Improving the thermal reliability of photonic chiplets on multicore processors

Abstract

Silicon photonics is a promising technology for high-performance inter/intra-chip networks. Although chiplet technologies help to integrate photonic network with processors and memories, thermal reliability remains a major challenge for photonic networks. Thermal variations are common in chiplet-based systems. Due to thermo-optic effects, photonic switches, such as microresonators (MRs), suffer from temperature-dependent wavelength shifts. To improve the thermal reliability of photonic chiplets, we propose an effective bridging and tuning method, called BAT. Our method enhances the temperature channel transition among MRs and minimize bit errors during temperature changes. BAT is integrated with thermal tuning, electrical tuning, laser tuning, and receiver tuning. We optimize MR design, control logic, tuning circuit, thermal floorplan, and network protection mechanism for BAT. Compared to state-of-the-art methods, BAT achieves a 10−12 bit-error-rate for photonic chiplets, saves up to 70.0% of the thermal tuning power, and improves up to a 1.17X system energy efficiency.