Year: 2026
Type: article
Abstract: ABSTRACT Compared with conventional planar chips, monolithic three-dimensional (3D) integration offers significant improvement in chip area, latency, and energy efficiency. However, the realization of such vertically stacked architectures is challenging due to the stringent thermal budget. Meanwhile, although resistive random-access memory (RRAM)-based in-memory computing (IMC) has emerged as a promising paradigm to overcome the memory bandwidth and energy bottlenecks in deep neural network inference, its practical deployment is hindered by the pronounced device-to-device variation (DDV) and cycle-to-cycle variation (CCV), which degrade the inference accuracy. To address these challenges, we experimentally demonstrate a monolithic 3D IMC platform based on a 1–transistor–4–vertical RRAMs (1T–4R) structure, in which each two-dimensional (2D) molybdenum disulfide field-effect transistor drives four vertical RRAMs, with the whole structure fabricated under a low thermal budget below 300°C. This architecture significantly improves memory density and reduces data movement overhead for in-memory multiply-and-accumulate operations. Furthermore, to mitigate inference accuracy degradation caused by DDV, a DDV-aware optimal mapping strategy for binary coding is proposed, which minimizes weight representation error without introducing additional memory cells or area overhead. Inference simulations on Vision Transformer (ViT) show that the top-5 accuracy degradation remains within 5% under the worst variation conditions considering both DDV and CCV. Hardware performance estimation of the 3D structure further demonstrates substantial reductions in chip area, inference latency, and energy consumption, with energy efficiency improved by more than 62% compared with conventional planar neural network accelerators.
Source: Chip
Institution: Shanghai Jiao Tong University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,000
OpenAlex ID: W7202374364