USC Builds Atomic Selector for Denser AI Memory Chips
A new atomic gatekeeper allows memory cells to stack vertically, potentially enabling offline AI on phones.
Key points
- USC researchers developed a selector device 100 times more efficient than current standards.
- The atomic structure allows memory cells to be stacked vertically for higher density.
- The innovation enables running large AI models directly on phones without cloud access.
A team at the University of Southern California has developed a new component that solves a decades-old problem in semiconductor design. By creating a more efficient gatekeeper for memory cells, the innovation allows for significantly denser and more energy-efficient storage. This breakthrough could shift artificial intelligence processing away from remote data centers and onto personal devices.
The device, published in Nature Electronics, outperforms previous best-in-class selectors by a factor of 100. It enables memory arrays to be stacked vertically, similar to building a high-rise, rather than spreading out horizontally. This structural change is critical for fitting the massive memory requirements of modern AI models into compact hardware like smartphones.
Atomic layers replace standard transistors
Current chips rely on silicon transistors as gatekeepers, which are too large to pack densely and cannot be stacked in three dimensions. The new selector uses five atomically thin layers of van der Waals materials. These two-dimensional crystals are so thin that a human hair is roughly 10,000 times thicker, allowing for extreme miniaturization.
The key innovation lies in the shape of the electrical barrier. Previous designs used a flat barrier that electricity had to climb over. The USC team used a graded, triangular shape that becomes both thinner and lower when voltage is applied. This dual effect significantly increases current flow, making the switching process faster and more reliable.
Performance metrics show durability
The new device distinguishes between open and closed states 100 times better than previous technology. It also survived more than one trillion switching cycles without failure. According to Professor J. Joshua Yang, the device has no theoretical endurance limit because only electrons move during operation, preventing physical wear on the material.
Consistency is another major advantage. The selector performs nearly identically from one memory cell to the next due to atomic precision in fabrication. This uniformity is essential for chips containing billions of cells, ensuring that the entire memory array functions predictably under varying temperatures and loads.
Enabling offline AI on phones
Currently, large language models require too much memory and power to run on phones, forcing users to send data to remote servers. This creates latency, costs, and privacy concerns. With dramatically denser memory chips, it may soon be possible to run entire AI models locally on a device without an internet connection.
This shift would allow for private, on-device processing of sensitive data. It also opens the door for computing systems in places where connectivity is impossible, such as deep space missions. The technology offers a practical path to reducing the energy footprint of global AI infrastructure.
