Alphabet's 'Optical Net' Challenges NVIDIA's Dominance in Data Center Interconnect Architecture

Dolphin Research is putting Alphabet's (GOOGL.US) 'Optical Net' under scrutiny, comparing it to NVIDIA's (NVDA.US) dominance in data center interconnect architecture. The research highlights Google's unique approach, which includes a vast Scale-Up domain with over ten thousand chips and a highly distinctive topology using direct-connect instead of switches.

Google's network uses a proprietary TPU interconnect protocol called ICI, which aggregates 2,400 GB/s bidirectional per TPU, doubling the v7 generation. The company also leverages Jupiter DCN, its general-purpose Ethernet across all servers for search, ads, and storage, to carry cross-Pod traffic.

The Optical Net uses a 3D block-building topology that scales from chip → tray → cube (i.e., rack) → Pod/Superpod. Intra-Tray Scale-Up is achieved with physical 1×4 layout and logical 2×2 topology, while Tray-to-Tray Scale-Up involves direct-connect topology providing connectivity.

The Optical Net's OCS switch aggregates optical signals and decides whether a port pair self-loops into a 64-chip torus or assembles into a larger torus. Unlike EPS (electronic packet switching), OCS establishes static optical paths between ports, pre-configured in software to avoid spending seconds reconfiguring under low-latency synchronous workloads.