From IBM Research, Breaking von Neumann Bottleneck Using PCM

By Fausto Martelli, research staff member, IBM Research laboratory, UK

The world’s information keeps expanding. In 2018, global data storage reached 33ZB (33×10²¹ bytes). Put another way, it would take 33 billion 1TB HDDs to store one zettabyte of data.

As difficult as it may be to wrap your head around that amount of data, it’s expected to swell to 175ZB by 2025. Even today, storing and extracting this increasingly massive amount of data represents a remarkable challenge in terms of accuracy, efficiency and sustainable energy cost.

IBM Research is responding to this challenge by studying new materials that could be the basis for faster, more energy-efficient architectures. One of the most mature is phase-change materials (PCMs) that store and delete information based on changes in their atomic structure from crystalline to a disordered, or amorphous, state.

My recently published research ⁽¹⁾ produced in collaboration with the Chinese Academy of Sciences, reports, for the first time, exactly how PCMs crystallize at the molecular level.

Material change
Today’s electronic devices feature von Neumann architectures, which store information via a sequential data exchange between physically separated CPU and memory or storage units. But these architectures have limited throughput, because instructions can be carried out only one at a time and sequentially.

This von Neumann bottleneck is especially limiting for artificial intelligence and deep learning applications.

“PCMs are important because they can switch very rapidly and reversibly
for a virtually countless number of cycles (up to one trillion, or 1012, cycles)“.

Heating PCM crystals to transform them to their softer, more amorphous form deletes information quickly. Unfortunately, cooling PCMs down to crystallize them again – in order to store information – is at least 1,000x slower. That’s a huge bottleneck in throughput, and a barrier to the development of next-generation electronic devices.

Inefficient energy use is another downside of PCMs. Heating them consumes a lot of energy and requires efficient thermal insulation to prevent heat loss.

My work is centered on understanding and describing what’s happening on the atomic level during these PCM transformations. The goal is to develop more efficient PCM-based storage devices and, ultimately, boosting the speed and efficiency of computing architectures.

Simulation technique reveals molecular mystery
Common theories predict the crystallization occurs as a random event initiated by the formation of a crystal-like nucleus (commonly known as critical nucleus). Using a method I introduced a few years ago, ⁽²⁾ I showed that the critical nucleus forms from a smaller nucleus. I observed that this smaller nucleus is formed from the encounter of atoms with different mobility, which allows them to adjust in space in a very stable configuration, stable enough to resist melting and to further grow to generate the critical nucleus.

The goal is to understand which atoms in a PCM move faster and how atoms interact to build a stable nucleus. Knowing this provides us with a rational to either modify the chemical composition of the PCM favoring crystallization or to develop new technologies able to speed the crystallization process.

Breaking through the von Neumann bottleneck is far from the only application for this work. It could also spur advances in neuromorphic computing, which models the brain’s interconnected network of synapses. Faster and more efficient PCMs could fire up IBM Research’s development of new materials and devices for electronic and photonic neuromorphic computing systems.

Eventually, these novel materials could be the basis for more complex AI that ‘thinks’ like the human brain.

⁽¹⁾ Song, W., Martelli, F., Song, Z. Observing the spontaneous formation of a sub-critical nucleus in a phase-change amorphous material from ab initio molecular dynamics. Materials Science in Semiconductor Processing. Volume 136. 2021. 106102. ISSN 1369-8001.
⁽²⁾ Martelli, F., Ko, H., Oğuz, E., Car, R. Local-order metric for condensed-phase environments. Phys. Rev. B 97, 064105. Published 12 February 2018.

Article: Observing the spontaneous formation of a sub-critical nucleus in a phase-change amorphous material from ab initio molecular dynamics

Materials Science in Semiconductor Processing has published an article written by Wen-Xiong Song, State Key Laboratory of Functional Materials for Informatics, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China, Fausto Martelli, IBM Research, Hartree Centre, Daresbury, WA4 4AD, United Kingdom, and CNR, Institute of Complex Systems, Department of Physics, Sapienza University of Rome, P.le Aldo Moro 2, 00185 Rome, Italy, and Zhitang Song^, State Key Laboratory of Functional Materials for Informatics, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China^.

Abstract: “Phase-change materials (PCMs) are finding wide applications in emerging technologies such as nonvolatile phase-change random-access memory and in-memory computing devices by utilising the ultrafast and reversible amorphous-to-crystal transition. The crystallisation of the amorphous state is much slower compared to the crystal melting, hence representing a bottleneck in the further development of new technologies. Here, we disclose the detailed crystallisation pathway of amorphous Ge₂Sb₂Te₅ (GST) via ab initio molecular dynamics simulations. By probing the local order with a highly sensitive metric, we detect the formation of a sub-critical nucleus formed by the spontaneous aggregation of highly ordered octahedral-like atoms. Specifically, we observe that Sb atoms recover octahedral-like geometry quicker than Ge atoms, implying that the different mobility of different species plays a central role in the overall process. With respect to other less locally-ordered (hence transient) domains, this stable precursor is characterised by lower energy, is resilient to melting, and acts as a skeleton over which the critical nucleus further develops. The detailed understanding of the kinetics of homogeneous nucleation in GST opens the doorway for the development of more efficient PCM-based storage devices by a rational composition design and, ultimately, for the boosting of the speed and efficiency of computing architectures.“