Neuromorphic Computing Platform In Perovskite Nickelates (UCSD, Rutgers)

Why It Matters

Merging processing and memory in one material cuts latency and power for edge AI, accelerating the path toward scalable neuromorphic chips.

Key Takeaways

• Protonic nickelate devices operate at nanosecond speeds
• Energy consumption ~0.2 nJ per input event
• Combines transient dynamics with multilevel resistance memory
• Achieves real‑time spoken digit and seizure detection
• CMOS‑compatible platform for scalable neuromorphic hardware

Pulse Analysis

Neuromorphic engineering seeks to emulate the brain’s ability to process information through tightly coupled temporal and spatial dynamics. Traditional hardware separates logic and storage, leading to latency and energy penalties that hinder real‑time artificial intelligence at the edge. Recent advances in material‑based computing have explored memristors, phase‑change alloys, and spintronic devices, yet few have demonstrated simultaneous short‑term dynamics and long‑term weight programming within a single substrate. The UC San Diego–Rutgers collaboration addresses this gap by leveraging the unique proton‑conducting properties of perovskite nickelates, offering a unified platform where computation and memory coexist at the device level.

The core of the platform consists of hydrogen‑doped NdNiO₃ junctions fabricated on a common wafer. Symmetric junctions act as volatile nodes, delivering ultrafast, proton‑mediated resistance transients that encode temporal spikes in nanoseconds. Asymmetric junctions provide stable, multilevel resistance states that serve as reconfigurable synaptic weights. Proton redistribution across neighboring devices creates emergent spatial coupling, enabling network‑level feature transformation without external interconnects. Measured energy per input is roughly 0.2 nJ, comparable to state‑of‑the‑art CMOS neurons but with orders‑of‑magnitude lower latency, and the process remains compatible with existing silicon manufacturing lines.

Demonstrations of spoken‑digit classification and early seizure detection illustrate the platform’s capability to handle both pattern‑recognition and time‑critical medical diagnostics. By collapsing the von Neumann bottleneck, these nickelate networks promise substantial power savings for edge devices such as autonomous sensors, wearables, and low‑power robotics. Commercial adoption will depend on scaling the wafer‑level integration and establishing design‑automation flows, but the CMOS‑compatible chemistry lowers the barrier to entry for semiconductor foundries. As AI workloads continue to migrate toward on‑device inference, protonic nickelate neuromorphic chips could become a cornerstone of next‑generation energy‑efficient intelligence.