Academic News
The deep learning neural networks have rapidly reached every corner of life; a greater number of embedded memories are needed to store millions of parameters of neural networks and perform in-memory computing. Embedded non-volatile memories (eNVMs) with extremely high density that can be integrated into logic technology have been applied in this field. Compared with volatile memories such as static random access memories (SRAMs) and dynamic random access memory (DRAMs), the eNVMs have better performance in boosting energy efficiency and reducing power consumption. Embedded non-volatile ferroelectric random access memories (FeNVMs) have attracted many researchers’ attention in recent years. Compared with traditional FRAMs made from lead zirconium titanate, which require a thin film with a thickness thicker than 100 nm to generate significant ferroelectric hysteresis, the new type of FeNVMs made from zirconium-doped hafnium oxide (HZO) can generate significant ferroelectric hysteresis with a thin fin thinner than 10 nm. Its simple stacking structure in the form of a sandwich (metal-ferroelectric-metal, MFM) can be easily integrated into the back-end-of-line of advanced complementary metal-oxide-semiconductor (CMOS) logic technology. When ZrHfO is applied to in-memory computing, frequent loading of data and transfer of information are not required because HZO-based ferroelectric hysteresis is non-volatile. Researchers have further discovered that HZO has a multi-lattice crystalline structure. When HZO generates the effect of ferroelectric memory , it can partly flip the electric dipole moment in a polycrystalline structure, making it possible for a memory to store multiple bits of data. The research team led by Assistant Professor E Ray Hsieh in the Dept. of Electrical Engineering at NCU has developed a new type of three-dimensional Fe Fin Field-Effect Transistor (FeFinFET) NVM, which is highly efficient and energy-saving. It can operate under 3V and take only 80 nanoseconds to complete operation. Furthermore, it has a linear-tuning window of conductance that is 30,000 times wider (Fig. 2), its continual endurance cycles exceeded one million times (Fig. 3), and it has the ideal activation function of AI (Fig. 4). These characteristics make FeFinFET NVMs very suitable for highly dense storage with multi-bit-per-cell capacity and AI inference applications based on in-memory computing. This research has been selected by an international flagship conference—2021 Symposia on VLSI Technology—to be covered in its Focus Session.
