Nvidia's NVSwitch Dominates Scale-Up AI Networks, but Ultra Ethernet Looms

Nvidia's NVSwitch has emerged as the leading interconnect technology for scale-up AI networks, echoing the dominance of InfiniBand in the past. The story begins with the dawn of time when two rival future I/O schemes, InfiniBand and Ethernet, were competing to become the standard for PCs and servers.

As the Dot Com boom subsided, the industry opted for Ethernet as a point-to-point interconnect, leaving InfiniBand in the shadows. However, Voltaire and Mellanox Technologies revived InfiniBand with their switch ASICs and NICs, respectively. The technology gained traction in HPC ModSim workloads until it was eventually acquired by Intel.

In recent years, the Ultra Ethernet Consortium has been formed to marry the low latency and high bandwidth of InfiniBand with the scalability and quality of service features of Ethernet. Nvidia's NVSwitch interconnects have also gained prominence, especially in AI clusters that support 1 million endpoints.