Zero-Counting and Adaptive-Latency Cache Using a Voltage-Guardband Breakthrough for Energy-Efficient Operations

Abstract

In modern embedded processor systems, energy efficiency is a critical issue. Unfortunately, to avoid cache memory (SRAM) faults from dynamic variations, caches generally operate at an elevated voltage to build a safety guardband that decreases energy efficiency. To address this issue, tolerating SRAM faults to eliminate the safety of a guardband without frequency scaling may be a viable solution. This study investigates the characteristics of low-voltage 8 T SRAM faults and demonstrates that most SRAM faults are typically caused by insufficient access times with variation effects and significantly reduced voltages. Thus, we propose an access-time fault-tolerant cache design based on a type of 8 T SRAM known as zero-counting and adaptive-latency cache (ZCAL cache), which can tolerate numerous access-time faults. ZCAL caches detect access-time faults dynamically using a lightweight error detection code ("0" counting) because access-time faults occur only when reading "0" bits on the 8 T SRAM; the cache then adapts its access time to tolerate the access-time faults with new cache management processes. With the ZCAL cache, the experimental results from the MiBench benchmarks indicate that the energy efficiency is improved by 17% on average and that the energy consumption is reduced by 22% from 0.76 to 0.63 V.