A team of researchers from the University of California, San Diego (UCSD) have made a surprising discovery that could lead to a RRAM technology breakthrough. They shared their results to the cutting-edge IEEE International Electron Device Meeting (IEDM). This innovation involves running a learning algorithm on a new type of RRAM, which is poised to address the challenges faced by neural networks in processing data efficiently.
Duygu Kuzum, Principal Investigator
In her team’s innovative configuration, Kuzum has assembled an unusual cabal. By bridging several eight-layer stacks, they built it into a 1-kilobyte array that operates without requiring selectors. This development greatly increases the capability of parallelizable matrix operation like those needed for today’s neural networks. Kuzum points out that in conventional filamentary RRAM, limitations such as poor reliability, limited scalability, and transition variability hinder effective implementation in such operations.
Kuzum explained that their innovative eight-layer stack design allows each cell to take on any of 64 resistance values. Notably, this conversion occurs with only one identical voltage pulse. Besides making the operational workflow much more straightforward, this capability significantly boosts the computation efficiency of workload neural networks.
“We actually redesigned RRAM, completely rethinking the way it switches,” – Duygu Kuzum
Kuzum and her colleagues had already taken steps to minimize their RRAM device down to just 40 nanometers across. This nanoscale strategy further enhances the device’s resistance to the megaohm range. This high output resistance greatly facilitates the effective operation in parallel. In comparison, conventional filament-based RRAM cells usually show resistance capped at kiloohms.
As far as Kuzum is concerned, bulk RRAM has distinct advantages for neural network models run on edge devices. These devices need to figure out their surroundings all on their own. This additional requirement renders a more efficient memory technology than today’s SRAM essential. Lately, San Diego researchers have announced stunning test results. With an accuracy rate of 90 percent, their RRAM comes within spitting distance of the performance metrics of digitally implemented neural networks.
“We are doing a lot of characterization and material optimization to design a device specifically engineered for AI applications,” – Duygu Kuzum
The progress achieved by the UCSD research team foreshadows a major development in the complexity of memory technologies. The San Diego group’s RRAM also removes filamentary paths and fine-tunes the high and low resistance states. Consequently, it is able to conduct more sophisticated maneuvers than its conventional brethren. This makes their technology a particularly attractive solution for the requirements presented by the latest generation of artificial intelligence applications.
The collaborative research team adds a unique twist that increases the power of RRAM exponentially. This breakthrough unlocks new creative opportunities for embedding RRAM into AI workflows. This technology is much more than a performance booster. Its promise is so great, it could fundamentally change the way all edge devices sense and learn about the world and respond to it.
“I think that any step in terms of integration is very useful.”
Even more impressive breakthroughs could be just around the corner as researchers keep pushing the limits of what’s possible with bulk RRAM. The potential to tube resistance levels through a wider array could unlock even more system-level efficiencies and functionality.
As researchers continue to explore the potential of bulk RRAM, further advancements may soon emerge. The ability to tune resistance levels across a broader spectrum could lead to even greater system-level efficiencies and functionality.
“We can actually tune it to anywhere we want, but we think that from an integration and system-level simulations perspective, megaohm is the desirable range,” – Duygu Kuzum
