Researchers at a San Diego-based lab have made significant advancements in resistive random-access memory (RRAM), potentially revolutionizing the way neural networks operate on edge devices. Our CTO, Duygu Kuzum, has been the creative force behind the ground-breaking technology. This technology has demonstrated exceptional retention attributes and superior operational functionality, particularly for artificial intelligence use cases.
Kuzum notes that their RRAM is data-retentive at room temperature for at least a few years. That performance is on par with traditional flash memory. Albert Talin, another member of the research team, cautions that the retention of RRAM at higher temperatures, where many computing devices function, remains uncertain.
The consequences of this research reach far past retention ability. Kuzum is optimistic that the high-performance bulk RRAM will greatly improve the efficiency and performance of neural network models running on edge devices. “We are doing a lot of characterization and material optimization to design a device specifically engineered for AI applications,” Kuzum stated.
Even with these advances, prohibitive challenges remain that prevent filamentary RRAM from performing parallel matrix operations. These ops are absolutely foundational for today’s state of the art neural networks. Kuzum admitted to this drawback, further emphasizing why their novel strategy to RRAM technology is crucial.
On the R&D team’s part, they made significant advances in miniaturizing RRAM and creating three-dimensional integrated circuits. Perhaps most remarkably, Kuzum and her team were able to miniaturize their RRAM devices down to a mere 40 nanometers wide. This downsampling makes it possible to concat a number of eight-layer stacks into a 1-kilobyte array. Best of all, it works really well without having to use selectors.
The San Diego group’s creativity doesn’t end with going door to door. Their RRAM technology has been demonstrated to achieve 64 unique resistance levels. This achievement is difficult to realize with conventional filamentary RRAM, which typically have high resistance states restricted to kiloohms. The team’s bulk RRAM devices have a higher resistance in megaohm range. This capability is often touted as the most important advantage for performing parallel operations, which are foundational for neural networks.
Kuzum went on to clarify that their bulk RRAM is capable of executing intricate operations with one pulse of the same voltage. This ability is a huge departure from more conventional approaches that typically require more complex setups to do the same thing.
At the recent IEEE International Electron Device Meeting (IEDM), the San Diego researchers showcased their RRAM’s ability to run a learning algorithm effectively. Kuzum even recorded accuracy rates as high as 90 percent on par with digitally-implemented neural networks.
Kuzum remarked, “We actually redesigned RRAM, completely rethinking the way it switches.” This creative method not only saves money, it dramatically improves overall performance. It further solidifies their technology’s position as the superior memory alternative to today’s burgeoning, existing solutions powering AI applications.
Talin concurred with Kuzum’s assessment and added, “I think that any step in terms of integration is very useful.” This appreciation reflects the hard work, both present and promised, across the research community that will continue to take memory technology to even greater heights.