DALA: Distribution-Agnostic Level Allocation for Multiple Bits-Per-Cell RRAM​

Speaker: Anjiang Wei, PhD Student, Stanford University
Date: December 14, 2022

Recently, researchers have demonstrated multiple-bit-per-cell (MBPC) data storage in resistive memories. In MBPC storage, a level allocator identifies an allocation that maps resistance ranges to bit sequences. State-of-the-art level allocators, such as SBA, deploy technology-specific algorithms that use statistical measures (e.g., variance) extracted from characterization data. These approaches discard important analog behaviors and cannot be easily generalized to other memory technologies. We present DALA, a distribution- and technology-agnostic level allocation algorithm capable of working with highly irregular resistance distributions. We show that DALA level allocations have 29.6%-71.0% lower bit-error rates and 21.6%-40.9% lower ECC storage overheads than SBA on a fabricated RRAM storage array.