Abstract
Error correcting codes (ECCs) are commonly used to protect against soft errors. Single error correcting and double error detecting (SEC-DED) codes are generally used for this purpose. This paper proposes a design method for SEC-DED checker circuits that selectively reduces power consumption, delay, and area. By employing a multi-objective genetic algorithm (MOGA), the design problem is framed as a non-linear optimization. This approach provides a set of non-dominated solutions, allowing users to choose the optimal ECC configuration based on their specific constraints for power, delay, and area. The method is validated using the odd-column weight Hsiao code, demonstrating its effectiveness in generating efficient hardware implementations.
Introduction: Addressing Soft Errors in Memory Systems
As semiconductor technology scales to smaller feature sizes and lower voltages, memory systems become increasingly susceptible to "soft errors"—bit flips caused by cosmic rays or electromagnetic interference. While Error Correcting Codes (ECCs) like SEC-DED are standard solutions, the checker circuits required to implement them introduce overhead in terms of power, silicon area, and latency.
Traditionally, designing these circuits involves trade-offs that are difficult to balance manually. Most existing methods focus on optimizing a single metric or use fixed weightings for different objectives. This research addresses the need for a flexible design framework that can explore the entire trade-off space between power, area, and delay for memory ECC checkers.
Methodology: Multi-objective Genetic Algorithm (MOGA)
The authors focus on Hsiao codes (a type of odd-column weight SEC-DED code) and optimize the construction of the parity-check matrix ($H$-matrix). The $H$-matrix defines how parity bits are generated and checked. A Multi-objective Genetic Algorithm (MOGA) is used to maintain a population of potential $H$-matrices, evolving them to find a set of non-dominated solutions.
A key technical component is the novel chromosome encoding, which represents the $H$-matrix in two parts: a permutation part (mapping memory bits to columns) and a selection part (choosing which high-order odd-weight columns to include).
To maintain validity during evolution, the algorithm uses specialized operators: cycle crossover for the permutation segment and arithmetic crossover for the selection segment. Furthermore, the fitness function weights for power, delay, and area are randomized in each generation to ensure thorough exploration of the Pareto front. The method was evaluated using 64-bit memory data sets with varying bitwise profiles, reflecting typical memory usage patterns where switching activity—and thus power consumption—varies across bits.
Results: Navigating the Pareto Front
The primary innovation of this approach is the shift from providing a single design to providing a "menu" of optimal designs, known as non-dominated solutions. In a multi-objective space, a solution is non-dominated if no single objective can be improved without degrading another.
The MOGA successfully identified multiple solutions that outperformed previous state-of-the-art methods. For a 64-bit architecture, the tool produced a variety of configurations. For example, it found solutions reducing power consumption by roughly 5-10% compared to standard implementations. While area (gate count) remained relatively stable, the GA minimized variance between XOR tree depths, leading to more predictable timing. This flexibility allows a designer for a mobile device to select a low-power solution, while a server designer might prioritize the lowest possible delay.
Limitations and Future Directions
While effective, genetic algorithms are computationally expensive, and the paper does not deeply discuss convergence times for larger memory widths (e.g., 128-bit). Additionally, the power and delay estimations rely on simplified gate-level models; real-world ASIC or FPGA implementations might introduce routing overheads not fully captured here. Future work could extend this multi-objective approach to other code types or additional hardware constraints like peak current or soft-error masking rates.
Summary updated