AI’s hunger for computational power and energy appears insatiable. Even the most powerful data centers are using more processors and energy to meet the growing demand.
To address this, Intel announced it has developed what it calls the Optical Compute Interconnect (OCI) chiplet, which uses light rather than electricity to transmit data between processors. Intel believes this is the first such device that can cope with the computational workload required by the AI market. The company recently demonstrated the device, integrated with one of its CPUs, at the Optical Fiber Communication Conference.
“There’s a number of challenges that are known but accelerated and exacerbated by AI,” said Thomas Liljeberg, senior director of product management and strategy at Intel, at a media briefing prior to this announcement. “This product is what we believe is very compelling and a scalable solution to the challenges.”
Information constantly travels between servers as it’s being processed at a data center, typically connected by electric signals traveling on a copper cable. While this allows for the high-bandwidth, low-energy transmission essential for efficient processing, the cables are limited in length to just over 3 feet – which places limitations on how you can connect things together. A widely used alternative is relying on pluggable modules to perform data transmission, but these devices are often inefficient and have latency issues.
Intel’s integrated chiplet addresses these problems because its optical signaling – where laser light travels through fiber optic connections – makes it possible for processors to communicate over longer distances—almost 100 times longer than copper. This allows for more flexible and distributed data center architectures. With this, processors can handle more powerful AI models while running with reduced latency and improved efficiency, said Liljeberg.
In addition to the optical connections, the OCI is fitted with an integrated circuit to process information, an electrical circuit to handle electric signals, and a path to attach an optical connector that acts like a port for the light signals. It can also be expanded to move more data at quicker speeds, enabling more scalability for data transmission between processors. That modularity potentially makes it easier for data centers to add them to meet demand.
Intel is not the only company looking into photonic technology for semiconductors. Companies like Global Foundries and Ayar Labs also have boasted the use of photonic integration to aid their AI architecture and products. However, questions abound about what scale these products need to be used at to make economic sense, as optical technologies are more expensive than conventional ones.
When it comes to Intel’s solution, “I think the price is justified because of the benefits and the use cases that are associated with it,” said Bhavin Manjee, partner and board member at consultant firm Simon Kucher. “It’s ultimately much more reliable and flexible as well.”
Another advantage for Intel is that its new product won’t be limited to its own hardware. It is going to be compatible with most processors, even those from other vendors like Nvidia and AMD, Liljeberg said.
“We think we have a very compelling solution to that problem and we’ll bring that to the broader market,” he said. “We will work with the ecosystem to ensure interoperability and make sure that our photonics chip can talk to a photonics chip from another vendor.”
Moving forward, Intel plans to deliver the chiplet in large volumes for integration with processing decks. The company is also running pilot programs with select customers to begin integrating the technology as it scales up production.
Once manufacturing is scaled up, Lijeberg said that the OCI “can deliver both on cost and manufacturability.” He added that as more data centers integrate this technology, “it’s going to continue to scale and continue to support bandwidth requirements for generations ahead.”
