Why are silicon photonics and co-packaged optics replacing copper interconnects in 1-megawatt AI server racks?
Electrical copper traces have hit a physical wall at 224 Gbps per lane, turning network interconnects into the primary power and latency bottleneck for frontier AI clusters. To scale beyond 100,000 accelerators without melting rack power budgets, hyperscalers are replacing copper with silicon photonics and co-packaged optics.
At 1.6-terabit transceiver speeds, traditional Direct Attach Copper (DAC) cables suffer severe signal attenuation beyond one meter, consuming up to 30 picojoules per bit just to drive retimers and DSPs. By integrating optical engines directly onto the switch and compute interposer via platforms like TSMC's COUPE and GlobalFoundries' Fotonix, Co-Packaged Optics (CPO) slashes interconnect power draw below 5 picojoules per bit while extending lossless reach across entire data halls.
However, converging optical waveguides with 2048-bit HBM4 memory stacks on advanced CoWoS-L packaging introduces acute thermal and mechanical hazards. Because continuous-wave indium phosphide (InP) lasers drift in wavelength above 70°C while adjacent GPU dies dissipate over 1,400 watts, data center architects must deploy direct-to-chip two-phase liquid cooling plates engineered to isolate optical engines from compute hotspots.
Will packaging yields and laser reliability slow down Co-Packaged Optics adoption, or is copper permanently obsolete for next-generation AI clusters?
Comments
Be respectful. Comments reported by several people are hidden until a moderator reviews them.
You do not need an account to read, only to comment. Sign in with a code sent to your email or with your Google account. We only show the name you choose, never your email.