are important considerations in a network design that optimizes the widest possible application characteristics. Ultra Ethernet compatible Omni-Path will continue to advance the state of the art in network management, resource allocation, and workload scheduling for HPC, and advancements pioneered by Omni-Path will be crucial for Ultra Ethernet’s success.