AI Has Hit the Copper Wall — Light Is the Way Out
Frontier AI models have outgrown electrical wiring. In multi-thousand GPU clusters, copper cables generate massive heat and signal loss over just 2 meters — eating up to 30% of total datacenter energy just moving data.
Frontier AI models have outgrown electrical wiring. In multi-thousand GPU clusters, copper cables generate massive heat and signal loss over just 2 meters — eating up to 30% of total datacenter energy just moving data.
Silicon photonics replaces electrons with laser photons.
The shift:
Copper: 15–25 pJ/bit, degrades past 2m, heavy rigid cables blocking airflow
Optics: 1–2.5 pJ/bit, reaches 100–500m, lightweight fiber ribbons
Three breakthroughs making it mainstream:
Co-Packaged Optics (CPO) — Mounting the optical engine directly on the same substrate as the GPU. Electrical path drops below 1mm, cutting I/O power by 70%.
Multi-wavelength micro-ring resonators — A single fiber carries 16–32 laser wavelengths at 100 Gbps each. That's 3.2 Tbps per strand with zero cross-talk.
External Laser Sources (ELS) — Lasers moved off the hot compute package to cool front-panel modules. Failed diodes hot-swap in seconds without stopping GPUs.
Why hyperscalers need this:
Gigawatt AI campuses are capped by local grid allocations. Trimming 15–20 pJ/bit across petabytes of traffic saves tens of megawatts — redirected straight into training compute.
It also unlocks true memory disaggregation: GPUs accessing distant HBM pools as if they were local.
Full read: https://www.nizover.com/en/blog/tech-silicon-photonics-optical-interconnects-ai
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.