Implementation of offset corrected AGAD algorithm on 130-nm CMOS technology-based RRAM array for analog neural network training.

We present the first implementation of the analog gradient accumulation with dynamic reference (AGAD), reported as the most advanced and highest-performing version of the TT (Tiki-Taka) algorithm, on an HfO2-based resistive random-access memory (RRAM) array for analog neural network training. Through comparative simulations with recent versions of the TT algorithm, we verify that only AGAD enables rapid and accurate gradient accumulation, even in the presence of reference errors. This robustness is enabled by the chopper technique and moving-average-based dynamic digital reference. For demonstration, we propose and fabricate a novel 4T1R unit-cell crossbar array that offers excellent retention while eliminating half-select disturbances. Notably, we experimentally implement the bidirectional weight update behavior of AGAD on hardware and extend it to an array-level weight transfer task, achieving successful weight convergence with a low loss of 0.005, essential for neural network training. This study emphasizes that AGAD can provide a viable pathway toward practical in-memory computing under realistic device limitations.