Fast Accessing Non-volatile, High Performance-High Density, Optimized Array for Machine Learning Processor

Abstract

Traditional memory technologies for example; FLASH memory, SRAM (Static Random Access Memory) & DRAM (Dynamic Random Access Memory) are not able to cope with machine learning processor because of high density & low power requirement. FLASH memories have already attained their physical limit and therefore can’t be scaled further due to its bounded tenacity. Non-volatile memories have gained tremendous popularity after being in theoretical study for more than 30 years. Among all various types of Nonvolatile memories, Memristors are the one which are compact, highly dense, fast and nonvolatile. It combines nonvolatile nature of flash memories, speed of SRAMs and high dense nature of DRAMs. Earlier one transistor one memristor based array have been designed but it consisted of CMOS transistor that now has attained fundamental limits and nothing new can be improvised in it because it can’t be scaled further. So, in place of transistor FinFET have been used to form the memory element 1F1M, which fulfill the requirement of machine learning processors. FinFET is the most promising transistor known for its superior controllability of short channel effects and robust threshold voltage (Vth) through a double gate. The main challenge in this time is chip design with scaling of Integrated circuits (IC’s) at Low Power. The thinner Wfin FinFET shows less performance loss than the thicker Wfin FinFET. In VLSI, the Power, Area and Delay are key factors for improving circuit functionality, when any of these can be decreased then the circuit output can be improved. Due to different threshold voltage, memory types in FinFET offer Data Retention Voltage variables compared to CMOS based memory.