Soft-Error Tolerant Cache Architectures

The problem of soft errors caused by radiation events are expected to get worse with technology scaling. This thesis focuses on mitigation of soft errors to improve the reliability of memory caches. We survey existing mitigation techniques and discuss their issues. We then propose 1) a technique that can mitigate soft errors in caches with lower costs than the widely-used Error Correcting Code (ECC), 2) a technique to mitigate soft errors in Content Addressable Memories, and 3) a cost-effective cache architecture achieving both variation-induced defect and soft-error tolerance. ECC is widely used to detect and correct soft errors in memory caches. Maintaining ECC on a per-word basis, which is preferred for caches with word-based access, is expensive. Chapter 3 proposes Zigzag-HVP, a costeffective technique to detect and correct soft errors for such caches. ZigzagHVP utilizes horizontal-vertical parity (HVP). Basic HVP can detect and correct a single bit error (SBE), but not a multi-bit error (MBE). By dividing the data array into multiple HVP domains and interleaving different domains, a spatial MBE can be converted to multiple SBEs, each of which can be detected and corrected by the corresponding parity domain. Vertical parity update and error recovery in Zigzag-HVP can be performed efficiently by modifications to the cache data paths, write-buffer, and Built-In Self Test. Evaluation results indicate that the area and power overheads of Zigzag-HVP caches are lower than those of ECC-based ones. Chapter 4 proposes STCAM, a soft-error tolerant Content-Addressable Memory (CAM). Soft-error mitigation in a CAM is difficult due to the unavailability of data outside the cell array in a CAM access. Since CAMs are used in several components of a processor, making those CAMs being resilient against soft errors is required to attain high processor’s reliability. STCAM can successfully detect and correct false hits and false misses caused by soft errors in a CAM. This is achieved through subdividing a CAM and