NBTI-aware design of NoC buffers

Davide Zoni, William FornaciariIna-ocmc@hipeac 2013 · DOI

TL;DR

This paper tackles the challenge of NBTI (Negative Bias Temperature Instability) degradation in Network-on-Chip (NoC) buffers, a critical component for modern multi-core processors. They propose two power-gating-based techniques to mitigate this degradation, demonstrating significant improvements in buffer reliability with minimal area overhead, even in the presence of process variations.

What Problem Does This Paper Address?

As chip technology scales down, devices become more susceptible to reliability issues like NBTI. NBTI causes a gradual increase in the threshold voltage (Vth) of PMOS transistors, leading to performance degradation and reduced device lifetime. This is particularly problematic for NoC buffers, which are constantly active and can experience prolonged periods of stress. Furthermore, process variations (PV) during manufacturing introduce unpredictable differences in device characteristics, exacerbating these reliability concerns. Existing solutions often don't fully account for PV or the specific traffic patterns within NoCs, leaving a gap in effectively mitigating NBTI in these crucial components.

Methods

The authors propose two NBTI-aware mitigation strategies that leverage power-gating to reduce the stress period on NoC buffers. Power-gating allows selectively shutting off power to idle buffers, allowing them to enter a recovery state and partially reverse NBTI-induced Vth degradation.

  • Sensor-less Round-Robin Approach (rr-no-sensor): This method cyclically selects buffers to be powered down for recovery, irrespective of their degradation level. It relies on the upstream router to signal which buffer should remain active for incoming traffic, while others can be powered off. This approach is simpler as it doesn't require dedicated NBTI sensors.
  • Sensor-wise Approach: This technique utilizes NBTI sensors within the downstream router to identify the most degraded buffer. The upstream router then uses this information to prioritize the recovery of that specific buffer, while still allowing other idle buffers to recover. This approach aims for more targeted and efficient NBTI mitigation.

Both techniques account for the power-gating delay overhead and are evaluated considering process variations.

Datasets

The paper evaluates their proposed techniques using simulations based on the GEM5 simulator with the Garnet network model. They simulate multi-core architectures (4-core and 16-core 2D-mesh) with varying injection rates and virtual channel configurations. The technology parameters are based on a 45nm node. They also mention using real traffic patterns (though not detailed due to space limitations) in addition to synthetic traffic.

Key Insights and Ideas

  • Power-gating is a viable NBTI mitigation strategy for NoC buffers: By strategically powering down idle buffers, significant recovery from NBTI degradation can be achieved.
  • Process variation significantly impacts NBTI mitigation effectiveness: The proposed techniques are designed to be robust against PV, which is crucial for real-world applications.
  • Targeted recovery is more effective: The sensor-wise approach, which identifies and prioritizes the most degraded buffers, generally yields better results than the sensor-less round-robin approach, especially when process variations are significant or many virtual channels are present.
  • Traffic patterns influence mitigation strategy choice: While sensor-wise is generally better, the gap between the two approaches narrows with increased traffic load, as fewer buffers remain truly idle.

Key Results

  • The proposed techniques achieve a net NBTI Vth saving of up to 54.2% compared to a baseline NoC that doesn't account for NBTI.
  • The area overhead for the proposed techniques is below 5%, with specific implementations achieving less than 1% and 3.20% for the sensor-less and sensor-wise approaches, respectively.
  • The sensor-wise policy consistently achieves lower stress period values than the rr-no-sensor policy.
  • The stress gap between sensor-wise and rr-no-sensor decreases with increasing traffic load but increases with the number of virtual channels.

Limitations and Future Work

  • Optimistic Power Gating Assumptions: The analysis of power gating overhead assumes an ideal predictor for idle intervals and a low wake-up latency. Real-world predictors might introduce delays, impacting the effectiveness.
  • Limited Scope of Analysis: While the paper discusses both synthetic and real traffic, the detailed results presented are primarily for synthetic traffic patterns.
  • Future Work: The authors suggest further investigation into the impact of NBTI sensors on power consumption and the exploration of more advanced NBTI mitigation techniques that can adapt to dynamic traffic conditions. They also highlight the need for more comprehensive analysis of real traffic patterns.

Summary updated

BibTeX
@inproceedings{zoni2013nbti,
  author = {Davide Zoni and William Fornaciari},
  title = {{NBTI-aware design of NoC buffers}},
  booktitle = {Ina-ocmc@hipeac},
  year = {2013},
  doi = {10.1145/2482759.2482766},
}
Figure