Challenges in the implementation of artificial intelligence hardware based on phase change devices

Cross-point arrays built with phase change tunable resistors were suggested for a computationally efficient implementation of artificial neural networks (ANN) [1,2]. By storing the ANN weights as the conductance in the array elements and using Ohm's law and Kirchhoff’s current law the multiply-accumulate operation (MAC) can be realized using an analog computation. To maximize computational accuracy while maintaining low energy consumption, the phase change cells in the array need to have a low reset current, low resistance drift, and low read/write noise. In this paper we will discuss recent advances in textured heterostructure superlattice PCM devices and textured homostructure PCM devices which shown to exhibit low resent current and low resistance drift, and therefore may be used in ANN hardware [3-7]. We will also review limitation to the cross-point array size imposed by noise, and methods to improve the computation accuracy in the presence of noise.

References:

1. H. Tsai et al., “Recent progress in analog memory-based accelerators for deep learning”, J. Phys. D: Appl. Phys. 51 283001, (2018).
2. A. Sebastian et al., “Tutorial: Brain-inspired computing using phase-change memory devices”, J. Applied Physics 124, 111101 (2018).
3. R. E. Simpson et al, “Interfacial phase-change memory”, Nature Nanotechnol., 244, p.501, (2011).
4. K. Ding, et al., “Phase-change heterostructure enables ultralow noise and drift for memory operation”, Science, 366, p. 210, (2019).
5. A. I. Khan et al., “Electro-Thermal Confinement Enables Improved Superlattice Phase Change Memory” Electron Device Lett., 43, p. 204, (2022).
6. X. Wu et al., “Novel nanocomposite-superlattices for low energy and high stability nanoscale phase change memory”, Nature Communications, 15:13, (2024).
7. G. M. Cohen et al., “Low RESET Current Mushroom-Cell Phase-Change Memory Using Fiber-Textured Homostructure GeSbTe on Highly Oriented Seed Layer”, Phys. Status Solidi RRL, 2300426, (2024).