San Diego scientists have been making big waves with resistive random-access memory (RRAM). This advance has the potential to revolutionize memory technology, improving the energy efficiency of neural networks by more than 100x. Through their work, Dr. Kuzum and his team have made great miniaturization progress with the RRAM devices. They’ve built three-dimensional integrated circuits, creating new opportunities for use in edge computing.
It was a terrific breakthrough by the San Diego team. They were then able to integrate several eight-layer stacks of RRAM cells into a single small 1-kilobyte array—again, without any selectors. Each of these eight-layer stacks can manage 64 individual resistance values. At first glance, they can all be controlled with a single pulse of the same voltage. With this unique and creative approach, data processing and storage has been streamlined even further.
The San Diego group got phenomenal resistance levels, with a megaohm plus range and block response. This increase propulsion device capability thus makes a marked improvement on how devices perform during parallel runs. Dr. Kuzum tells us that this increased resistance is particularly advantageous for executing complicated, parallelized matrix operations in lockstep. These operations are second nature to the building of today’s neural networks. She pointed out traditional filamentary RRAM’s drawbacks in this space. This replication is very expensive, and as such, makes it less attractive for the performance requirements of advanced computing.
The researchers have demonstrated that their RRAM devices can reliably retain data at room temperature. This remarkable discovery bodes very well for future data storage in the decades to come. As Dr. Kuzum admitted, the retention abilities at such high temperatures—in which most computers would be used—is still unknown and needs more scrutiny.
The San Diego team’s accomplishments reach into miniaturization, with their RRAM devices only 40 nanometers wide. This decrease in size has been imperative in testing the devices’ performance. As a consequence, they were able to realize a remarkable accuracy with 90 percent—on par with digitally realized neural networks.
The San Diego group was not the first to build bulk RRAM devices. Yet, despite these myriad challenges, their work is a huge leap forward in the field. Compared to conventional RRAM, bulk RRAM can realize more sophisticated operations. This feature is particularly useful for neural network models deployed on edge devices. In order to achieve these breakthroughs, these models should learn from their environment. By doing so, they can get real-time response times and operational efficiencies even without cloud access.
Yet traditional memory technologies like DRAM can’t support the pace at which memory must grow with these large models. Yet, forward-thinking solutions from San Diego may hold the key in closing this gap. Edge devices are now smarter, capable of storing and processing more complex tasks. This progress only increases capability in applications such as artificial intelligence and real-time data processing.