THz-SynC: Using terahertz wireless links with optical fabric to speed and save energy in AI datacenters
This paper introduces THz-SynC, a system that combines terahertz (THz) wireless links with an optical wired fabric to improve how large AI training jobs move data between racks. The authors focus on “collective” communication patterns that dominate distributed training, such as All-to-All and AllReduce. THz-SynC tries to cut the time it takes for these collectives to finish while also lowering transmission energy.
The work has two main parts. First, the system builds collective-specific communication structures — tuned layouts and schedules that respect the rules of All-to-All and AllReduce operations. These synthesized structures consider uneven compute speeds across racks (so-called stragglers) and allow parallel wireless transmissions without breaking collective semantics. Second, a contextual-bandit coordinator makes on-the-fly decisions about which pieces of a collective should go over the optical fabric and which should use THz wireless links. The coordinator also sets per-rack THz power budgets based on real-time observations of network state and the collective’s needs.
In practice, wired chunks of a transfer use the optical network, while wireless chunks go through multi-round scheduling and wireless resource allocation. The THz overlay is modeled as directional, single-hop rack-to-rack links that can be set up on demand. The authors tested THz-SynC in trace-driven simulations that include training dependencies similar to Megatron-LM. According to those traces, THz-SynC finds a better tradeoff between delay and energy than baselines that use only wired, only wireless, or simpler hybrid approaches.
Important caveats remain. THz channels can fluctuate and optical links can become congested. Compute heterogeneity across racks also complicates scheduling. The results reported come from simulations rather than a real hardware deployment, so practical issues such as control overhead, physical interference or blockage, and implementation costs could affect real-world gains. The paper also assumes a specific rack layout (concentric rings) and directional transceivers, which may not match all datacenter designs.