Low-power Charge-domain In-FeRAM Computing

The growing energy demands of large-scale model training and scientific computing have inspired interest in low-power computing paradigms. In-memory computing (IMC) reduces data movement overheads, but resistive devices suffer from high static power and Joule-heating. Here, we introduce a charge-domain computing paradigm based on Hf0.5Zr0.5O2 ferroelectric random-access memory (HZO FeRAM), in which computation is enabled by linear accumulation of charge associated with the anti-orientated ferroelectric polarization. We experimentally demonstrate this concept using a 16-kbit HZO FeRAM array, where key computational characteristics execute directly within this array, gradient calculations are performed with errors below 6×10 − 3, and in-memory difference operations solve partial differential equations with 98% accuracy after only 21 iterations. Hybrid FeRAM-digital training of a single-layer neural network for letter recognition is also shown. Furthermore, a 1 μm2 HZO ferroelectric capacitor exhibits a measured switching energy of only 1.66×10 − 12 J, implying the potential of HZO FeRAM for energy-efficient charge-domain computing.