Cheng Yang, Shu Qingyuan, Lee Albert, He Haoran, Zhu Ivy, Chen Minzhang, Chen Renhe, Wang Zirui, Zhang Hantao, Wang Chih-Yao, Yang Shan-Yi, Hsin Yu-Chen, Shih Cheng-Yi, Lee Hsin-Han, Cheng Ran, Wang Kang L
Department of Electrical and Computer Engineering, University of California, Los Angeles, CA, USA.
Department of Physics, The Ohio State University, Columbus, OH, USA.
Nat Commun. 2025 May 30;16(1):5022. doi: 10.1038/s41467-025-58932-x.
Neuromorphic diffusion models have become one of the major breakthroughs in the field of generative artificial intelligence. Unlike discriminative models that have been well developed to tackle classification or regression tasks, diffusion models aim at creating content based upon contexts learned. However, the more complex algorithms of these models result in high computational costs using today's technologies. Here, we develop a spintronic voltage-controlled magnetoelectric memory hardware for the neuromorphic diffusion process. The in-memory computing capability of our spintronic devices goes beyond current Von Neumann architecture, where memory and computing units are separated. Together with the non-volatility of magnetic memory, we can achieve high-speed and low-cost computing, which is desirable for the increasing scale of generative models in the current era. We experimentally demonstrate that the hardware-based true random diffusion process can be implemented for image generation and achieve comparable image quality to software-based training as measured by the Fréchet inception distance (FID) score, achieving ~10 better energy-per-bit-per-area over traditional hardware.
神经形态扩散模型已成为生成式人工智能领域的重大突破之一。与已得到充分发展以处理分类或回归任务的判别模型不同，扩散模型旨在基于所学上下文创建内容。然而，这些模型更为复杂的算法在当今技术下会导致高昂的计算成本。在此，我们为神经形态扩散过程开发了一种自旋电子电压控制磁电存储器硬件。我们的自旋电子器件的内存计算能力超越了当前内存和计算单元分离的冯·诺依曼架构。结合磁存储器的非易失性，我们能够实现高速且低成本的计算，这对于当前时代生成模型规模的不断扩大是非常理想的。我们通过实验证明，基于硬件的真实随机扩散过程可用于图像生成，并且通过弗雷歇 inception 距离（FID）分数衡量，能实现与基于软件的训练相当的图像质量，与传统硬件相比，每比特每面积的能量提高了约 10 倍。
