NEW YORK — What happened
Major data center hardware providers have deployed the first commercially available AI server racks featuring Co-Packaged Optics at scale, using optical fiber links directly connected to processing nodes.
Why it matters
Electrical copper interconnects lose substantial energy as signal speeds increase over short distances. At modern data transfer rates, driving electrical signals across backplanes and switches consumes up to 30% of total cluster power. Moving light-based optical transceivers directly onto the package silicon drastically reduces power dissipation and latency.
Deep dive
In traditional setups, electrical signals travel from the GPU die, through substrate traces, across circuit boards, and into pluggable optical transceivers at the switch front panel. Co-Packaged Optics removes these long copper pathways by placing tiny silicon photonic engines directly on the same substrate as the GPU compute die. Light travels directly through fiber cables connected near the chip, reducing signal degradation and cutting interconnect power consumption by up to 40%. The technical challenge lies in thermal management: laser light sources are highly sensitive to temperature fluctuations caused by nearby high-wattage GPU dies. As a result, current systems utilize external, field-replaceable laser sources routed via optical fibers into the package.
Report check
Claims assert that Co-Packaged Optics eliminates all copper connections inside AI data centers. What is verified: Optical interconnects reduce copper usage on board traces and backplanes for long rack connections. What is still rumor: Claims that short-distance on-chip communication will shift to optics in this hardware generation remain inaccurate; on-die logic relies on copper.
Open questions
Will the cost of external laser engines drop fast enough for Co-Packaged Optics to be adopted outside of ultra-large hyperscale AI data centers?