Hardware/Software Codesign for Energy Efficiency and Robustness: From Error-Tolerant Computing to Approximate Computing

Abbas Rahimi, Rajesh K. GuptaEmbedded 2020 · DOI · OpenAlex

Abstract

Voltage scaling, as the most important knob for energy efficiency, is limited by leakage and variability. Variability arises from various sources including static manufacturing process, dynamic voltage and temperature fluctuations, and temporal changes over time. To address these variations, designers resort to excessive margins. These margins are increasing rapidly and eventually obliterating any gains due to device scaling. As a consequence, reduction of margins in design has become an important research challenge. We demonstrate how to recover part of these margins through hardware/software codesign with examples in many-core GPUs and FPGAs. This naturally leads to a departure from traditional error-tolerant computing to approximate computing.

Introduction: The Crisis of Excessive Margins

As CMOS technology scales into the nanometer regime, hardware reliability becomes increasingly volatile. Modern chips face three primary categories of variability: static manufacturing defects, temporal aging effects (such as Negative Bias Temperature Instability), and dynamic fluctuations in operating conditions like temperature and supply voltage.

Traditionally, the industry has countered these uncertainties through "worst-case" design. By implementing excessive safety margins—running processors at higher voltages or lower frequencies than required for typical operation—engineers ensure functional correctness even under extreme, rare conditions. However, these margins have grown so large that they now consume the energy efficiency gains promised by Moore's Law. This paper proposes a paradigm shift: instead of overdesigning hardware, we should "underdesign" it and leverage hardware/software codesign to manage the resulting errors.

Methodology: A Taxonomy of Error Tolerance

The authors categorize cross-layer contributions to error management using a "Clockwise Y-Chart" taxonomy. This framework spans from predicting and preventing errors to accepting them as a trade-off for efficiency.

Figure 4: The Clockwise Y-Chart taxonomy of error tolerance, moving from positive margins (predict/prevent) to negative margins (accepting errors).
  1. Positive Margin (Predict & Prevent): This approach uses software to anticipate hardware failure. For GPUs, the authors developed an aging-aware compiler. By monitoring threshold voltage ($V_{th}$) shifts via hardware sensors, the compiler identifies "fatigued" Processing Elements (PEs). It then exploits the inherent underutilization of VLIW (Very Long Instruction Word) slots to rotate workloads, shifting instructions to "younger" PEs to equalize wear and extend the chip's lifetime.
  2. Zero Margin (Detect & Correct): This middle ground involves operating at the edge of failure and using architectural or software mechanisms to catch and fix errors as they occur.
  3. Negative Margin (Accepting Errors): This is the domain of Approximate Computing. For FPGAs, the authors introduced "Grater," a workflow that uses a Genetic Algorithm to identify non-critical code segments that can tolerate reduced precision. By shrinking operation bit-widths, the hardware footprint of kernels is reduced, allowing for higher parallel density.

Case Studies and Results

GPU Reliability through Adaptive Execution

The adaptive compiler was tested on the AMD Evergreen (Radeon HD 5870) architecture. By equalizing the workload across PEs based on their health, the system achieved a 34% reduction in $V_{th}$ shifts. This effectively mitigates the "weakest link" problem where a single prematurely aged PE limits the entire device's lifespan.

Figure 5: Comparison between naive and adaptive healthy kernel execution on GPUs, showing how workload equalization reduces threshold voltage shifts and recovers margins.

FPGA Throughput via Disciplined Approximation

Using data-intensive OpenCL benchmarks (e.g., Sobel filters) on a Stratix V Altera FPGA, the Grater workflow demonstrated that "area is the new speed." By reducing precision in a way that maintained quality loss below 1%, the researchers were able to fit significantly more kernel instances on the same silicon. This resulted in throughput improvements ranging from 1.4x to 3.0x.

Discussion and Limitations

While hardware/software codesign offers a path out of the "margin trap," it introduces new complexities. The GPU aging mitigation strategy relies on per-PE NBTI sensors, which are currently not standard in commercial hardware. Furthermore, approximate computing is best suited for "error-resilient" domains like image processing, signal processing, and machine learning. Applying these techniques to general-purpose control logic remains a challenge, as a single bit-flip in a pointer or branch condition could lead to system failure.

Conclusion

The transition from error-tolerant computing to approximate computing represents a fundamental shift in how we view hardware reliability. By treating hardware errors as a software-manageable resource rather than a catastrophic failure, we can reclaim the energy efficiency and performance scaling that traditional "worst-case" design has surrendered.

Summary updated

Abstract

Abstract Voltage scaling, as the most important knob for energy efficiency, is limited by leakage and variability. Variability is arisen from various sources including static manufacturing process, dynamic voltage and temperature fluctuations, and temporal changes over time. To address these variations, designers resort to excessive margins. These margins are increasing rapidly and eventually obliterating any gains due to device scaling. As a consequence, reduction of margins in design has become an important research challenge. We demonstrate how to recover part of these margins through hardware/software codesign with examples in many-core GPUs and FPGAs. This naturally leads to a departure from traditional error-tolerant computing to approximate computing.
BibTeX
@inproceedings{rahimi2020hardware,
  author = {Abbas Rahimi and Rajesh K. Gupta},
  title = {{Hardware/Software Codesign for Energy Efficiency and Robustness: From Error-Tolerant Computing to Approximate Computing}},
  booktitle = {Embedded},
  year = {2020},
  doi = {10.1007/978-3-030-52017-5\_22},
}
Figure