- IBM has introduced a co-packaged optical module that uses polymer optical waveguide technology to enhance data centre bandwidth and speed up AI training.
- The module could boost bandwidth between chips by up to 80 times and reduce data centre energy use by 80%, improving AI processing efficiency.
What happened: IBM researchers introduce new optical module
IBM researchers have unveiled a new co-packaged optical (CPO) module that could revolutionise AI training by enabling fibre optic-level speeds in data centres. The module uses polymer optical waveguide (PWG) technology to provide high-speed optical connectivity, significantly increasing data centre bandwidth. The innovation aims to reduce GPU downtime and accelerate AI processing speeds. IBM’s Dario Gil, SVP and director of research, explained, “As generative AI demands more energy and processing power, co-packaged optics can future-proof data centres.”
Unlike traditional data centres, which rely on copper-based electrical wires inside racks, IBM’s module applies optics directly to these racks, potentially boosting bandwidth between chips by up to 80 times compared to electrical connections. The new technology could also allow AI hardware makers to add six times as many optical fibres to silicon photonics chips, enabling faster, more energy-efficient data transmission. This breakthrough could cut energy use in data centres by up to 80% and speed up AI model training, ushering in a new era of sustainable, high-speed communication.
Also read: IBM reports growth in AI bookings, beating earnings estimates
Also read: Discovering the power of fibre optic networks
Why it is important
This innovation is crucial as it addresses the growing demands of AI, particularly in data centres. By enabling fibre optic speeds, IBM’s new co-packaged optical (CPO) module enhances data centre bandwidth and reduces GPU downtime. Unlike traditional copper-based connections, the module uses polymer optical waveguide (PWG) technology to offer faster, more energy-efficient data transmission.
The ability to boost bandwidth between chips by up to 80 times could significantly accelerate AI processing speeds. Additionally, this technology could reduce energy consumption in data centres by up to 80%, making operations more sustainable. With generative AI requiring more power and efficiency, IBM’s solution future-proofs data centres, ensuring they can handle the increasing demands of AI workloads.
Overall, the breakthrough could not only improve AI model training speeds but also promote responsible, high-speed communication, paving the way for faster, more sustainable AI advancements.
