SMART WORLDJOURNALSubmit Manuscript
← All technology news

Digital Infrastructure

Lambda Enhances AI Training Capacity with NVIDIA Photonics Optical Networking

Lambda integrates NVIDIA Photonics co-packaged optics switches into its AI factory infrastructure, improving power efficiency, network resilience, and latency across over 10,000 GPUs to support large-scale AI model training.

NEWS

Training leading AI models requires large-scale GPU clusters interconnected by networks that can manage massive data flows with minimal latency and maximum uptime. Lambda, a cloud services provider specializing in AI infrastructure, has adopted NVIDIA Photonics co-packaged optics (CPO) switches to advance these networking demands in its deployment of NVIDIA GB300 NVL72 racks, totaling more than 10,000 GPUs.

This deployment uses CPO switches featuring 144 ports operating at 800Gb/s, placed alongside the GPU cluster's switch chips with optical engines integrated into the same package. By bringing optical transceivers onto the switch package, Lambda reduces the electrical path length from inches to millimeters, minimizing signal loss and eliminating the need for power-consuming retimer chips.

One key benefit of this technology is power efficiency: co-packaged optics deliver up to five times better power performance and up to 15 times higher reliability compared to traditional pluggable optics. Furthermore, the switches use four times fewer lasers, simplifying maintenance and reducing potential failure points.

Lambda employs liquid cooling to maintain a median optics temperature of 36°C at the switches, significantly cooler than the 51°C typical of conventional optics. Cooler operating temperatures contribute to signal stability and overall hardware longevity, supporting the network's robustness during large-scale AI training.

Resiliency is crucial because network interruptions in AI factories can stall thousands of GPUs, leading to costly delays. NVIDIA Photonics architecture enables technicians to replace light-source modules individually in the field, unlike standard optics which require replacing entire modules upon laser failure, facilitating faster repairs and sustained operations.

Tests by Lambda show network latency reductions of approximately 4% at rack scale and up to 10% at pod scale, without sacrificing throughput compared to conventional optics. These improvements help maintain consistent data flow between GPUs, essential for synchronized AI training workflows.

While co-packaged optics provide notable advantages, the integration complexity and adoption at scale require careful thermal management and infrastructure design. Lambda’s deployment demonstrates viable performance and reliability gains, signaling a critical step towards the networking infrastructure needed for next-generation AI development.

Source & further reading

Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics ↗

Source published 2026-11-16. Claims and limitations should be checked against the original announcement or research.