TL;DR
This paper introduces Molecular-scale Network-on-Chip (mNoC), a novel nanophotonic interconnect architecture that replaces traditional ring resonators and external lasers with quantum dot LEDs and chromophores. This innovation promises significantly reduced latency and power consumption, enabling more scalable and efficient on-chip communication for future multi-core processors.
What Problem Does This Paper Address?
As the number of cores on microprocessors continues to grow, traditional electrical interconnects (wires) face limitations in terms of power consumption and latency, especially for long distances across the chip. While nanophotonic networks-on-chip (NoCs) offer a promising alternative, existing designs (ring resonator-based NoCs or rNoCs) still suffer from high static power consumption due to inefficient external laser sources and the power required for tuning ring resonators. Furthermore, rNoCs face scalability challenges. This paper aims to overcome these limitations by exploring the use of emerging molecular-scale devices to create a more efficient and scalable nanophotonic NoC.
Methods
The core innovation of this paper lies in the replacement of key components in traditional nanophotonic NoCs with molecular-scale devices:
- Transmitters: Quantum Dot LEDs (QD_LEDs) replace the external laser source and ring modulators. QD_LEDs are small, efficient, and only emit light when a '1' bit is transmitted, significantly reducing static power.
- Receivers: Chromophores replace ring resonators for filtering and coupling optical signals to photodetectors. Chromophores are tiny, non-linear, and do not require trimming power. They absorb light at one wavelength and emit at another, enabling signal filtering and coupling to infrared photodetectors.
- Waveguides: The paper proposes using subwavelength-diameter silica waveguides, which are more compact than traditional silicon waveguides.
The authors explore how these molecular components can be integrated into various crossbar structures, including Single Writer Multiple Reader (SWMR), Multiple Writer Single Reader (MWSR), and Multiple Writer Multiple Reader (MWMR). They also analyze the implications of these mNoCs on router microarchitecture and system design, particularly for high-radix routers.
Datasets
Not discussed in the paper. The evaluation uses synthetic benchmarks and the PARSEC benchmark suite.
Key Insights and Ideas
- Power Efficiency through Traffic-Aware Emission: QD_LEDs only emit light when transmitting a '1', drastically cutting down on static power compared to continuously operating external lasers in rNoCs.
- Scalability without Trimming Power: Chromophores eliminate the need for power-hungry ring resonator tuning, a significant bottleneck in rNoC scalability.
- Cost-Effective Broadcast: The mNoC architecture supports broadcast and multicast operations with little to no extra cost, a significant advantage for cache coherence protocols.
- High-Radix Router Enabler: The inherent scalability and efficiency of mNoC make it a strong candidate for building high-radix routers, which are crucial for future many-core systems.
- Simplified Arbitration: For MWSR and MWMR, mNoC proposes using a special wavelength as a busy signal, simplifying arbitration compared to token-based schemes in rNoCs.
Key Results
- Latency Reduction: Preliminary evaluations show that a 256-node SWMR mNoC crossbar reduces average packet latency by over 50% compared to ring-based alternatives.
- Power Consumption Reduction: The same mNoC design reduces average power consumption by 40% compared to a radix-64 clustered rNoC.
- Scalability: The mNoC architecture, even with a waveguide loss of 2dB/cm, can scale up to a radix-256 crossbar.
- Traffic Tolerance: mNoC demonstrates higher tolerance to traffic, especially under hotspot traffic patterns, outperforming rNoC and even electrical mesh networks.
- PARSEC Benchmark Performance: While speedup is bounded by communication fractions in benchmarks like PARSEC, mNoC still shows significant latency advantages, being approximately one-third the latency of rNoC and less than one-third of Mesh.
Limitations and Future Work
- Experimental Validation: While the paper mentions building a prototype using chromophores on optical fibers and demonstrating filtering and evanescent coupling, detailed experimental results are not presented due to space limitations.
- Cache Coherence Protocol Exploration: The authors acknowledge that further research is needed to determine which broadcast-based cache coherence protocols best leverage mNoC's capabilities. Developing customized protocols is an ongoing effort.
- MWSR and MWMR Evaluation: The paper focuses its preliminary evaluation on SWMR crossbars, with MWSR and MWMR evaluations planned for future work.
- Complex Optics for MWMR: Implementing MWMR in mNoC might require more complex optics to inject the right amount of power into two directions for writers.
Summary updated