Improves AI with optimized phase change memory
(Nanowerk Highlights) Phase change memory (PCM) is a type of non-volatile memory technology that stores data at the nanoscale by changing the phase of a particular material between crystalline and amorphous states. In the crystalline state, the material exhibits low electrical resistance, while in the amorphous state, it has high resistance. By applying distinct heat pulses and rapid cooling, phases can be switched, allowing data to be written and read as binary values (0s and 1s) or continuous analog values based on the resistance of the material.
Phase change memory is an emerging technology with great potential for advancing computation in analog memory, particularly in deep neural networks and neuromorphic computing. Various factors, such as resistance values, memory windows and resistance drift, affect PCM performance in these applications. So far, it has been a challenge for researchers to compare PCM devices for in-memory computing based solely on their various device characteristics, which often have compromises and correlations.
Another challenge is that computations in analog memory can greatly increase speed and reduce power consumption for AI computations, but may suffer from reduced accuracy due to imperfections in analog memory devices.
New research, published in Advanced Electronic Materials (“Optimization of Projected Phase Change Memory for Computational Inference In Analog Memory”), addresses this issue by 1) performing extensive benchmarking of PCM devices in large neural networks, offering valuable guidance for optimizing these devices in the future, and 2) improving and optimizing analog memory devices fabricated with phase-change materials, ultimately improving AI computational accuracy.
Ning Li, then at IBM Research in Yorktown Heights and Albany (now Associate Professor at Lehigh University), the study’s first author, and her IBM colleagues explain: “First, we found that many device characteristics can be systematically tuned systematically using the liner layers introduced in our previous work.Second, we found a way to optimize the characteristics of these devices from a system point of view using extensive system-level simulations.” These two advances together allowed the team to identify the best tools.
In this work, the team created a model to represent the drift and noise behavior of PCM devices. They used this model to assess the performance of these devices in neural network inference applications. They evaluated the performance of large neural networks with tens of millions of weights (i.e., parameters in a neural network that determine the strength of connections between neurons; in the case of PCM-based analog in-memory computations, weights are stored as resistance values in the PCM device) using PCM devices both with and without liners. projection (an additional layer is inserted into the PCM device structure, which is made of non-phase changing materials), testing various deep neural networks (DNNs) and data sets at multiple time steps.
“We report two core findings in our paper,” Li explained to Nanowerk. “Firstly, we found that many device characteristics can be tuned systematically by adding additional liner layers in the device structure. Second, we found a way to optimize these device characteristics from a system point of view using extensive system-level simulations. These two breakthroughs together allowed us to identify the best device.”
This study found that devices with projection liners performed well in various types of DNNs, including recurrent neural networks (RNNs), convolutional neural networks (CNNs), and transformer-based networks. The researchers also examined the impact of different device characteristics on network accuracy and identified various target device specifications for liner PCM that could lead to further improvements.
Unlike previous reports on PCM tools for AI computing, this work ties the device results to the end result of computing chips with large and useful deep neural networks. Dr. Li explained that PCM devices for in-memory computing are difficult to compare for AI applications using only device characteristics. This study provides a solution to this problem by offering an extensive comparison of PCM devices across different networks under various weight mapping conditions and a guide for optimizing PCM devices.
By being able to show that device characteristics can be tuned continuously, and that these characteristics are correlated with each other, systematic device optimization becomes possible.
Using their optimization strategy, the researchers demonstrated that they could achieve much better accuracy for both short and long term programming. They significantly reduce the effects of PCM drift and noise on deep neural networks, increasing initial accuracy and long-term accuracy.
“Potential applications of our work include increasing speed, reducing power, and reducing costs in language processing, image recognition, and even broader AI applications, such as ChatGPT,” said Li.
As a result of this work, the researchers imagine computing large neural networks will become faster, greener, and cheaper. The next stage in their investigation includes further optimizing PCM devices and implementing them in computer chips.
“The future direction for this research field is to enable tangible products that customers find useful,” Li concluded. “While analog systems use imperfect analog devices, they offer significant advantages in terms of speed, power and cost. The challenge lies in identifying the appropriate applications and enabling them.”
Michael is the author of three books by the Royal Society of Chemistry:
Nano-Society: Pushing the Boundaries of Technology,
Nanotechnology: A Small FutureAnd
Nanoengineering: Skills and Tools for Making Technology Invisible
Become a Spotlight guest writer! Join our large and growing group of guest contributors. Have you recently published a scientific paper or have other interesting developments to share with the nanotechnology community? Here’s how to publish on nanowerk.com.