Advanced Micro Devices Inc. today announced that it has acquired Enosemi Inc., a low-profile startup focused on developing photonics chips.

The terms of the deal were not disclosed. AMD did, however, detail that the acquisition will support internal engineering initiatives focused on artificial intelligence hardware.

Large-scale AI clusters comprise multiple server racks. Those racks, in turn, include not only graphics cards but also networking gear, cooling equipment and a range of other components. Historically, companies had to assemble the individual components manually into server racks.

AMD rival Nvidia Corp. offers a line of systems called the DGX series that eases the task. DGX appliances combine graphics cards, network equipment and the other hardware needed to run AI models into an integrated package. That removes the need for customers to assemble everything on their own.

The chipmaker reportedly plans to take a similar approach as Nvidia. Last year, it spent $4.8 billion to acquire ZT Systems Inc., a company that specializes in designing rack-scale data center systems. The purchase of Enosemi will support AMD’s efforts in this area.

Turning the components of a server cluster into a single functioning system requires connecting them over a network. In AI-optimized server clusters, that network is often based on fiber-optic cables. Sacramento-based Enosemi has developed a series of chips for powering fiber-optic data center networks.

“As AI models grow larger and more complex, the need for faster, more efficient data movement is accelerating,” Brian Amick, AMD’s senior vice president of technology and engineering, wrote in a blog post today. “To meet these evolving demands, particularly at the rack scale, optical interconnects offer a compelling path forward.”

Enosemi will support AMD’s efforts to develop co-packaged optics, or CPO, technology. This is an emerging type of hardware that can boost efficiency of AI clusters’ fiber-optic network equipment.

Servers turn data into light before sending it over a fiber-optic network and then turn it back into electrical signals. This task is performed with the help of devices called pluggable transceivers. The more servers there are in an AI cluster, the more transceivers are required. This means that the latter devices can represent a major expense for companies building large-scale AI infrastructure. 

CPO technology removes the need for pluggable transceivers. It does so by integrating a transceiver into the chips that power an AI cluster’s switches. In addition to reducing hardware costs, this arrangement can boost network performance. When a transceiver is integrated directly into a switch, data takes less time to travel between the transceiver and the switch’s other components, which lowers latency.

“Enosemi has collaborated with us as an external development partner on photonics, and this acquisition extends that successful relationship,” Amick wrote. “Now as part of AMD, the team will help us immediately scale our ability to support and develop a variety of photonics and co-packaged optics solutions across next-gen AI systems.”

Image: AMD