IBM is advancing its co-packaged optics (CPO) technology, which aims to replace traditional electrical interconnects in chips, significantly enhancing AI model training and energy efficiency in data centers.

This innovative CPO technology is projected to reduce energy consumption by over five times compared to mid-range electrical interconnects, while also extending cable lengths from one meter to hundreds of meters.

Utilizing polymer materials instead of glass fiber optics, the CPO prototype promises lower costs, improved flexibility, and reduced interference, making it a viable option for modern data centers.

The development and testing of the CPO technology took place at IBM's facilities in Albany, New York, and Bromont, Quebec, underscoring the company's leadership in semiconductor innovation.

IBM's research team is now focused on bringing this technology to production, collaborating with component suppliers and gathering client feedback to refine the CPO modules.

According to IBM, the CPO module can enhance chip-to-chip communication bandwidth by up to 80 times compared to existing electrical connections, which is crucial for handling the increasing demands of AI workloads.

Tests indicate that using polymer optical waveguides in data centers could reduce power usage by five times compared to conventional systems, paving the way for more flexible architectures with high data transfer rates.

This breakthrough technology could drastically shorten the training time for large language models from three months to just three weeks, while also saving energy equivalent to the annual consumption of 5,000 U.S. homes per model.

Dario Gil, IBM's SVP and director of research, emphasized that integrating CPO technology will make data centers more sustainable and capable of meeting the processing demands of future chips.

The push for CPO technology reflects a broader trend in the tech industry aimed at improving connectivity solutions to meet the escalating demands for AI processing and data transmission.

The rising demand for efficient data management in data centers is largely driven by the growth of Generative AI, which now accounts for nearly 75% of data center traffic.

Industry analysts highlight that the expansion of AI clusters, potentially reaching up to one million accelerators, will heighten the need for efficient interconnection technologies like CPO.