Published: 01 April 2024
Summary
New networking technologies like Ethernet, InfiniBand, PCIe, CXL and optical interconnects play a key role in boosting speed, efficiency and scalability for GenAI workloads. Product leaders must use this trend analysis for near-term solutions and integrate these trends into their future product plans.
Included in Full Research
Overview
Implications
Interchip optical interconnects enable data centers to scale up compute clusters for generative AI (GenAI) workloads, with enhanced power efficiency and higher bandwidth.
Ethernet in its standard configurations is not ideal for AI training.However, remote direct memory access (RDMA) over Converged Ethernet (RoCE) using lossless Ethernet, or Ultra Ethernet may prove to be viable alternatives given the open ecosystem and reliability of such networks.
InfiniBand is a viable alternative to Ethernet for GenAI workloads, given the dominance of graphics processing units (GPUs) by a single supplier; however, InfiniBand networks are priced higher and do not have an open ecosystem.
Peripheral Component Interconnect
Clients can log in to view the entire
document.