New IBM optic tech
IBM has developed a new optics technology to train AI models at light speed, while saving loads of energy. It said by applying its breakthroughs in data centers, the energy saved when training an AI model is equivalent to the annual energy usage of 5,000 US homes.
The company explained that while data centers connect to the outside world with fiber optic cables, internally, it says they're still using copper wires. The wires connect GPU accelerators which spend a lot of time idle waiting for data from other devices while using up energy and driving up costs.
Commenting on the development, Dario Gil, SVP and Director of Research at IBM, said:
"As generative AI demands more energy and processing power, the data center must evolve – and co-packaged optics can make these data centers future-proof. With this breakthrough, tomorrow’s chips will communicate much like how fiber optics cables carry data in and out of data centers, ushering in a new era of faster, more sustainable communications that can handle the AI workloads of the future."
IBM outlines its new co-package optics (CPO) prototype in a technical paper. By significantly increasing the bandwidth in data centers, GPU downtime could be minimized, accelerating AI processing.
IBM explains that the training of large language models (LLMs) could be reduced from three months to three weeks. At the same time, the increased energy efficiency would cut down on energy usage and reduce costs associated with training LLMs.
Source: IBM Newsroom