講演情報
[15a-PB1-8]Large-scale integration of MoS2–Hf0.5Zr0.5O2 ferroelectric memristors for in-memory image denoising in neural networks
〇(D)jeehwan lee1,2, Do Kyung Yoon1, Woo Jong Yu1 (1.SKKU Univ., 2.Samsung Electronics)
キーワード：
Ferroelectric memristor array、Transition metal dichalcogenide channels、In-memory Computing
In-memory computing using floating-gate memristor (FGMEM) arrays with 2D channels enables large-scale parallelism, low energy consumption, efficient hardware-level matrix–vector multiplications. However, tunneling-based FGMEM faces high operating voltage, slow programming, low on/off ratio, and a small memory window. Here, we demonstrate a 16×16 ferroelectric memristor (FeMEM) array integrating a 2D MoS2 channel with a hafnium–zirconium oxide (HZO) layer for in-memory image denoising inference. The array achieves 91% yield and excellent uniformity in ON–OFF current and forward/reverse threshold voltage with Gaussian distributions within ±3σ of the mean. The FeMEM exhibits a large memory window (32.3%) and a high Ion/Ioff ratio of 1.02×10^5-about 2 and 1000 times higher than an FGMEM, respectively. It shows high linearity (β=0.14/0.68 for long-term potentiation/depression) and operates with ±1.5 V, 10 μs pulses-significantly lower and faster than FGMEMs (>±4 V, ~100 ms). Parallel-multiply (PM) operations between input voltages (V) and conductance (G) yield an output current map (I = G×V) corresponding to a denoised image achieving quality gains of +13.03 dB PSNR, +0.226 SSIM, and 20 times MSE reduction after denoising-approaching the quality of ideal images. As an in-memory preprocessing for a CNN, this denoising boosts classification accuracy from 55.7% (noisy inputs) to 95.5% (denoised inputs)