Exploiting Read/Write Asymmetry to Achieve Opportunistic SRAM Voltage Switching in Dual-Supply Near-Threshold Processors

Yunfei Gu, Dengxue Yan, Vaibhav Verma, Pai Wang, Mircea R. Stan, Xuan Zhang

TL;DR

This paper tackles the energy efficiency challenge in near-threshold computing (NTC) by exploiting the asymmetric reliability of SRAM cells. They propose a dual-supply voltage switching scheme that dynamically lowers memory voltage during read operations and uses a write aggregation buffer to minimize performance penalties, achieving significant energy savings and performance improvements.

What Problem Does This Paper Address?

Modern processors are increasingly focused on energy efficiency, especially for applications like green data centers and the Internet of Things. Near-threshold computing (NTC) is a promising technique to achieve this by lowering the supply voltage (VDD) close to the transistor's threshold voltage. However, NTC presents a dilemma: logic cores and on-chip memory (like SRAM) have different optimal supply voltages for peak energy efficiency and reliable operation. In a single-supply NTC system, a compromise must be made, leading to suboptimal performance and energy usage for one or both components. Specifically, memory often dictates a higher VDD than ideal for the logic core, wasting energy. Furthermore, NTC can degrade single-thread performance due to increased critical path delays. Dual-supply architectures have been proposed to address this by allowing temporary voltage boosts, but this paper aims to further exploit energy savings by leveraging a specific characteristic of SRAM.

Methods

The core of this paper lies in identifying and exploiting the asymmetric reliability of 6T SRAM cells at near-threshold voltages. Through circuit-level simulations, the authors found that SRAM read operations can tolerate significantly lower supply voltages than write operations while maintaining reliability.

Based on this insight, they propose an opportunistic dual-supply switching scheme for the memory subsystem. This scheme dynamically switches the memory's supply voltage:

  • Lower voltage (VL): Used during read-only operations to save power.
  • Higher voltage (VH): Used during write operations to ensure reliability.

To mitigate the performance overhead associated with frequent voltage switching (especially the transition from VL to VH), they introduce a Write Aggregation Buffer (WAB) and Write Back Aggregation Buffer (WBAB). These buffers accumulate write operations, delaying their release to the memory hierarchy. This strategy reduces the frequency of write operations, thereby decreasing the number of disruptive read-to-write transitions and allowing the memory to stay at the lower voltage (VL) for longer consecutive read periods. The system also includes voltage comparators and control logic to manage these transitions.

Datasets

The authors evaluated their proposed technique using:

  • SPEC2006 benchmarks: A suite of general-purpose computational-intensive workloads.
  • PARSEC benchmarks: A suite of data-intensive workloads.

They used the Gem5 simulator to generate cache traces and McPAT for power analysis. Circuit-level simulations were performed using a 16 nm FinFET technology process and a 7 nm ASAP PDK.

Key Insights and Ideas

  • SRAM Read/Write Asymmetry is Key: The most significant insight is that SRAM read operations are much more tolerant to low voltages than write operations. This asymmetry is not just a minor difference but a substantial opportunity for energy savings.
  • Voltage Switching is Powerful, but Costly: Dynamically switching voltage can save energy, but the transition overhead (latency and energy) can negate the benefits if not managed carefully.
  • Aggregation is Crucial for Performance: The write aggregation buffer is vital. It doesn't just store writes; it strategically delays them to create longer sequences of read operations, maximizing the time the memory can operate at the energy-efficient low voltage.
  • Workload Dependency: The potential for energy savings varies significantly across different benchmarks, with data-intensive workloads (like PARSEC) showing more promise due to their inherent longer consecutive read sequences.

Key Results

The proposed dual-supply switching scheme with write aggregation buffers achieved:

  • Energy Efficiency Improvement: More than 21.45% on average across various benchmarks.
  • Performance Speed-up: Approximately 10.19% on average.

The authors also showed that the effectiveness of the WAB/WBAB is influenced by buffer size, with larger buffers generally leading to better Long Consecutive Read Sequence Potential (LCRSP) improvements, though saturation effects were observed.

Limitations and Future Work

  • Transition Overhead: While the WAB/WBAB significantly mitigates the latency and energy overhead of voltage switching, some overhead still exists due to combinational logic and bypass mechanisms.
  • Workload Dependence: The energy savings are not uniform across all benchmarks, indicating that workload characteristics play a crucial role in the effectiveness of the proposed technique.
  • Future Work: The paper suggests that this technique can be combined with existing voltage boosting capabilities for NTC processors, implying further potential for optimization. The authors also highlight the need to consider the impact of process variation and temperature fluctuations on the reliability of these low-voltage operations.

Summary updated

Abstract

Energy-efficient microprocessors are essential for a wide range of applications. While near-threshold computing is a promising technique to improve energy efficiency, optimal supply demands from logic core and on-chip memory are conflicting. In this paper, we perform static reliability analysis of 6T SRAM and discover the variance among different sizing configuration and asymmetric minimum voltage requirements between read and write operations. We leverage this asymmetric property i n near-threshold processors equipped with voltage boosting capability by proposing an opportunistic dual-supply switching scheme with a write aggregation buffer. Our results show that proposed technique improves energy efficiency by more than 21.45% with approximate 10.19% performance speed-up.
BibTeX
@article{gu2018exploiting,
  author = {Yunfei Gu and Dengxue Yan and Vaibhav Verma and Pai Wang and Mircea R. Stan and Xuan Zhang},
  title = {{Exploiting Read/Write Asymmetry to Achieve Opportunistic SRAM Voltage Switching in Dual-Supply Near-Threshold Processors}},
  journal = {Journal of Low Power Electronics and Applications},
  year = {2018},
  doi = {10.3390/jlpea8030028},
}
Figure