IBM reinvents the future: artificial intelligence propelled at the speed of light, leaving copper cables behind

In a major technological shift, IBM is redefining the future of data centers by replacing copper cables with optical fibers, using the speed of light to power data.artificial intelligence towards new frontiers. Thanks to a optical module recently presented, AI training could be accelerated up to five times, revolutionizing data center architecture and making traditional electrical interconnections obsolete.

IBM recently announced a significant advancement in the data center industry with the development of a new optical module. This revolutionary technology promises to replace cables in copper traditional, using light for data transmission. It could radically transform the performance ofartificial intelligence (AI), by increasing the training speed of language models, while reducing the energy consumption of data centers. This move to optical fiber should also make it possible to increase the bandwidth between the chips by 80.

The Limitations of Copper in Data Centers

Historically, data centers have relied heavily on wires in copper for the interconnection of electronic components. However, these cables represent a considerable limitation in terms of data transmission capacity and energy consumption. These drawbacks hamper the performance of modern electronics, especially when it comes to training complex AI models.

The use of copper often requires structural compromises, forcing manufacturers to place components as close together as possible to minimize signal loss and improve efficiency. However, this approach shows its limitations as the scale and complexity of IT systems increase.

A revolution through light: IBM’s optical module

To overcome these limitations, IBM has developed a new prototype of an optical module. Using a polymer waveguide for transferring optical signals, this module connects photonic integrated circuits to each other as well as to external connections. This system avoids the dead times caused by waiting for data, which affect the performance of graphics chips used as AI accelerators. The proposed technology offers increased bandwidth, capable of increasing the data transmission capacity between chips by up to 80 times. In terms of connection speed, it reaches impressions of 32 Gbps over 64, which revolutionizes the way in which tomorrow’s chips will communicate with each other.

Major impact on training AI models

Thanks to this advance, IBM claims that training large language models for AI could be up to five times faster. This means that processes that previously took three months could be completed in three weeks. This valuable time saving paves the way for the development of even more massive models, using a greater number of

graphic processors .This speed of execution also allows for greater flexibility in the design of data center architectures, making it possible to deploy interconnections several hundred meters long, unlike the current limitation of one meter with copper.

To read LinkedIn : le grand ménage débute, place aux posts authentiques sans IA

Towards more sustainable data centers

In addition to increasing performance, the transition to the use of optical fibers significantly reduces

power consumption facilities. With each AI model trained, the energy savings are comparable to the annual consumption of 5,000 American homes. This development not only contributes to improving the performance of computer systems, but it also places IBM in a more eco-responsible approach. As Dario Gil, director of research at IBM, points out, this innovation opens a new era of

communications faster and more durable, capable of meeting the growing demands of AI workloads.