Multiscale Modeling Platform Development: a Neuromorphic Memory Case

F. Nardi, A. Padovani, M. Pesic, L. Larcher
Applied Materials,
United States

Keywords: modeling and simulation, non-volatile memories, dielectrics, charge and ion transport


The development of a new technology requires huge efforts in terms of materials screening, process tuning and device design. This silicon learning process is often based on a brute-force approach that requires to fabricate and electrically characterize multiple test wafers. This comes along with significant costs both in economic and time-to-market terms. Fig. 1 shows the schematic representation of the Material to System (M2S) workflow that we are currently developing at Applied Materials (AMAT) where both simulation and physical paths are run in parallel accelerating the technology development from material to system. The interdependency of these two paths is achieved through the Material Database, a collection of a wide variety of attributes enabling the use of silicon-proven simulations for hardware-level benchmarks and predictions. This infrastructure is currently leveraged for internal R&D and planned to be ported to external customers where thin films deposited by AMAT’s High Volume Manufacturing (HVM) tools can be linked to electronic device performances (and vice-versa) by AMAT’s multiscale modeling software Ginestra™ [1]. Fig. 2 is a schematic illustration of the software where the electrical device response is linked to the material properties through a multiscale platform. Key atomic properties of the materials are calculated using ab-initio simulations [2] and a link to compact model and system level benchmarks is established. AMAT’s software includes all the physical mechanisms relevant for the operations and reliability of logic (FinFET, GAA) and memory (3D NAND, RRAM, FeRAM, PCM, etc.). Thanks to the software compatibility with High Performance Computing (HPC) infrastructure we can perform time-consuming statistical and device optimization simulations in few hours/days thus providing a fast, virtual silicon learning path toward technology design and optimization. As an example of the versatility of the simulation platform, we used Ginestra™ to explore the neuromorphic performances (potentiation/depression pulse sequences for linear weight update) of different memory technologies, i.e. RRAM and 3D-NAND. As shown in Fig. 3, a comparison between 1-layer and 2-layer RRAM synapses showed better performances of the latter stack [3]. This is due to the improved field redistribution and control of O ion diffusion, allowing more gradually modulating the conductance of the barrier/CF. Another example is reported in Fig. 4 for the optimization of a 3D-NAND synapse [4]. A wide parameter space comprised of 288 virtual stack variations and 100 pulsed operating conditions has been covered in few days of calculations using 1000 nodes of HPC (70x time reduction and 120x cost reduction as reported in Table I). Unique Ginestra™ functionalities (in addition to the memory modules) were used in these studies. Fig 5a-b respectively show the Optimization Tool (allowing the screening of optimum materials & geometry targeting key device performances) and the Defect Spectroscopy Tool (allowing to automatically extract material/device and defects properties from multiple electrical curves).