Microsoft Research Develops Laser-Free Optical Interconnect for AI Data Centers

Microsoft Research has developed a laser-free optical interconnect called MOSAIC to reduce energy consumption in AI data center networking ^[1].

This development addresses the escalating power demands of artificial intelligence workloads. As data centers expand to support larger models, the energy required to move data between chips becomes a significant operational cost and environmental burden.

MOSAIC utilizes MicroLED-based technology rather than traditional lasers to facilitate optical interconnects ^[1]. By removing the need for lasers in the cabling process, the system aims to streamline how data travels across networking environments ^[1]. This shift in architecture targets the specific inefficiencies found in current high-speed data transmission methods used for AI.

The impact on power efficiency is substantial. Microsoft Research said the technology can reduce cabling energy consumption by up to 50% ^[1]. Such a reduction could lead to significantly lower power bills for operators of large-scale AI clusters.

Traditional optical networking relies on laser sources that can be power-intensive and difficult to integrate at scale. The MicroLED approach used in MOSAIC provides a different path for scaling interconnects without the proportional increase in energy overhead ^[1]. This allows for denser networking configurations within the data center without exceeding thermal or power limits.

The research focuses specifically on the networking environments where AI workloads are most intensive ^[1]. By optimizing the physical layer of data transmission, the company seeks to ensure that the growth of AI compute does not lead to unsustainable energy requirements.