“While the scale-up domain today is largely serviced by passive copper, data rates and rack densities are necessitating a shift to alternatives,” Naji wrote. “While many of the optical providers like Marvell (following its acquisition of Celestial AI), Broadcom, and Nvidia believe that co-packaged optic is the right solution, others point to pluggable copper, microLEDs, linear optics, and near-packaged optics as potentially better alternatives for scaling up (for their power efficiency, flexibility, and cost). We expect that as connections slowly shift away from copper over the five years, there will be a broad mix of these solutions that succeed, depending on the use-case and scale of a particular system or deployment.”
The need for speed drives bandwidth density
Woodside Capital Partners predicts that 1.6T is expected to surpass 800G as the primary port speed in new AI backend fabrics by 2027. “The shift to 224G lanes, the push toward 448G, and the unprecedented production ramp rate of 1.6T optical modules highlight how aggressively the roadmap is being pulled forward,” the firm wrote.
One example of the speed ramp-up is Broadcom’s Taurus BCM83640, a 400G/lane optical DSP built for 1.6T transceiver applications. The Taurus platform of DSPs will enable the next-generation 3.2T optical transceiver modules.
400G/lane technology is the next evolution of 200G/lane architectures, enabling a critical step in scaling bandwidth for high-performance networking and AI infrastructure, according to Broadcom. “1.6T pluggable modules using the Taurus BCM83640 double the bandwidth per optical lane, effectively enabling 102.4T switching capacity in a 1RU system to improve bandwidth density in AI optical interconnects. Further, the adoption of 400G/lane optical interfaces lays the foundation for the eventual deployment of 3.2T module solutions with 400G/lane electrical interfaces for 204.8T switches,” Broadcom stated.
“We expect more than 100 million units of 1.6T and 3.2T optical transceivers to be shipped over the next 5 years with close to half of these using 400G optics,” said Vladimir Kozlov, CEO and founder of analyst firm LightCounting, in a statement related to Broadcom’s Taurus announcement. “High speed optical interconnects are essential for operation of AI clusters. Doubling of the lane rates has been a proven strategy to keep up with the bandwidth growth and it is great to see the first 400G per lane solutions becoming available.”
Energy efficiency for the AI era
AI’s energy requirements are a challenge for enterprise and hyperscaler organizations. A number of OFC exhibitors looked to address this issue in different ways. For example, Cisco rolled out its Open Transport 3000 Series, a multi-rail open line system that integrates optical components for multiple fiber rails into a single line card, providing improvements in power and density for hyperscalers, neocloud operators and very high-end enterprise AI applications, the vendor stated.
