faster AI model training and significant energy savings. The technology is expected to reduce power consumption by over five times compared to electrical interconnects and could decrease the time required to train large language models (LLMs) from three months to just three weeks. Additionally, the energy savings per AI model trained equate to the annual power consumption of 5,000 U.S. homes.

The breakthrough addresses the growing demand for efficient data center connectivity as AI workloads scale. IBM’s CPO modules deliver an 80-fold increase in bandwidth compared to electrical connections, enhanced by advanced optical channel density and multi-wavelength transmission. The modules have passed rigorous durability tests, affirming their readiness for mass production. Developed in IBM’s Albany, NY, and Bromont, Quebec facilities, this innovation continues IBM’s legacy in semiconductor leadership. By integrating optics into AI workflows, IBM aims to enable faster, more sustainable, and scalable generative AI processing, setting a new benchmark for the industry.

Read more