Polarization-coupled complete operators in sliding ferroelectric semiconductors for native in-memory attention computing
Abstract
Abstract Attention computing, central to Transformer-based generative AI, calls for native in-memory hardware with linear and nonlinear operators intrinsically coupled. Current compute-in-memory (CIM), however, remains tailored to linear algebra, lacking integrated nonlinear operators. Here, we demonstrate complete-operator hardware that intrinsically couples both operator types in situ via ferroelectric polarization. Our device integrates a vertical ferroelectric tunnel junction for linear multiply-accumulate operations and a lateral ferroelectric semiconductor field-effect transistor for exponential functions in sliding-ferroelectric semiconductor ReS 2 . Tuning lateral transport length and vertical barrier thickness yields designable exponential operators with tunable sensitivity and operating range. A fabricated (9×3)×9 array realizes the first in-memory hardware execution of full Transformer attention primitives, including intrinsic Softmax, with device-level 600-ps latency and 320-aJ/operation energy. A closed-loop hardware prototype enables real-time text generation, with experimentally calibrated simulations supporting neural-signal-to-language scalability. This work embeds the missing nonlinearity directly into memory, advancing CIM toward complete, native hardware for generative AI.
Read the paper
Related studies
- Diarylethenes for Memories and Switches
- Long-Term Potentiation and Memory
- Social Memory Studies: From “Collective Memory” to the Historical Sociology of Mnemonic Practices
- The Cognitive Neuroscience of Working Memory
- Memory aging and brain maintenance
- Effector and Memory CTL Differentiation
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.