Hybrid Nanophotonic NOC Design for GPGPU

Abstract

Due to the massive computational power, Graphics Processing Units (GPUs) have become a popular platform for executing general purpose parallel applications. The majority of on-chip communications in GPU architecture occur between memory controllers and compute cores, thus memory controllers become hot spots and bottle neck when conventional mesh interconnection networks are used. Leveraging this observation, we reduce the network latency and improve throughput by providing a nanophotonic ring network which connects all memory controllers. This new interconnection network employs a new routing algorithm that combines Dimension Ordered Routing (DOR) and nanophotonic ring algorithms. By exploring this new topology, we can achieve to reduce interconnection network latency by 17% on average (up to 32%) and improve IPC by 5% on average (up to 11.5%). We also analyze application characteristics of six CUDA benchmarks on the GPGPU-Sim simulator to obtain better perspective for designing high performance GPU interconnection network.