Abstract
Analog in-memory arrays can reduce the energy of neural inference, but efficient deployment of pretrained networks on hardware is still limited by slow, sequential weight programming and chip-specific mismatch. Weight import and fine-tuning are often handled by different update modes, which increases end-to-end latency and energy in practical workflows. Here we demonstrate a unified outer-product pulse engine in a bidirectional dual-gate charge-trap flash transistor array that executes both low-rank SVD-based weight import and gradient-descent fine-tuning using a single matrix-parallel update primitive. In fabricated arrays, parallel outer-product programming achieves 93% agreement within a 2σ target window and transfers a rank-2 pattern in two update steps. In a closed-loop on-wafer Iris classifier, low-rank import reduces the required fine-tuning iterations from 69 to 3 and cuts training latency by 23×. Scaling analysis projects > 6.4× lower transfer latency, > 64× lower training-update latency, and > 209× lower total training energy for a 128×128 array near 90% accuracy. Together, these results establish an end-to-end, array-native workflow that unifies pretrained weight import and on-device fine-tuning through a single matrix-parallel outer-product primitive.

This publication has 37 references indexed in Scilit: