UC Irvine

Light-driven computing systems are closing the gap with state-of-the-art digital accelerators in terms of energy efficiency, compute throughput, density and latency. Here we map where photonic accelerators can matter in modern artificial intelligence infrastructure, and what must change for their impact at scale. We distinguish cloud-oriented, general-purpose photonic accelerators from application-specific, edge-oriented systems. In the cloud, the central challenge is scalability under tight energy budgets: heterogeneous systems with large vector inputs suffer from losses and electro–optic interfaces that dominate power and impede throughput, preventing the commercial deployment of photonic tensor cores. At the edge, opportunities arise wherever ultralow latency or high spatial parallelism are required and tight energy constraints dominate—for instance, in on-fibre optical processing and vision—yet progress is limited by the underwhelming scalability of all-optical nonlinearities, the lack of efficient reconfigurability, and a sizable optics footprint. We conclude that photonics is a viable near-term strategy within the existing digital ecosystem, and with focused advances along these axes, it can enable more sustainable artificial intelligence and open the door to broader adoption over the next decade.