Lambda Deploys NVIDIA Co-Packaged Optics for 10,000+ Blackwells
7 hour ago / Read about 9 minute
Source:Converge Digest
Lambda says NVIDIA Photonics switches connect more than 10,000 Blackwell Ultra GPUs in Québec, with 10.1% lower pod-level network latency than conventional pluggable optics in its benchmarks.

Lambda has deployed NVIDIA Photonics switches connecting more than 10,000 Blackwell Ultra GPUs at its Québec campus, the AI cloud provider said November 16. Lambda describes it as the world’s first large-scale deployment of co-packaged optics (CPO). In Lambda’s comparison with conventional pluggable optics, the CPO fabric delivered 3.8% lower network latency at rack scale and 10.1% lower latency at pod scale, with the same throughput across rack, scalable-unit and pod tests.

The switches connect NVIDIA GB300 NVL72 racks. Each liquid-cooled switch has 144 ports running at 800 Gb/s and 18 removable external light sources. The optical engines sit alongside the switch ASIC, shortening the electrical path from inches to millimeters and eliminating the need for retimers on that path. Lambda reported more stable signal integrity and a median optics temperature of 36°C, compared with 51°C for the pluggable configuration.

Lambda said the Québec campus has secured 142 MW of capacity from nearly all renewable sources and can use free cooling for up to 80% of the year. The site carries a guaranteed power usage effectiveness (PUE) of 1.2, and server cooling uses no process water. NVIDIA plans to feature the deployment in its SC26 special address.

“Networking has to keep pace with the demands of increasingly large AI clusters,” said Dave Ward in Lambda’s announcement.

Québec CPO deployment at a glance
• Cluster More than 10,000 Blackwell Ultra GPUs in NVIDIA GB300 NVL72 racks.
• Switch hardware 144 × 800 Gb/s ports per switch, liquid cooling and 18 removable external light sources.
• Latency and throughput 3.8% lower latency at rack scale and 10.1% lower at pod scale. Throughput matched pluggables at all three test scales.
• Optics 36°C median for CPO versus 51°C for pluggables. Lambda also reported more stable signal integrity.
• Campus 142 MW of secured capacity from nearly all renewable sources and a guaranteed PUE of 1.2.

Analysis

Lambda disclosed a deployment of Quantum-X800 InfiniBand Photonics at the 10,000-GPU scale in March. In June, Converge Digest covered Lambda’s evaluation of the liquid-cooled Quantum-X Photonics Q3450-LD. The November disclosure identifies the Québec site and gives operators a rack-to-pod comparison. The latency advantage widened as the test configuration grew while throughput held steady. Lambda did not disclose the absolute latencies or the test traffic profile, which limits comparisons with other fabrics.

The external laser modules give operators a field-replaceable light source even though the optical engines are integrated into the switch package. NVIDIA cites up to 15× higher reliability for its photonics technology. The Québec disclosure provides comparative temperature and signal-integrity results, with no field failure rates, observation period or measured fabric power savings. Those figures would help quantify how the design affects interruptions and power use in day-to-day operation.

Company Profile | Lambda
Business GPU cloud infrastructure for AI model training and inference.
Headquarters San Jose, California, United States.
Founded and ownership 2012 · Privately held.
Leadership Michel Combes, CEO · Stephen Balaban, co-founder and CTO.
Platforms Lambda Cloud, including Superclusters and 1-Click Clusters.