InfiniteHBD: Datacenter-Scale High-Bandwidth Domain for LLM with Optical Circuit Switching Transceivers#

Chenchen Shou et al. (Peking U. / StepFun / Lightelligence)

Why: HBD architecture (NVL-72 vs TPU-style) is a live topology/cost/fault-domain debate; this proposes a genuinely different point in that space.

What: A transceiver-centric HBD that embeds optical circuit switching into each transceiver, giving reconfigurable point-to-multipoint links and variable-size ring topologies at ~31% of NVL-72 cost with node-level fault isolation.

How: A silicon-photonic OCS transceiver (OCSTrx), a reconfigurable k-hop ring topology, and an HBD-DCN orchestration algorithm; evaluation claims near-zero GPU waste and 3.37x MFU vs DGX.

Where applicable: Aimed at LLM-training fabrics; depends on custom silicon-photonic transceivers, so it's a direction/architecture signal rather than commodity-deployable today.