Fault tolerance of L1 data cache memory induced by intrinsic parameters fluctuation in sub 10nm UTB-SOI MOSFETs

Currently, the development of models at higher level of abstractions (system-level) to be able to incorporate effects at lower levels of abstractions (process /transistor) is in demand. This thesis addresses issues to enabling computer system simulation model in the presence of cell failures in L1 data cache corresponding to the impact of Intrinsic Parameters Fluctuation (IPF). These time-independent transistor-level sources of variation are randomly characterized in nature. This makes it difficult for the designer to include IPF impact in the design plan to overcome. This computer model is vital to analyze and evaluate credibly the effectiveness of L1 cache fault tolerance techniques in controlling the implications of IPF cell failures on microprocessor reliability and yield. The objectives of this thesis are (i) to devise a framework to simulate system-level environment in the presence of L1 data cache cell failures corresponding to the impact of IPF, (ii) to introduce an evaluation method for deduce the effectiveness of L1 cache fault tolerance techniques in handling the actual error pattern caused by IPF cell failures in computer system under test and workload conditions, and (iii) to investigate the implications of L1 data cache faults induced by the individual and combined impact of IPF sources on reliability of a general-purpose microprocessor. The case study of this thesis is the impact of cell failures in the data array of L1 data cache in Intel Strong ARM@SA-1110 microprocessor. The failure models are generated corresponding to the individual and combined impact of Random Discrete Dopants in the source/drain regions (RDD), Line Edge Roughness (LER) and Body Thickness Variation (BTV) as the main sources of IPF in next nanometre-scale Ultra-Thin Body Silicon-on Isolator (UTB-SOI) transistor generations on Six-Transistors Static Random Access memory (6T SRAM) cell stability. The L1 cache fault tolerance techniques evaluated are hardware redundancy, parity check,Hamming single error correction double error detection (SECDED), and Hamming triple error detection (TED). It was found that the rate of read faulty cells will rapidly increase in 6T SRAM cache with continued scaling of UTB-SOI device beyond 10 nm gate length. L1 cache conventional fault tolerance techniques, i.e. hardware redundancy, parity check, and SECDED, might be able to hold the implications of IPF cell failures in L1 data cache based 7.5 nm and 5 nm UTB-SOI device, particularly when 6T SRAM is designed with cell ratio of two. However, the effectiveness of these techniques was found to be sensitive to the existence of any faulty word in cache. Hence, their immunity against any transient fault that might occur during system operation will significantly degrade. Experimental results showed that in L1 data cache based on 5 nm UTBSOI device, hybrid hardware redundancy with TED would achieve 68.2 percent of microprocessor chip yield in applications tolerate 10 percent performance loss bound. This indicates that employing these techniques in industry will assist to keep 6T SRAM cache scalability even with the increasing impact of IPF