1 min read
Add as a preferred source on Google

Lambda Connects More Than 10,000 GPUs With NVIDIA Photonics Switches

Lambda’s tests showed latency fell about 4% at rack scale and up to 10% at pod scale, while throughput matched conventional pluggable optics.

Fiber cables connect a photonics switch beside liquid-cooling lines / TokenPost.ai
Fiber cables connect a photonics switch beside liquid-cooling lines / TokenPost.ai

Lambda deployed NVIDIA Photonics switches to connect more than 10,000 NVIDIA GB300 NVL72 GPUs, with tests showing lower latency at rack and pod scale while throughput remained unchanged.

Lambda’s tests recorded about 4% lower latency at rack scale and up to 10% at pod scale compared with conventional pluggable optics. Throughput stayed the same.

The switches have 144 ports running at 800 gigabits per second and 18 removable external light sources. The co-packaged optics design shortens the electrical path from inches to millimeters and removes the need for retimer chips.

The system uses liquid cooling, with median optics temperatures of 36°C compared with 51°C for conventional optics. Lambda says it uses one-quarter as many lasers and offers up to five times better power efficiency and 15 times higher reliability than conventional pluggable optics.

Networking is one part of an AI factory’s broader mix of compute, storage, power and cooling, as described in TokenPost’s earlier coverage of NVIDIA’s modular AI-factory systems.

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

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…