Optical processing units vs. GPUs and TPUs: conditions for performance and energy advantage
Year: 2026
Type: article
Abstract: The conditions under which optical processing units (OPUs) can outperform conventional electronic accelerators, such as graphics processing units (GPUs) and tensor processing units (TPUs), are investigated. Because integrated system-level implementations remain unavailable, experimental validation is infeasible; a modeling-driven approach is therefore adopted to identify favorable configurations prior to fabrication. An analytical design space exploration framework is developed that extends roofline-based latency modeling to photonic cores and identifies total computational parallelism—core size squared times core count—as the governing parameter of OPU competitiveness. The latency model, validated against representative GPU and TPU measurements, enables efficient exploration of key variables, including core dimensionality, modulation frequency, and numerical precision. Optical loss, modulation bandwidth, and signal-to-noise limitations of integrated photonic platforms constrain the accessible design space. Physically realizable parameter regimes are identified in which OPUs achieve performance and energy-efficiency parity with, or advantage over, their electronic counterparts. These results establish the conditions under which photonic computing can provide a quantifiable benefit, clarify the corresponding architectural constraints, and yield concrete specifications for subsequent high-fidelity simulation and prototyping.
Source: Optics Express
DOI: 10.1364/oe.604976
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,270
OpenAlex ID: W7204549465