A shorter electrical path
Japanese telecom operator KDDI and networking-chip developer Broadcom began a joint optical-networking trial for AI data centers on October 22. Its central piece of hardware is the network switch, which directs data between connections. A test environment at KDDI’s AI data center is to combine Broadcom’s Tomahawk 6 switching chip with Davisson optical engines. An optical engine is a small component that converts electrical signals into light and back again. Placing it immediately beside the switching chip shortens the electrical transmission path; the companies aim to reduce the energy consumed by communication.[1]
That arrangement is the physical mechanism behind co-packaged optics, or CPO: optical components sit in the same package as the switching chip. In the network connecting graphics processors, data travels part of its route as an electrical signal and part as light. The design changes where that conversion occurs. For an infrastructure engineer, the useful starting point is this physical placement. Assessing its benefit under a workload also requires knowing how much data the switch actually carries. That makes the trial’s decision to examine network behavior alongside power consumption consequential.[1]
Useful traffic beside power consumption
KDDI plans to evaluate packet loss, congestion control and effective throughput. Packets are the small units of data carried across the network; congestion control regulates their flow under heavy traffic. The planned conditions cover distributed model training and inference, the operation of an already trained model. Switch power is therefore being tested alongside the communication demands of AI workloads. This scope gives the evaluation a practical purpose: establishing the amount of useful traffic the network can sustain as the partners pursue lower power consumption.[1]
Selecting just one of those measures would make the energy benefit difficult to interpret. My comparison would be energy consumed for useful data carried under the same workload. If a switch draws less power while throughput also falls, the benefit per unit of work could differ. Sustaining the same useful traffic with lower energy consumption would offer the operator a more direct gain. These are conditional engineering comparisons. KDDI has announced the measurements it intends to make, with no results yet establishing which outcome the trial delivers.[1]
Keeping the comparison at the network boundary
The system boundary also matters: the announced test concerns the switch and communication between graphics processors. KDDI supplies no energy-saving percentage measured in this joint trial. Extending a networking result to the entire data center would fold computing hardware and other facility loads into a measurement that does not cover them. At this stage, I would keep the comparison at the network boundary: how power consumption and useful traffic change together when optical conversion moves closer to the chip, under the stated training and inference conditions.[1]
KDDI intends to use the findings to establish an optical-electrical network architecture for its own AI data centers. The work begun today is a verification step connecting a test environment to a future network design. The announcement provides neither a commercial deployment schedule nor measured savings. The next concrete milestone I would look for is a set of results showing switch consumption alongside packet loss and effective throughput for the same workload. That would make the operating value of the shorter electrical path intelligible in terms of the work the network carries.[1]