IBM has unveiled a revolutionary advancement in optical technology that promises to transform how data centers train and execute generative artificial intelligence models. The company has developed an innovative process for co-packaged optics (CPO), enabling connections within data centers at the speed of light and complementing the short-range electric cables currently in use.
With this advancement, IBM introduces the first prototype module based on a polymer waveguide (PWG) that redefines how high-speed data flows between chips, circuit boards, and servers. This development could mark the beginning of a new era of faster and more sustainable communications within data centers.
Optical Technology Takes Center Stage in Data Centers
Although optical fibers already transport data at high speeds over long distances, data centers still heavily rely on copper cables for their internal communications. This creates inefficiencies, such as downtime for GPU accelerators, which can spend more than half their time waiting for data from other devices.
IBM has taken a step forward by demonstrating how optics can significantly accelerate these internal communications. According to a technical study, their CPO module can increase data transmission speeds by up to 80 times compared to current electrical connections. This advancement not only reduces wait times in AI model training processes but also promises:
- Reduction in energy costs: An over 80% decrease in energy consumption compared to current mid-range electrical interconnects.
- Increased AI model training speed: It would allow for training language models (LLMs) up to five times faster, cutting training time from three months to just three weeks.
- Improved energy efficiency: Each trained AI model could save energy equivalent to the annual consumption of 5,000 U.S. households.
Dario Gil, Senior Vice President and Director of Research at IBM, stated, “As generative AI demands more energy and processing power, data centers must evolve. Co-packaged optics can make these centers more efficient and future-ready.”
Beyond the Limits of Electrical Connections
IBM’s advancement is based on the design and assembly of a high-density PWG capable of incorporating six times more optical fibers at the edge of a silicon photonic chip than current technologies. Each fiber, only three times the width of a human hair, can transmit terabits of data per second over distances ranging from centimeters to hundreds of meters.
Moreover, IBM’s CPO modules have passed all necessary stress tests for manufacturing, including extreme temperatures (-40°C to 125°C), high humidity, and mechanical durability. These features make the technology suitable for mass implementation in the industry.
The Impact on Data Centers and Generative AI
The implementation of this optical technology in data centers not only enhances performance but also addresses the increasing energy demands of generative AI. In recent years, the transistor density in chips has significantly increased; for example, IBM’s 2-nanometer technology contains over 50 billion transistors on a single chip. By introducing optical connections, IBM tackles the challenge of scaling these interconnection densities without the limitations of electrical pathways.
A Legacy of Semiconductor Innovation
This advancement solidifies IBM’s leadership in semiconductor research and development. Previous achievements include the creation of the first 2 nm chip, the implementation of 7 nm and 5 nm process technologies, and advances in nanosheet and vertical transistors.
The design, modeling, and simulation work for the CPO was carried out in Albany, New York, while the prototypes and modules were assembled at IBM’s facilities in Bromont, Quebec. The latter, part of the Northeast Semiconductor Corridor between the U.S. and Canada, has been a global benchmark in chip packaging for decades.
Conclusion
IBM continues to set the pace for innovation in computing and semiconductors. Its new optical technology promises not only to accelerate the training of AI models but also to transform the energy efficiency and capabilities of data centers. As the demand for AI grows, this advancement positions IBM as a key player in creating faster, more sustainable, and future-ready infrastructures.