← 返回 JSSC 论文列表
📄 下载 JSSC 原文 PDF
JSSC 2020第1期Memory40nmCIM

A2 × 30k-Spin Multi-Chip Scalable CMOS Annealing Processor Based on a Processing- In-Memory Approach for Solving Large-Scale Combinatorial

开发了首个基于内存计算方法的2×30k自旋多芯片CMOS退火处理器,用于解决大规模组合优化问题。
40nm CMOS, 22μs退火时间, 比传统CPU快2.6×10^4倍
CMOS退火处理器多芯片内存计算组合优化自旋电路
可扩展高精度自旋算子用于本地通信
使用直接访问SRAM的高度集成自旋电路
不影响退火过程运行时或结果的低延迟芯片间接口
Abstract
The world’s first 2 × 30k-spin multi-chip CMOS annealing processor (AP)—based on the processing-in-memory approach for solving large-scale combinatorial optimization problem—was developed. To expand the bit width of coefficients and enhance the scalability of the AP, it has three key features: an expandable and high-accuracy spin operator for local com- munication, a highly integrated spin circuit using direct access to SRAM, and a low-latency inter-chip interface that does not affect the runtime or results of the annealing process. The AP is fabricated on the basis of 40-nm CMOS technology. It was exper- imentally demonstrated that the spin-flip ratio of the processor agrees well with theoretical values based on the Gibbs distribution over a wide temperature range. As a result, under two-chip operation with 2 × 30k spins, the AP achieves an annealing time of 22 μs, which is 455 times and 2.6 × 10 4 times faster than those achieved by our previous CMOS-AP and a conventional CPU, respectively. Moreover, its energy efficiency is 1.75 × 105 times higher than that of a conventional CPU-based algorithm.