Budget the complete optical communication path

Wavelength-division multiplexing places independently encoded optical channels on the same physical path under a designed transceiver system. Count the supported channels, symbol rate, bits represented per symbol, and payload efficiency. Do not multiply by a channel count already included in an advertised aggregate. A bidirectional total can combine two opposite one-way rates, so label which rate serves the actual communication phase. Directional workloads cannot consume unused reverse-direction capacity automatically.

A usable rate also depends on errors, coding, framing, and the receiver’s operating conditions. More bits per symbol require the corresponding signal quality and implementation; they are not a free arithmetic multiplier for an unchanged physical link. The original TeraPHY demonstration describes a particular electro-optic implementation and boundary. Its role here is to illustrate integrated conversion and multiplexing, while a workload estimate must still use the actual deployed transceiver, protocol, and sustained payload measurements.

A bulk transfer’s serialization time falls as payload bandwidth rises, but source queuing, electrical-optical conversion, optical propagation, reception, and completion logic can dominate short messages. State which terms are already included in a quoted endpoint latency. Fiber propagation depends on distance and group velocity; it is not zero because the carrier is light. Additional switches or regeneration can add terms whose inclusion depends on the topology and device design.

Optical circuit switching changes connectivity without making the matrix arithmetic itself optical. A reconfigured topology can improve placement or collective routes while introducing configuration and scheduling constraints. The TPU v4 paper studies a particular optically reconfigurable system; its application results remain attached to its workload, machine, and software. Separate that system-level evidence from a transceiver’s rate, and include topology changes or circuit setup where they are exposed on the application’s critical path.

Define whether power includes the laser source, transmitter drivers, receiver circuits, clock recovery, control, package interfaces, and allocated shared infrastructure. A device-only switching-energy number and a complete link power measurement are different quantities. Divide average included power by the useful delivered bit rate over the same interval, and include failed or retransmitted traffic according to the chosen boundary. High raw capacity does not guarantee low energy per useful bit when the application uses only a fraction of it.

Idle and shared-source power can make utilization decisive. If included power stays roughly constant while payload use falls, energy per useful bit rises even though the active channel’s device physics has not changed. Measure both directions and allocate shared power consistently when reporting a directional result. For an AI workload, retain conversion, communication, and compute timing together; faster links help only where the actual graph is communication-limited. State performance and energy as measured workload outcomes or explicitly conditional estimates.