New RRAM Memory Breaks Through AI Performance Limitations

In the world of artificial intelligence, there is a persistent problem known as the "memory wall." Even the fastest AI models face limitations caused by the time and energy required to transfer data between the processor and memory. One promising solution to this problem is the use of resistive memory (RRAM), which allows computations to be performed directly in memory cells, bypassing traditional bottlenecks.

However, most types of RRAM are unstable and complex to manage. Fortunately, researchers from the University of California, San Diego may have found a solution. At the IEEE International Electron Device Meeting (IEDM) conference, they presented a new type of RRAM capable of executing machine learning algorithms.

"We completely redesigned RRAM, rethinking the switching principle," says Duygu Kuzum, an electrical engineer from the University of California, San Diego, who led the project. Traditional RRAM stores data using low-resistance filaments in a dielectric material. Forming these filaments requires high voltage, which makes it difficult to integrate with CMOS circuits. Moreover, the filament formation process is random and noisy, which negatively impacts data storage, especially in neural networks where weight stability is required.

The new development, called "bulk RRAM," differs in that it switches an entire material layer between high and low resistance. This eliminates the need for high-voltage filament formation and removes the geometry-limiting selector transistor. The San Diego group was not the first to create bulk RRAM devices, but they achieved significant success in reducing their size and creating three-dimensional circuits.

The researchers were able to reduce the RRAM size to the nanometer scale (40 nm) and create eight-layer stacks. Each layer can take on 64 resistance values, which is difficult to achieve with traditional filamentary RRAM. The stack's resistance is in the megaohm range, which, according to Kuzum, is better suited for parallel operations. The increased number of resistance levels and higher resistance allow bulk RRAM to perform more complex operations than traditional RRAM.

The San Diego team tested an array of eight-layer stacks with a volume of 1 kilobyte using a continuous learning algorithm. The chip classified data from wearable sensors, determining whether the user was sitting, walking, climbing stairs, or performing another action. The accuracy was 90%, which is comparable to the performance of a digitally implemented neural network. Kuzum believes that 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 Laboratory, notes that the ability to integrate RRAM into an array is an important step forward. However, he emphasizes that preserving data over extended periods of time may become a problem, especially at the high temperatures at which computers operate. If engineers can prove the reliability of this technology, it could benefit all types of models. The "memory wall" is getting higher as traditional memory cannot keep up with the growing demands of large models. Any solution that allows models to work directly with memory could be a long-awaited breakthrough.