This review discusses the development of high-speed emerging memories for AI hardware accelerators, focusing on their potential to replace traditional static random-access memory (SRAM) as global buffer memories. AI accelerators require two types of memory: weight memory for AI model parameters and buffer memory for intermediate data. Emerging memory technologies such as capacitorless gain cell-based embedded dynamic random-access memory (eDRAM), ferroelectric memories (FeFET, FeRAM), and spin-transfer torque (STT) and spin-orbit torque (SOT) magnetic random-access memories (STT-MRAM, SOT-MRAM) are evaluated for their suitability as global buffer memories. These technologies offer high speed, high endurance, and low power consumption, making them suitable for AI hardware at the edge where minimizing stand-by leakage power is critical. However, they face challenges such as charge injection, threshold voltage instability, and high write energy. The review also presents a benchmarking analysis of these memory technologies in TPU-like AI accelerators, showing that emerging memories can outperform SRAM in terms of energy efficiency and area. The review highlights the potential of BEOL stacked 2T gain cells and FeFETs for high-density global buffer solutions. The study concludes that while SRAM remains competitive for cloud-based AI hardware, emerging memories offer significant advantages for edge AI applications. The review also discusses the industrial development and technological challenges in buffer memory applications, and suggests that further research is needed to improve the performance and reliability of emerging memories for AI hardware accelerators.This review discusses the development of high-speed emerging memories for AI hardware accelerators, focusing on their potential to replace traditional static random-access memory (SRAM) as global buffer memories. AI accelerators require two types of memory: weight memory for AI model parameters and buffer memory for intermediate data. Emerging memory technologies such as capacitorless gain cell-based embedded dynamic random-access memory (eDRAM), ferroelectric memories (FeFET, FeRAM), and spin-transfer torque (STT) and spin-orbit torque (SOT) magnetic random-access memories (STT-MRAM, SOT-MRAM) are evaluated for their suitability as global buffer memories. These technologies offer high speed, high endurance, and low power consumption, making them suitable for AI hardware at the edge where minimizing stand-by leakage power is critical. However, they face challenges such as charge injection, threshold voltage instability, and high write energy. The review also presents a benchmarking analysis of these memory technologies in TPU-like AI accelerators, showing that emerging memories can outperform SRAM in terms of energy efficiency and area. The review highlights the potential of BEOL stacked 2T gain cells and FeFETs for high-density global buffer solutions. The study concludes that while SRAM remains competitive for cloud-based AI hardware, emerging memories offer significant advantages for edge AI applications. The review also discusses the industrial development and technological challenges in buffer memory applications, and suggests that further research is needed to improve the performance and reliability of emerging memories for AI hardware accelerators.