ONE Summit 2024

There are several factors to drive AI networks to expand from supporting 25 thousand GPUs to 1 million GPUs, from interconnecting clusters within a datacenter to interconnecting clusters spanning across tens of datacenters, these include 1) the sustainable power that can be delivered to a single datacenter 2) the growing number of homogenous and heterogenous AI clusters and the desire to bring AI resources together to higher aggregated performance. In the 2nd half of year 2023, non-minimal path routing in Dragonfly and Dragonfly+ topologies were proposed. In this talk we analyze Clos and Mesh topologies based on a model of path diversity from network congestion point of view. We review the findings from AI workload optimization (such as the one from Meta and MIT). We also review the two facets of the non-minimal path routing on (e)BGP protocol. Finally, we come back to work on the Dragonfly and Dragonfly+ topologies with what we just learned in this talk and tentatively propose several large-scale AI networks with likely good characteristics to meet the needs of low network latency, low congestion, zero packet loss, and large bisectional bandwidth.