No abstract

Domain-specific neural network accelerators have seen growing interest in recent years due to their improved energy efficiency and inference performance compared to CPUs and GPUs. In this paper, we propose a novel cross-layer optimized neural network accelerator called CrossLight that leverages silicon photonics. CrossLight includes device-level engineering for resilience to process variations and thermal crosstalk, circuit-level tuning enhancements for inference latency reduction, and architecture-level optimization to enable higher resolution, better energy-efficiency, and improved throughput. On average, CrossLight offers 9.5× lower energy-per-bit and 15.9× higher performance-per-watt at 16-bit resolution than state-of-the-art photonic deep learning accelerators.

MEMS technology offers some unique product opportunities for optical networking applications. One of the critical factors that determine the commercial success of any optical MEMS components is the cost for packaging and optical fiber alignment. Small form factor MEMS components such as protection switches and variable optical attenuators require the integration of optical fibers and lenses with the MEMS silicon chip itself. This work uses a MEMS 2 × 2 optical switch as a test vehicle to demonstrate a fiber optic lens design, specifically engineered for use with the micromachined silicon chip, that enables the use of passive fiber optic lens alignment to significantly reduce component assembly cost.

The approximate computing paradigm advocates for relaxing accuracy goals in applications to improve energy-efficiency and performance. Recently, this paradigm has been explored to improve the energy efficiency of silicon photonic networks-on-chip (PNoCs). In this paper, we propose a novel framework (LORAX) to enable more aggressive approximation during communication over silicon photonic links in PNoCs. Given that silicon photonic interconnects have significant power dissipation due to the laser sources that generate the wavelengths for photonic communication, our framework attempts to reduce laser power overheads while intelligently approximating communication such that application output quality is not distorted beyond an acceptable limit. To the best of our knowledge, this is the first work that considers loss-aware laser power management and multilevel signaling to enable effective data approximation and energy-efficiency in PNoCs. Simulation results show that our framework can achieve up to 31.4% lower laser power consumption and up to 12.2% better energy efficiency than the best known prior work on approximate communication with silicon photonic interconnects, for the same application output quality.

No abstract