Multi‐Bit Floating‐Gate Memory with an Ultrawide Programmable Window

Why It Matters

The breakthrough delivers unprecedented density and multilevel capability in a fast, durable memory, accelerating the adoption of neuromorphic and edge‑AI systems that require integrated optoelectronic control.

Key Takeaways

• 199 V memory window, 90.5% ratio.
• Storage density exceeds 10^13 cm⁻².
• 50 ns switching, 10⁶ endurance cycles.
• 6‑bit electrical, 7‑bit optical multilevel storage.
• 95% CIFAR‑10 accuracy via in‑memory computing.

Pulse Analysis

The explosion of AI workloads and high‑resolution imaging has stretched conventional memory architectures to their limits. Traditional floating‑gate transistors offer reliable charge storage but suffer from modest storage densities and limited multilevel capability, making them ill‑suited for next‑generation neuromorphic or optoelectronic systems. Researchers have turned to van der Waals heterostructures—layered crystals bonded by weak inter‑layer forces—to engineer novel charge‑trapping interfaces. By stacking molybdenum ditelluride (MoTe₂), hexagonal boron nitride (hBN), and multilayer graphene, the new FGFET creates a pristine, atomically thin floating‑gate channel that overcomes many of the legacy constraints.

The resulting device delivers a p‑type memory window of roughly 199 V and a window ratio of 90.5 %, translating to an unprecedented charge‑storage density above 10¹³ cm⁻². Switching occurs in just 50 ns, and the structure endures more than one million program/erase cycles while retaining data for a projected decade. Electrical pulsing enables over six bits of multilevel storage, and the intrinsic negative photoconductance permits laser‑driven programming that pushes the level count beyond seven bits. Moreover, the transistor mimics synaptic plasticity, positioning it as a versatile building block for analog‑digital hybrid computing.

Beyond raw metrics, the heterojunction’s ability to perform in‑memory computation was demonstrated on the CIFAR‑10 benchmark, achieving classification accuracy exceeding 95 %—on par with fully digital networks. This convergence of ultra‑dense charge storage, ultrafast operation, and optoelectronic programmability opens a pathway for compact neuromorphic processors and edge AI accelerators that can learn and retain information locally. Industry players seeking to shrink form factors while boosting energy efficiency may adopt this technology for high‑performance storage class memory, sensor‑fusion modules, and next‑generation smart cameras.