SENTRY: Lifetime Secure Technique for Industry Embedded Non-Volatile Random-Access Memory

Emerging non-volatile memories (NVMs), including resistive random-access memory (RRAM) and magnetic random-access memory (MRAM), have been promising solutions for intelligent embedded systems. With the non-volatility and low leakage power consumption, the industry-based embedded NVMs can be used to develop energy-efficient artificial intelligent processors. However, industry-based NVMs typically have low endurance with only 10^4 to 10^6 maximum write times, shortening the lifetime of embedded systems. In addition, non-uniform write distribution to physical addresses onto NVMs may further reduce system lifetime. Furthermore, several intentional attacks, such as malicious writing, may impair embedded NVMs. In this paper, we present a novel wear-leveling technique based on row-column data movement, SENTRY, to extend system lifetime. We explore industry-based embedded NVM chips and analyze practical data distributions of several tasks. We design a coordinate system based moving method with negligible storage overhead for more efficient data movement. The experiment shows that SENTRY achieves a 96.07% life utilization rate with 1.47% data movement overhead. In addition, SENTRY increases the endurance margin of memory by 6048x compared with an unprotected baseline (without SENTRY) under malicious address attacks.