Amnesia, or memory loss, is a cliché plot device frequently used in movies as seen in Memento, Total Recall, or The Bourne series to name a few. It is a convenient tool to stimulate the audience's curiosity at the beginning, twist the storyline during plot development, and present thrilling experiences at the ending. And while we rarely encounter it in our everyday lives, we do often find ourselves forgetful. Our real lives seem to run somewhere between clear perception and complete oblivion. Recognition, which relies entirely on flimsy memory, is often vulnerable to mistakes or memory distortion. Details may fade, losing their vivid original color, and then be repainted to adhere to one’s confirmation bias. Time plays a crucial role in accelerating this process.

In flash memory, like in the human brain, data retention is not endless; portions of stored data are gradually lost with time and retrieved data might not be exactly the same as it was when originally written. The longer the storage period is, the larger the flash memory data loss becomes. In order to compensate for this limit in reliability, protective measures are devised and implemented. Error correction code (ECC) is an example of a protective measure that is designed to detect and correct the corrupted data by comparing the integrity of the retrieved data to its original form. Even with such efforts, if the overlap of two supposedly distinctive sets of data outgrows the limit that ECC can handle, the retrieved data will be faulty. In flash memory devices where the ON/OFF states are set by switching the threshold voltage (Vth) of a memory cell, the overlap of two data sets means the Vth of one state is somehow shifted to the Vth level of the other, resulting in unintended ‘1 to 0’ or ‘0 to 1’ data conversion. The overlap of memory states, or data retention failure, is analogous to memory loss of the human brain.

It is interesting to highlight that simple digital single-level cell (SLC, 1 bit/cell, 2 states ON/OFF) has been evolved to triple-level cell (TLC, 3 bits/cell, 8 states) or quad-level cell (QLC, 4 bits/cell, 16 states) where multiple digits are stored in the same single physical cell. In other words, two distinct Vth distributions of SLC are now split into 8 or 16 sets of rather analog-like TLC or QLC Vth distribution. Flash memory is not alone when it comes to the directional trend of memory evolution towards analog-like multi-states distribution. Various non-volatile memories (NVMs) including resistive memory (RRAM), phase-change memory (PCM), and magnetic memory (MRAM) have been considered for analog computing-in-memory (Analog CIM) architecture exploiting the multi-states characteristic of the NVMs. Recent interest in CIM architecture for deep neural networks (DNNs) inspires such application, and taking advantage of the analog nature of the once-digital cell is not uncommon. Undoubtedly, reliability of these multi-level cells should be more carefully controlled.

As defying the scaling limits of DRAM and NAND technologies is one of our core visions, Revolutionary Technology Center (RTC) is naturally destined to explore new materials and architectures. We demonstrated an MLC-capable ferroelectric 3D-NAND device in IMW 2022 (https://doi.org/10.1109/IMW52921.2022.9779278), where we introduced new ferroelectric material as a novel storage element and integrated it into the current 3D-NAND architecture. Whenever new material is adopted into the existing system, it almost always accompanies new reliability concerns like endurance or data retention issues with different failure mechanisms. Thus, introducing foreign materials like ferroelectric film requires engineers to be overly cautious about unpredicted device behaviors.

Now, with overwhelming amounts of data and tremendous data processing requirements, researchers are asked to think innovatively with an open mindset to work across any technology boundaries, such as the digital-to-analog, logic-to-memory, device-to-system boundaries and many more. Who knows? Someone might find a way even to exploit troubling data retention failure in the NVM device for specific neuromorphic computing applications, as forgetting trivial details in the human brain could sometimes help extract concepts for efficient decision-making.