互连
光子学
计算机科学
计算机体系结构
材料科学
光电子学
电信
作者
Guanglong Li,Yaoyao Ye
标识
DOI:10.1109/tcad.2023.3328828
摘要
In pursuit of higher inference accuracy, the complexity and parameter size of recent deep neural networks (DNNs) have increased significantly. Due to the increasing demand for computing power, the chiplet-based accelerator has been an important computing platform that can handle these DNN models more efficiently. In widely used DNN models, the feature sizes and the number of channels vary greatly among different convolutional layers. Existing chiplet-based accelerators typically adopt consistent optimization strategy for all of the convolutional layers regardless of their sizes, which would limit the inference performance. In this work, we carry out communication-aware customized optimization for convolutional layers with different sizes. First, we propose a reconfigurable high-performance photonic interconnect (HPPI) architecture to facilitate the communication in chiplet-based DNN accelerators. Second, we propose a customized dataflow as the mapping framework and provide four communication patterns of the photonic interconnect with different ways of spatial mapping. Third, we propose a lightweight back propagation neural network to efficiently select the optimal communication pattern for each convolutional layer. The proposed photonic interconnect can be switched between the four communication patterns to enable communication-aware customized optimization for each convolutional layer in the DNN model. As compared to Simba (a representative chiplet-based accelerator with electronic interconnect), HPPI reduces the execution time by 72.23% on average, while saving the energy consumption by 25.49% on average. As compared to ASCEND (a state-of-the-art chiplet-based accelerator with photonic interconnect), HPPI reduces the execution time by 34.04% on average, while saving the energy consumption by 10.34% on average.
科研通智能强力驱动
Strongly Powered by AbleSci AI