Journal of Semiconductors, Volume. 46, Issue 4, 040401(2025)
A leap forward in compute-in-memory system for neural network inference
Fig. 1. (Color online) The AnDi architecture based on a DDFP processor is designed to seamlessly integrate data from both the analogue and digital domains, enabling FP computation[5]. With this architecture, the hybrid mapper has the capability to dynamically allocate matrix multiplication tasks across these domains with fine-grained control at the single row and column level, tailored to the specific requirements of the task. The system outperforms a purely analogue CIM system in terms of floating-point compatibility, precision, and the efficient use of spatial array resources.
Get Citation
Copy Citation Text
Liang Chu, Wenjun Li. A leap forward in compute-in-memory system for neural network inference[J]. Journal of Semiconductors, 2025, 46(4): 040401
Category: Research Articles
Received: Feb. 24, 2025
Accepted: --
Published Online: May. 21, 2025
The Author Email: Liang Chu (LChu), Wenjun Li (WJLi)