Lambda cut network delays by up to about 10% in tests of optical switches connecting its artificial intelligence computing equipment, Nvidia said in a blog post published Nov. 16.
The switches maintained the same data throughput as conventional plug-in optics across three configurations, Nvidia, which supplies the equipment, said. Its post, “Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics,” described Lambda’s deployment and tests but gave no testing dates or duration.
Lambda, a cloud provider, is using the switches to connect Nvidia GB300 NVL72 racks containing more than 10,000 graphics processing units, or GPUs, the company said. The processors work together to train AI models, making the network responsible for moving data between them.
Network latency, the time needed to move data, fell about 4% in rack-scale tests and up to about 10% at the larger pod scale, Nvidia said. The throughput comparisons covered rack, scalable-unit and pod configurations. The post supplied no absolute latency measurements.
Nvidia said its photonics technology delivers up to five times the power efficiency and up to 15 times the reliability of conventional plug-in optics. Those figures describe the technology’s performance according to Nvidia, while the post separately identified the latency results as Lambda’s tests.
The technology also uses a quarter as many lasers, reducing the number of components that can fail, Nvidia said. The company said a single interrupted network link can stall thousands of GPUs during a large training job, costing computing time.
Lambda reported a median optical-component temperature of 36 degrees Celsius, compared with 51 degrees for conventional plug-in optics, Nvidia said. Lambda’s tests also found more stable signals. Liquid cooling removes heat from both the switch chip and the optical engines.
Each switch has 144 ports operating at 800 gigabits a second and 18 removable external light sources, Nvidia said. Optical engines beside the switch chip convert electrical signals into light, which carries data over fiber between parts of the computing cluster.
Technicians can replace the external light sources in the field, Nvidia said. Conventional plug-in optics require replacement of the entire module when a laser fails. Placing optical engines on the switch package also shortens the electrical path from inches to millimeters, reducing signal loss.
The shorter connection removes the need for chips that restore signal timing, cutting electricity use and the number of components in the signal path, Nvidia said. Dave Ward, Lambda’s president of cloud services, said networking must keep pace as AI computing clusters grow.


