The idea with the CPO module is to increase the bandwidth of data center communications and minimize GPU downtime while drastically accelerating AI processing, Khare said. The technology can reduce power consumption by 5x compared to mid-range electrical interconnects, he said.
Expectations for co-packaged optics
The CPO package is available for industry development, and IBM hopes to develop an ecosystem of customers and uses for the CPO technology.
“IBM has designed this technology for broad deployment within the industry. Given the compute and energy-efficiency demands for technologies like generative AI, and 5G and even 6G communications, we expect multiple applications,” said IBM researcher Griselda Bonilla.
In the big picture, the CPO technology IBM is developing is a push forward for the technology and AI deployments, according to industry watchers.
“If you look at the largest AI clusters built by hyperscalers today, they are already approaching 100,000 accelerators, with projections reaching up to one million in the near future. As these clusters continue to grow, they become increasingly distributed, potentially affecting latency and processing times. And maximizing the number of GPUs in a single rack can significantly reduce latency,” said Sameh Boujelbene, vice president, data center switch and AI networks market research with the Dell’Oro Group.
“The number of chips that can be connected on an electronic module has historically been limited by electrical pathways,” Boujelbene said. “However, CPO holds the promise of significantly increasing the interconnection density between accelerators and chips.
