In the rapidly evolving field of artificial intelligence (AI), the demand for efficient data processing has exposed limitations in traditional memory technologies. The paper “3D Stacked IGZO 2T0C DRAM Array with Multibit Capability for Computing in Memory Applications,” published in Science Advances on May 23, 2025, by Qijun Li, Qianlan Hu, Shenwu Zhu, Min Zeng, Wenjie Zhao, and Yanqing Wu, addresses these challenges through innovative use of indium gallium zinc oxide (IGZO) in dynamic random access memory (DRAM). This work demonstrates a 3D stacked, capacitorless DRAM array that promises higher density, longer data retention, and enhanced energy efficiency, particularly for computing-in-memory (CIM) paradigms.
Traditional DRAM, typically structured as one-transistor-one-capacitor (1T1C), suffers from high power consumption due to frequent refresh cycles necessitated by charge leakage. This issue is exacerbated in AI applications involving matrix operations for tasks like image recognition, where data movement between memory and processors creates a “memory wall.” The authors highlight how emerging memories such as spin-transfer torque magnetic RAM (STT-MRAM), resistive RAM (RRAM), phase-change RAM (PCRAM), and ferroelectric RAM (FeRAM) offer alternatives but fall short in cycling endurance, speed, or integration complexity. IGZO-based two-transistor-zero-capacitor (2T0C) DRAM emerges as a superior option due to its ultralow off-state leakage current—on the order of femtoamperes—which enables retention times exceeding 100 seconds without refresh, drastically reducing energy use.
The innovation lies in the monolithic 3D stacking enabled by IGZO’s low thermal budget, compatible with back-end-of-line (BEOL) processes. This allows vertical integration beyond planar scaling limits, increasing bit density. The paper details the fabrication of an 8 by 8 array using advanced techniques like electron beam lithography (EBL) for gate patterning, atomic layer deposition (ALD) for high-κ dielectrics, and reactive magnetron sputtering for the amorphous IGZO channel. The process involves layering read transistors (TRs) and write transistors (TWs) with interconnections via dry etching and metal filling, ensuring electrical isolation with SiO2 insulators. Electrical characterization was performed at room temperature in a vacuum environment using a Keysight B1500A analyzer, confirming optimized performance.
Key results showcase the array’s multibit capability, achieving 3-bit storage per cell with retention over 100 seconds. This is a significant leap, as multibit storage amplifies density and efficiency for AI workloads. The authors map a convolutional neural network (CNN) for handwritten digit recognition onto the array, where an 8 by 8 feature map from the convolutional layer is stored and processed in-memory. Each cell handles int4 weights, enabling vector-matrix multiplication directly within the memory, bypassing data shuttling. Simulations yield an impressive 94.95% accuracy on the MNIST dataset, demonstrating practical CIM viability. Compared to conventional architectures, this approach enhances energy efficiency by minimizing refresh operations and data transfers.
The discussion emphasizes IGZO 2T0C DRAM’s advantages over competitors. While 1T1C DRAM offers high speed and endurance, its short retention (milliseconds) demands constant power. IGZO’s long retention supports nonvolatile-like behavior in volatile memory, ideal for edge AI devices with power constraints. The 3D stacking addresses density bottlenecks, potentially scaling to larger arrays for complex neural networks. However, challenges remain, such as optimizing write/read disturbances and scaling fabrication for commercial viability. The authors suggest future integrations with logic circuits for fully embedded CIM systems.
In conclusion, this research paves a promising pathway for overcoming the memory wall in AI computing. By leveraging IGZO’s unique properties, the 3D stacked 2T0C DRAM array not only extends retention and density but also enables efficient in-memory computations, heralding a shift toward more sustainable and powerful AI hardware. As edge devices proliferate, such innovations could transform applications from autonomous vehicles to wearable tech, reducing global energy consumption in data centers. With further refinements, IGZO-based memories may redefine the DRAM roadmap, blending high performance with low power in an era of exponential data growth.
Also Read:
PDF Solutions Adds Security and Scalability to Manufacturing and Test
TSMC 2025 Update: Riding the AI Wave Amid Global Expansion
Alchip’s 3DIC Test Chip: A Leap Forward for AI and HPC Innovation
