Design and Analysis of an Adjacent Multi-bit Error Correcting Code for Nanoscale SRAMs
MetadataShow full item record
Increasing static random access memory (SRAM) bitcell density is a major driving force for semiconductor technology scaling. The industry standard 2x reduction in SRAM bitcell area per technology node has lead to a proliferation in memory intensive applications as greater memory system capacity can be realized per unit area. Coupled with this increasing capacity is an increasing SRAM system-level soft error rate (SER). Soft errors, caused by galactic radiation and radioactive chip packaging material corrupt a bitcell’s data-state and are a potential cause of catastrophic system failures. Further, reductions in device geometries, design rules, and sensitive node capacitances increase the probability of multiple adjacent bitcells being upset per particle strike to over 30% of the total SER below the 45 nm process node. Traditionally, these upsets have been addressed using a simple error correction code (ECC) combined with word interleaving. With continued scaling however, errors beyond this setup begin to emerge. Although more powerful ECCs exist, they come at an increased overhead in terms of area and latency. Additionally, interleaving adds complexity to the system and may not always be feasible for the given architecture. In this thesis, a new class of ECC targeted toward adjacent multi-bit upsets (MBU) is proposed and analyzed. These codes present a tradeoff between the currently popular single error correcting-double error detecting (SEC-DED) ECCs used in SRAMs (that are unable to correct MBUs), and the more robust multi-bit ECC schemes used for MBU reliability. The proposed codes are evaluated and compared against other ECCs using a custom test suite and multi-bit error channel model developed in Matlab as well as Verilog hardware description language (HDL) implementations synthesized using Synopsys Design Compiler and a commercial 65 nm bulk CMOS standard cell library. Simulation results show that for the same check-bit overhead as a conventional 64 data-bit SEC-DED code, the proposed scheme provides a corrected-SER approximately equal to the Bose-Chaudhuri- Hocquenghem (BCH) double error correcting (DEC) code, and a 4.38x improvement over the SEC-DED code in the same error channel. While, for 3 additional check-bits (still 3 less than the BCH DEC code), a triple adjacent error correcting version of the proposed code provides a 2.35x improvement in corrected-SER over the BCH DEC code for 90.9% less ECC circuit area and 17.4% less error correction delay. For further verification, a 0.4-1.0 V 75 kb single-cycle SRAM macro protected with a programmable, up-to-3-adjacent-bit-correcting version of the proposed ECC has been fab- ricated in a commercial 28 nm bulk CMOS process. The SRAM macro has undergone neu- tron irradiation testing at the TRIUMF Neutron Irradiation Facility in Vancouver, Canada. Measurements results show a 189x improvement in SER over an unprotected memory with no ECC enabled and a 5x improvement over a traditional single-error-correction (SEC) code at 0.5 V using 1-way interleaving for the same number of check-bits. This is compa- rable with the 4.38x improvement observed in simulation. Measurement results confirm an average active energy of 0.015 fJ/bit at 0.4 V, and average 80 mV reduction in VDDMIN across eight packaged chips by enabling the ECC. Both the SRAM macro and ECC circuit were designed for dynamic voltage and frequency scaling for both nominal and low voltage applications using a full-custom circuit design flow.