Nvidia, Cisco Develop New Optical Technology to Boost AI Network Capacity

Nvidia and Cisco are developing a high-density optical module designed to double data transmission speeds per lane and increase network capacity for artificial intelligence infrastructure without requiring additional layers of switches.

The technology, called Ultra-Dense Small Form Factor Pluggable (USFP), is designed to support data transmission rates of 400 gigabits per second per lane, up from 200 gigabits in existing modules, as operators expand data centres to accommodate growing AI computing and inference workloads.

The companies said the first planned modules would deliver aggregate capacity of 12.8 terabits per second, using 16 full-duplex lanes in a form factor comparable in size to existing Octal Small Form Factor Pluggable (OSFP) modules. The design could eventually accommodate 32 or more lanes.

Higher transmission capacity could allow data centre operators to connect more computing systems and 1.6-terabit-per-second endpoints without adding network tiers, which can increase latency, power consumption and operating costs.

Nvidia and Cisco said USFP would deliver four times the bandwidth and front-panel bandwidth density of comparable modules operating at 200 gigabits per second per lane, alongside approximately 30 per cent greater power efficiency.

The companies are targeting AI factories, large-scale computing infrastructure that links processors, accelerators and storage systems to train and run AI models. As demand for AI inference grows, operators face pressure to increase network throughput while limiting electricity consumption and infrastructure costs.

“Building the world’s largest AI factories starts with rethinking the smallest of connections,” said Gilad Shainer, Nvidia’s senior vice-president.

Cisco and Nvidia are also redesigning the module’s connector to support higher-speed electrical signals. The two-dimensional array connector uses a two-stage insertion mechanism intended to reduce contact wear, improve signal integrity and make insertion force more predictable.

The modules are designed to dissipate up to 100 watts without requiring a separate embedded cooling circuit inside the transceiver, potentially simplifying switch design. They will also support optical configurations intended to reduce latency and power consumption by simplifying signal processing.

The companies are working with TE Connectivity and Amphenol to develop cages, connectors and modules, with further partners expected to join the effort.

Nvidia and Cisco plan to establish an open industry specification through a community-led multi-source project, which would allow suppliers to develop compatible components.

The initiative builds on their existing work to integrate Nvidia Spectrum-X Ethernet networking with Cisco Silicon One-based Ethernet networks. The companies are targeting enterprises, neocloud providers and sovereign cloud operators seeking to deploy large-scale AI infrastructure.

Source: FSX Business News