This dissertation introduces a design trinity framework for three-dimensional (3D) phase-change memory (PCM) design in memory-centric computing, aiming for optimal PCM suitable for analog in-memory computing (AIMC). This framework comprehensively cons...
This dissertation introduces a design trinity framework for three-dimensional (3D) phase-change memory (PCM) design in memory-centric computing, aiming for optimal PCM suitable for analog in-memory computing (AIMC). This framework comprehensively considers and optimizes three axes: 3D structures, device-level characteristics, and application-specific requirements. Its validity is demonstrated through two case studies that apply the framework to distinct PCM device designs.
The first case study describes a novel 2-PCM device designed to overcome a fundamental limitation of PCM, which is the inverse relationship between memory window and data retention and is critical for embedded memory application. The device is realized by serially connecting two PCMs having contrasting properties, thereby combining their respective advantages. Experimental validation using monolithically integrated devices based on GST225 and Ga–Sb–Ge demonstrated about 100× on/off ratio at high temperatures and a ten-year data retention temperature of 153℃. This back-end of line (BEOL) compatible device is applicable as a digital storage medium in high-temperature embedded systems as well as in bit-sliced AIMC platforms.
The second case study demonstrates that 3D vertical PCM (3D V-PCM) can serve as a practical foundation for AIMC and develops hardware-friendly neural network models. By optimizing each fabrication step, highly reliable 3D V-PCM devices were successfully fabricated. The 3D memory array structure inherently enables concurrent matrix transposition during MAC (Multiply-Accumulate) operations by simply adjusting input and output directions, significantly reducing memory-movement overhead. Two mixer models—V-PCMixer, which incorporates windowed patch mixing tailored to the physical dimensions of the 3D memory array, and FracMixer, which utilizes discrete fractional Fourier transforms (DFrFT) with real-valued inference—were proposed and evaluated. These results highlight the importance of co-optimizing neural network model, non-idealities and structures of 3D memory to enhance the efficiency and scalability of AIMC system
The feasibility of the proposed 3D phase-change memory design framework is demonstrated through these two case studies. Moving forward, this work anticipates expansion toward integrated design methodologies across all engineering stacks, including the design of memory arrays, peripheral circuits, and systems suitable for analog in-memory computing.