Co-Packaged 1.6-Tbps Silicon Photonics Switches Cut MoE All-to-All Dispatch Tail Latency by 58 Percent Across 1,024-GPU Clusters

Silicon-photonics architects and distributed-training networking engineers across Santa Clara, San Jose, and Stanford University published production fabric telemetry on Saturday evening through the Optical Internetworking Forum (OIF) Co-Packaged Optics working committee, demonstrating that 51.2-Terabit-per-second (Tbps) radix switches equipped with 1.6-Tbps co-packaged silicon photonics (CPO) optical engines reduce Mixture-of-Experts (MoE) AllToAll token-dispatch 99th-percentile tail latency by 58 percent and cut interconnect wall-plug power by 62 percent across 1,024-GPU training clusters. As sparse frontier models scale to 64 or 128 routed experts per layer with top-4 token routing, electrical copper flyover cables and pluggable OSFP transceivers consume more than 30 watts per port in DSP retiming power and suffer frequent link-flap retransmissions.

In the validated CPO package architecture, thirty-two 1.6-Tbps silicon-photonic chiplets are flip-chip bonded directly onto the organic substrate surrounding the central 51.2-Tbps CMOS packet-switch ASIC, shrinking the electrical SerDes trace length from 280 millimeters to less than 9 millimeters. Eliminating the external pluggable retimer DSP lowers port-to-port serialization latency from 115 nanoseconds to 14 nanoseconds while reducing optical energy consumption to 4.1 picojoules per bit driven by external field-replaceable continuous-wave (CW) 8-wavelength DFB laser modules.

To prevent congestion collapse when skewed expert routing concentrates micro-bursts onto hot destination ranks, the switch ASIC implements credit-scheduled adaptive packet spraying across a flat two-tier Dragonfly+ topology. Nanosecond-scale wavelength-selective circuit reconfiguration dynamically reallocates optical lambdas toward oversubscribed expert groups during each forward and backward dispatch barrier.

Tested on a 1,024-accelerator MoE training cluster at 2400 Condensa Street in Santa Clara funded through a $520,000 ARPA-E LIGHTS photonics research grant, step-level communication exposure dropped from 31 percent to 12.8 percent of total iteration time. Open optical link-budget simulators and RoCEv2 adaptive-routing patches are available in the Open Compute Project Networking Repository.