Exploiting Memory Soft Redundancy for Joint Improvement of Error Tolerance and Access Efficiency

Shuo Wang,Lei Wang
DOI: https://doi.org/10.1109/TVLSI.2008.2001743
2009-08-01
IEEE Transactions on Very Large Scale Integration (VLSI) Systems
Abstract:Technology roadmap projects nanoscale multibillion- transistor integration in the coming years. However, on-chip memory becomes increasingly exposed to the dual challenges of device-level reliability degradation and architecture-level performance gap. In this paper, we propose to exploit the inherent memory soft (transient) redundancy for on-chip memory design. Due to the mismatch between fixed cache line size and runtime variations in memory spatial locality, many irrelevant data are fetched into the memory thereby wasting memory spaces. The proposed soft-redundancy allocated memory detects and utilizes these memory spaces for jointly achieving efficient memory access and effective error control. A runtime reconfiguration scheme is also proposed to further enhance the soft-redundancy allocation. Simulation results demonstrate 74.8% average error-control coverage ratio on the SPEC CPU2000 benchmarks with average of 59.5% and 41.3% reduction in memory miss rate and bandwidth usage, respectively, as compared to the existing memory techniques. Furthermore, the proposed technique is fully scalable with respect to various memory configurations.
Computer Science
What problem does this paper attempt to address?