Cisco has unveiled a new chip and optical modules to ease AI datacentre network bottlenecks.

SiliconANGLE reported that Cisco showcased open silicon and optical networking systems in a keynote speech at the Open Compute Project (OCP) Global Summit held on Oct. 12 local time.

Networks are cited as a key bottleneck in AI data centres. Frontier models require large clusters made up of thousands of GPUs. Datacentre operators need to move more data between chips without increasing hardware costs and power use. Cisco is pursuing a strategy to address this with a “scale-across” architecture that links distributed AI data centres into one.

A core element of the scale-across architecture is the Silicon One family of chips.

Cisco also introduced an optical module made with Nvidia. The ultra-high-density small detachable USFP optical module handles 400 gigabits of electrical signal per lane. Cisco said this can increase transmission speeds between Nvidia GPUs.

Gilad Shainer (길라드 샤이너), Nvidia's senior vice president, said, "Building the world's largest AI factories starts with rethinking the smallest connections."

Cisco also launched the Cisco 8001, an out-of-band management system designed to rapidly restore networks when failures occur. It added a "Live Protect" feature to a new version of the open-source SONiC network operating system. Live Protect finds and fixes vulnerabilities without stopping AI workloads in progress. It is set to be available from next year.

Kevin Wollenweber (케빈 월렌와이버), Cisco's senior vice president for the Data Center & Internet Infrastructure division, said, "The problem is no longer each individual data centre or cluster," adding, "The key is how to connect large-scale computing without harming performance and reliability."