A Hybrid Photonic Burst-Switched Interconnection Network for Large-Scale Manycore System

Quanyou FENG  Huanzhong LI  Wenhua DOU  

IEICE TRANSACTIONS on Information and Systems   Vol.E95-D   No.12   pp.2908-2918
Publication Date: 2012/12/01
Online ISSN: 1745-1361
DOI: 10.1587/transinf.E95.D.2908
Print ISSN: 0916-8532
Type of Manuscript: Special Section PAPER (Special Section on Parallel and Distributed Computing and Networking)
Category: Computer Architecture
manycore system,  network on chip,  photonic burst switching,  flattened butterfly,  3D stacking,  

Full Text: PDF>>
Buy this Article

With the trend towards increasing number of cores, for example, 1000 cores, interconnection network in manycore chips has become the critical bottleneck for providing communication infrastructures among on-chip cores as well as to off-chip memory. However, conventional on-chip mesh topologies do not scale up well because remote cores are generally separated by too many hops due to the small-radix routers within these networks. Moreover, projected scaling of electrical processor-memory network appears unlikely to meet the enormous demand for memory bandwidth while satisfying stringent power budget. Fortunately, recent advances in 3D integration technology and silicon photonics have provided potential solutions to these challenges. In this paper, we propose a hybrid photonic burst-switched interconnection network for large-scale manycore processors. We embed an electric low-diameter flattened butterfly into 3D stacking layers using integer linear programming, which results in a scalable low-latency network for inter-core packets exchange. Furthermore, we use photonic burst switching (PBS) for processor-memory network. PBS is an adaptation of optical burst switching for chip-scale communication, which can significantly improve the power efficiency by leveraging sub-wavelength, bandwidth-efficient optical switching. Using our physically-accurate network-level simulation environment, we examined the system feasibility and performances. Simulation results show that our hybrid network achieves up to 25% of network latency reduction and up to 6 times energy savings, compared to conventional on-chip mesh network and optical circuit-switched memory access scheme.