Due to its high density and close-to-SRAM read latency, spin transfer torque RAM (STT-RAM) is considered one of the most-promising emerging memory technologies for designing large last level caches (LLCs). However, in deep sub-micron region, STT-RAM shows read-disturbance error (RDE) whereby a read operation may modify the stored data value and this presents a severe threat to performance and reliability of STT-RAM caches. In this paper, we present a technique, named SHIELD, to mitigate RDE in STT-RAM LLCs. SHIELD uses data compression to reduce number of read operations from STT-RAM blocks to avoid RDE and also to reduce the number of bits written to cache during both write and restore operations. Experimental results have shown that SHIELD provides significant improvement in performance and energy efficiency. SHIELD consumes smaller energy than two previous RDE-mitigation techniques, namely high-current restore required read (HCRR, also called restore-after-read) and low-current long latency read (LCLL) and even an ideal RDE-free STT-RAM cache.
Index TermsNon-volatile memory (NVM), cache memory, STT-RAM, read disturbance error, data compression, reliability.
I. INTRODUCTIONRecent trends of increasing core-counts and LLC capacity have motivated researchers to explore low-leakage alternatives of SRAM for designing large LLCs. Due to their high density and near-zero leakage power consumption, non-volatile memories such as STT-RAM and ReRAM (resistive RAM) have received significant attention in recent years[1], [2]. Of these, STT-RAM is considered especially suitable for designing LLCs due to its close-to-SRAM read latency and high write endurance [3].Since the write latency/energy of STT-RAM are higher than those of SRAM, previous work has mainly focused on addressing this overhead to enable use of STT-RAM for designing on-chip caches [4]. However, a more severe issue of 'read disturbance error' (RDE) in STT-RAM caches has not been adequately addressed. Specifically, with feature size scaling, the write current reduces, however, read current does not reduce as much [5]. In fact, for sub-32nm feature size, the magnitude of read current becomes so close to the write current that a read operation is likely to modify the stored data and this is referred to as RDE (refer Section II for more details). Since reads happen on critical access path, RDE is likely to severely affect performance and reliability. In fact, it is expected that with ongoing process scaling, readability and not writability will become the most crucial bottleneck for STT-RAM [6], [7], [8], [9].Contributions: In this paper, we present a technique, named 'SHIELD', which shields STT-RAM caches against both RDE and high write latency/energy overhead. SHIELD works by using data compression to reduce the number of S. Mittal is with IIT Hyderabad, India.