NEW YORK — What happened
Cloud infrastructure providers have begun field testing server nodes where electrical copper traces between GPUs are replaced with co-packaged optics (CPO). By bringing silicon photonics modules directly onto the same organic substrate as the compute and memory chiplets, data is transmitted using light signals rather than electrical pulses.
Why it matters
As AI clusters grow to tens of thousands of GPUs, traditional copper cabling consumes an unsustainable portion of the datacenter's power budget just moving bytes across a rack. Copper cables also suffer from signal degradation at frequencies above 100 Gigabaud. Optical interconnects solve both issues, dramatically extending the distance and bandwidth over which GPUs can share memory pools.
Deep dive
Co-packaged optics integrate tiny lasers, optical waveguides, and silicon modulators directly alongside the main GPU silicon. Instead of pushing high-frequency electrical signals through centimeters of circuit board traces to a pluggable optical transceiver at the edge of a blade server, the conversion from electrons to photons happens millimeters away from the compute core. This cuts interconnect latency and slashes the power required for line-driving circuitry.
Report check
Claimed: Co-packaged optical interconnects decrease GPU-to-GPU communication energy consumption by up to 40 percent compared to copper retimers. Verified: Hyperscale operators have publicly demonstrated operational 3.2 Terabit-per-second optical chiplet links in lab conditions. Rumor: Reports that pluggable optical transceivers will be completely phased out of datacenter networking by 2027 are inaccurate; CPO remains expensive and limited to high-density compute fabrics.
Open questions
How will cloud operators handle the field-repairability of failed optical lasers when they are integrated directly inside an expensive GPU package? Will standardization efforts unify competing optical coupling interfaces across different foundry providers?