New RRAM Design Breaks Through the Memory Wall Holding Back AI

AI has a persistent problem known as the "memory wall." Even the fastest models run into limits set by the time and energy needed to move data between the processor and memory. One promising fix is resistive memory (RRAM), which computes directly inside memory cells and bypasses those traditional bottlenecks.

But most types of RRAM are unstable and hard to control. Fortunately, researchers at the University of California, San Diego may have found a solution. At the IEEE International Electron Device Meeting (IEDM), they presented a new type of RRAM capable of running machine learning algorithms.

"We completely redesigned RRAM by rethinking how it switches," says Duygu Kuzum, an electrical engineer at the University of California, San Diego, who led the project. Conventional RRAM stores data using low-resistance filaments in a dielectric material. Forming those filaments requires high voltage, which complicates integration with CMOS circuits. The filament formation process is also random and noisy, which hurts data storage, especially in neural networks, where weights must stay stable.

The new design, called "bulk RRAM," instead switches an entire layer of material between high and low resistance. This removes the need for high-voltage filament formation and eliminates the selector transistor that constrains geometry. The San Diego group is not the first to build bulk RRAM devices, but it made significant progress in shrinking them and building three-dimensional circuits.

The researchers shrank the RRAM to the nanometer scale (40 nm) and built eight-layer stacks. Each layer can take 64 resistance values, which is hard to achieve with conventional filamentary RRAM. The stack's resistance is in the megaohm range, which, according to Kuzum, is better suited to parallel operations. The larger number of resistance levels and the higher resistance let bulk RRAM perform more complex operations than conventional RRAM.

The San Diego team tested a 1-kilobyte array of eight-layer stacks using a continual learning algorithm. The chip classified data from wearable sensors, determining whether the user was sitting, walking, climbing stairs or doing something else. Accuracy reached 90%, comparable to a digitally implemented neural network. Kuzum believes bulk RRAM is particularly useful for neural network models on edge devices that need to learn from their surroundings without access to the cloud.

Albert Talin, a materials scientist at Sandia National Laboratories, says the ability to integrate RRAM into an array is an important step forward. He cautions, however, that retaining data over long periods may become a problem, especially at the high temperatures at which computers operate. If engineers can prove the technology is reliable, it could benefit all types of models. The "memory wall" keeps getting higher as conventional memory fails to keep pace with the growing demands of large models. Any solution that lets models work directly with memory could be a long-awaited breakthrough.