A Fully Optical Ring Network-on-Chip with Static and Dynamic Wavelength Allocation

Abstract

As the number of IPs (processor, memory) integration on a chip increases, chip multicore processors (CMPs), and system-on-chips (SoCs) will require high performance and low power consumption interconnection infrastructure. Traditional electronic network-on-chip (NoC) faces several problems, such as limited bandwidth, crosstalk, impedance mismatch, and huge power dissipation. To alleviate these challenges, optical NoCs have emerged as an attractive solution. Optical interconnects take advantage of light, and the multiple wavelengths within a single optical link (waveguide) to achieve high communication bandwidth at low power consumption cost. Toward this work we aim to propose a cost-performance and power efficient NoC. First, we proposed a low latency path setup network for conventional hybrid electronic-photonic Torus NoC using predictive switching. By lowering the path setup latency, we could achieve a considerable performance improvement. Second, a new hybrid architecture formed of an optical ring and electrical crossbar (OREX) has been proposed. OREX reduces the path setup network to a single electrical crossbar. Its optical network uses a ring topology more adapted for photonic interconnects. Using a cycle accurate simulator, our results show that OREX further improves hybrid electronic-photonic NoCs performance. Finally, to reduce power consumption, we have proposed a fully optical ring architecture. The proposed architecture combines static and dynamic wavelength allocation in the same network to fully take advantage of the low power and high performance optical interconnects. A different wavelength-channel is statically allocated to each destination node for light weight communication. Contention of simultaneous communication requests from multiple source nodes to the destination is solved by a token based arbitration. For heavy load communication, a shared multiwavelength-channel is available by requesting it in execution time from source node to a special node that manages dynamic wavelength allocation. Our architecture takes advantage of both wavelength allocation mechanisms by selecting the adequate one, depending on iv communication message sizes (baseline) and network congestion information (contention based and smart selection). Preliminary hardware cost comparison with several photonic architecture shows that our architecture can be an attractive costperformance interconnection infrastructure for future SoCs and CMPs. We further discuss performance of the proposed fully optical ring NoC architecture based on simulation using a photonic network simulator. Results show that our architecture allows considerable reduction of the network energy consumption compared to conventional hybrid NoCs and show reasonable bandwidth and latency performance using probabilistic traffic patterns.