Back to Stack Tree
Compute & Silicon
Optical & Co-Packaged Photonics
Overview
As chip clusters grow, copper cables become a major bottleneck due to resistance, heat, and physical size. Silicon photonics uses lasers and optical waveguides to transmit data between processors at the speed of light, drastically reducing latency and energy consumption.
Role in the Stack
Modern training runs are split across thousands of interconnected chips. If chip-to-chip bandwidth is slow, processors waste cycles waiting for data. Light-based networks unlock true linear scaling for mega-clusters.
Key Pioneers & Stakeholders
Broadcom
Luminate
Ayar Labs
Celestial AI
NVIDIA (NVLink Optical)
Key Statistics & Targets
10x over copper
Bandwidth Density Improvement
5x lower latency
Latency reduction
Up to 90% savings
Power reduction
Future Industry Directions
1
Co-packaged Optics (CPO)
Mounting optical transceivers directly on the processor substrate instead of external slots.
2
All-optical routing
Switching network packets using light paths, avoiding electrical-optical conversions.