New IBM Research Proposes Phase-Change Materials (PCMs) To Break The von Neumann Bottleneck For AI And Deep Learning Applications

865

We live in a world where significant technological developments in processing technology have dramatically transformed our way of life, with rapid improvements in computing capacity.

The world’s information continues to grow. In 2018, the total amount of data stored in the world was 33 zettabytes (33×1021 bytes). To put it another way, one zettabyte of data would require 33 billion one-terabyte.

As difficult as it is to wrap one’s head around that amount of data, it is expected to swell to 175 zettabytes by 2025. To this date, extracting and storing this increasingly massive amount of data represents a tremendous challenge in terms of efficiency, accuracy, and sustainable energy cost.

To meet this issue, IBM Research is researching novel materials to provide the foundation for faster, more energy-efficient systems. Phase-change materials (PCMs) are one of the most advanced, storing and deleting information based on changes in their atomic structure from crystalline to disordered or amorphous states.

By using the rapid and reversible amorphous-to-crystal transition, phase-change materials (PCMs) are finding wide applications in developing technologies such as nonvolatile phase-change random-access memory and in-memory computer systems.

Von Neumann architectures are used in today’s electronic gadgets to store data through sequential data exchange between physically separated CPU and storage or memory units. However, because instructions can only be executed one at a time and sequentially, these designs have restricted throughput.

The von Neumann bottleneck mainly constrains artificial intelligence and deep learning applications. Data must be transferred to and fro between the processor and memory and to long-term storage and peripheral devices because the von Neumann design isolates the processor from memory.

Information is quickly erased when PCM crystals are heated to their softer, more amorphous state. Unfortunately, cooling PCMs to re-crystallize them for data storage takes at least 1,000 times longer. This is a significant impediment in throughput and a roadblock to developing next-generation electronic gadgets.

Another disadvantage of PCMs is their inefficient energy utilization. It takes an enormous amount of energy to heat them, and excellent thermal insulation is required to prevent heat loss.

The research focuses on deciphering and characterizing what happens at the atomic level during PCM changes. The goal is to create more efficient PCM-based storage devices, ultimately improving the speed and efficiency of computing architectures.

Crystallization is thought to be a random process that begins with forming a crystal-like nucleus (commonly known as a critical nucleus). Researchers demonstrated that the critical nucleus originates from a smaller nucleus using a method developed a few years ago. They discovered that this smaller nucleus is generated by atoms with differing mobility colliding in space, allowing them to adapt in space in a very stable configuration, stable enough to resist melting and develop further to form the crucial nucleus.

The goal is to determine which atoms move faster in a PCM and how they combine to form a stable nucleus. Knowing this gives us a reason to either change the chemical composition of the PCM to promote crystallization or develop new crystallization-accelerating methods.

Breaking through the von Neumann bottleneck isn’t the only use for this technology. IBM Research’s development of novel materials and devices for electronic and photonic neuromorphic computing systems could be boosted by faster and more efficient PCMs. Neuromorphic computing, which stimulates the brain’s interconnected network of synapses, could also benefit from it.

Paper: https://www.sciencedirect.com/science/article/abs/pii/S1369800121004443?via%3Dihub

Source: https://research.ibm.com/blog/pcm-breaks-bottleneck