Software adaptation in quality sensitive applications to deal with hardware variability

Aashish Pant, Puneet Gupta, Mihaela van der SchaarACM Great Lakes Symposium on Vlsi 2010 · DOI

TL;DR

This paper proposes a novel approach to mitigate the impact of manufacturing variations in hardware by adapting software algorithms at runtime. By using "hardware signatures" (post-manufacturing performance characteristics), applications can dynamically adjust their behavior to maintain quality and improve manufacturing yield, reducing the need for costly hardware over-design.

What Problem Does This Paper Address?

Modern semiconductor manufacturing processes are plagued by variations, meaning that even chips designed to be identical can exhibit significant differences in performance, power consumption, and leakage. These variations can lead to a substantial number of chips failing to meet their specifications, resulting in lost yield and increased costs. Traditionally, this has been addressed by "over-designing" hardware, essentially building in a performance buffer to accommodate these variations. However, this comes at a significant cost in terms of area, power, and development time. The paper addresses the challenge of managing these hardware variations without resorting to expensive over-design, particularly for quality-sensitive applications like video encoding, where performance and quality are paramount.

Methods

The core of the proposed solution is hardware-signature based software adaptation. The authors introduce the concept of "hardware signatures," which are measured post-manufacturing characteristics of a specific chip (e.g., frequency, power). These signatures are then fed to the software application, allowing it to adapt its algorithm.

The adaptation process is framed using Quality-Complexity (Q-C) curves. These curves illustrate the trade-off between application quality (e.g., PSNR for video) and complexity (e.g., processing time). Manufacturing variations cause these Q-C curves to shift. By knowing the hardware signature, the software can select an operating configuration that optimizes quality within the given constraints, even on a variation-affected chip.

For H.264 encoding, the hardware signatures are defined as frequency variations of key components (motion estimation, transform, entropy coding). The paper also proposes a method to determine optimal signature measurement points by mapping the problem to a shortest path problem on a graph, aiming to maximize the benefit gained from adaptation with a limited number of signature measurements.

Datasets

The paper uses the H.264 encoder as a proof of concept. They employ a standard mobile video sequence for their experiments. The Q-C curves are constructed based on profiling the encoder's performance across various configurations and simulated hardware variations.

Key Insights and Ideas

  • Software as a "Smart Tuner" for Hardware: The fundamental idea is to leverage software's flexibility to compensate for hardware's inherent variability, turning a potential weakness into a manageable characteristic.
  • Hardware Signatures are Key: The concept of using post-manufacturing hardware characteristics as "signatures" is crucial. This allows for a data-driven, per-chip adaptation rather than a one-size-fits-all approach.
  • Q-C Curves as a Unified Framework: The Q-C curve provides a powerful abstraction to visualize and analyze the impact of hardware variations on application quality and complexity, and how adaptation can mitigate these effects.
  • Optimizing Adaptation Granularity: Not all hardware variations need to be measured or adapted to. The paper shows that there's an optimal granularity for signature measurement, and a graph-based approach can find these optimal points to maximize adaptation benefits.
  • Yield and Quality are Intertwined: The work demonstrates that by adapting software, one can simultaneously improve manufacturing yield (fewer chips discarded) and maintain or even improve application quality, breaking the traditional trade-off.

Key Results

  • Manufacturing Yield Improvement: With hardware-aware adaptation, manufacturing yield can improve significantly. For example, at 0% over-design, yield increased from 50% to 90%.
  • Reduced Over-design: To achieve a target yield of 80%, the need for hardware over-design can be reduced by 10% points.
  • Enhanced Application Quality: For the H.264 encoder, at 80% yield, adaptation resulted in a 2.6dB improvement in average PSNR compared to the non-adaptive case.
  • Optimal Signature Sampling: The proposed method for determining optimal signature measurement points improved PSNR by about 1.3dB over naive sampling for the H.264 encoder.
  • Graceful Degradation: Adaptation allows for a more graceful degradation of PSNR as yield requirements increase, unlike the staircase-like degradation seen in non-adaptive systems.

Limitations and Future Work

  • Granularity Analysis for Multi-component Systems: The current work focuses on one-component hardware or components with perfectly correlated variations. The authors acknowledge that extending this granularity analysis to multi-component, pipelined, or multi-application systems (potentially mediated by an operating system) is an area for future work.
  • Dynamic Voltage Scaling (DVS) and Other Adaptation Policies: The paper hints at extending the adaptation strategies to other scenarios like DVS for power optimization and exploring optimal signature-dependent adaptation policy perturbations for adaptive applications.
  • Real-time Implementation and Overhead: While the paper discusses the concept and benefits, the practical implementation details and overhead of signature measurement and adaptation in real-time systems would be a crucial next step.
  • Wearout Mechanisms: The paper briefly mentions that signatures could be measured periodically to account for wearout mechanisms, but a detailed analysis of this aspect is not presented.

Summary updated

BibTeX
@inproceedings{pant2010software,
  author = {Aashish Pant and Puneet Gupta and Mihaela van der Schaar},
  title = {{Software adaptation in quality sensitive applications to deal with hardware variability}},
  booktitle = {ACM Great Lakes Symposium on Vlsi},
  year = {2010},
  doi = {10.1145/1785481.1785501},
}
Figure