Photonic-accelerated mixed-precision computing

Modern AI and scientific computing demand performance beyond the limits of conventional electronics. Photonic computing offers ultrahigh bandwidth and intrinsic parallelism, but its adoption has been hindered by limited analog precision, especially at high operating speeds, due to electro-optic nonidealities and noises. We present a mixed-precision optoelectronic computing architecture that overcomes these limitations through hardware-software co-design. Using lithium-niobate photonics, we demonstrate coherent amplitude- and phase-based computation at up to 128 GS/s, achieving near-terascale operations per second per channel. An inner–outer iterative framework enables 12-bit-accurate solutions of partial differential equations, providing a scalable path toward energy-efficient, high-performance photonic computing.