ACED Fab: Ultrafast, low-power AI chip with a new class of MRAM for learning and inference at edge

ACED Fab：超快、低功耗 AI 芯片，配备新型 MRAM，用于边缘学习和推理

• 批准号: 2314591
• 负责人: Shan Wang
• 金额: $ 55万
• 依托单位: Stanford University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-07-01 至 2026-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2314591&HistoricalAwards=false
• 关键词: ACED; Fab; Ultrafast; low; power

The collaborative team of this project under Advanced Chip Engineering Design and Fabrication (ACED Fab) program will be working on an exciting advancement in the field of artificial intelligence (AI) for edge computing, such as secured machine learning based on personalized or sensitive data in smartphones (a type of edge devices) without resorting to a server at a remote data center. The project introduces a new class of Magnetoresistive Random Access Memory (MRAM) called Spin-transfer torque (STT) Assisted Spin-orbit torque (SOT)-MRAM (SAS-MRAM) which features ultralow power consumption and ultrafast write speeds. By co-designing SAS-MRAM with CMOS circuits, the project aims to create energy-efficient edge AI systems. SAS-MRAM's non-volatile nature eliminates standby leakage power, making edge-AI chips more energy-efficient at the system level compared to existing approaches using Static Random Access Memory (SRAM). The project’s activities extend beyond technological advancements, with plans of K-12 STEM outreach, undergraduate/graduate training, curriculum development in innovation and entrepreneurship, and broadening participation of underrepresented minority groups in the microelectronics STEM field and semiconductor industry. The team’s efforts in education and inclusivity will contribute to a diverse and innovative future of the microelectronics industry. The new SAS-MRAM with ultralow power and ultrafast write speed will be co-designed with CMOS circuits for energy-efficient edge AI applications. The SAS-MRAM will be fabricated on top of a TN40G CMOS die through a custom back-end-of-line (BEOL) process. The team will systematically perform micromagnetic simulation and HSpice simulation to build Process Development Kits (PDKs) required for co-designing SAS-MRAM and CMOS circuits. Furthermore, the project will leverage SAS-MRAM to design, optimize, and tape-out an In-Memory Computing (IMC) chip prototype for edge-AI, which could implement both on-chip inference and training computation. Finally, the project will develop new continual learning algorithms that could minimize the memory weight updates (i.e., memory writes) and computing complexity, allowing the AI system to learn new data without forgetting previously learned knowledge. The resulting edge-AI chips will be significantly more energy-efficient at system level than the prevalent counterparts based on SRAM due to zero standby leakage power for non-volatile MRAM. On-device training/learning based on SAS-MRAM is potentially ultrafast due to lower latency from denser bit cells and multi-bit writing with shared SOT write lines. The project can potentially revolutionize edge AI devices and systems by leveraging SAS-MRAM and in-memory computing to create energy-efficient AI systems with improved performance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

该项目的协作团队在高级芯片工程设计和制造（ACED FAB）计划下将致力于用于边缘计算的人工智能（AI）领域的令人兴奋的进步，例如基于智能手机（一种类型的边缘设备）的个性化或敏感数据的安全机器学习，而无需诉诸远程数据中心的服务器。该项目引入了一类新的磁磁性随机访问记忆（MRAM），称为自旋转移扭矩（STT）辅助自旋轨扭矩（SOT）-MRAM（SAS-MRAM），该扭矩（SAS-MRAM）具有超级功率消耗和超快写入速度。通过与CMOS电路共同设计SAS-MRAM，该项目旨在创建节能边缘AI系统。 SAS-MRAM的非易失性性质消除了备用泄漏功率，与使用静态随机访问存储器（SRAM）相比，在系统水平上使边缘芯片在系统级别上更加节能。该项目的活动超出了技术进步，随着K-12 STEM外展，本科/研究生培训，创新和企业家的课程开发的计划，并扩大了代表性不足的少数群体在微电脑STEM机构STEM领域和半导体行业中的参与。该团队在教育和包容性方面的努力将有助于微电子行业的多样性和创新的未来。具有超速功率和超快写速速度的新型SAS-MRAM将与CMOS电路共同设计用于节能边缘AI应用。 SAS-MRAM将通过定制的后端（BEOL）过程在TN40G CMO的顶部制造。该团队将系统地执行微磁模拟和HSPICE模拟，以构建共同设计SAS-MRAM和CMOS电路所需的过程开发套件（PDK）。此外，该项目将利用SAS-MRAM来设计，优化和胶带，用于Edge-AI的内存计算（IMC）芯片原型，该原型可以实现芯片推理和训练计算。最后，该项目将开发新的连续学习算法，这些算法可以最大程度地减少内存重量更新（即存储器写入）和计算复杂性，从而使AI系统可以学习新数据而不忘记先前学习的知识。由于非挥发性MRAM的零备用泄漏功率，在系统水平上所得的边缘芯片在系统水平上的节能效率将明显高于基于SRAM的普遍芯片。基于SAS-MRAM的设备训练/学习可能是由于较密集的位单元细胞的潜伏期和带有共享SOT写入线的多位写作的潜伏期。该项目可以通过利用SAS-MRAM和内存计算来创建具有改善性能的能源有效的AI系统来彻底改变Edge AI设备和系统。该奖项反映了NSF的法定任务，并被认为通过使用基金会的知识分子优点和更广泛的影响来评估来获得珍贵的支持。

项目成果

Slimmed Asymmetrical Contrastive Learning and Cross Distillation for Lightweight Model Training

用于轻量级模型训练的精简非对称对比学习和交叉蒸馏

• 发表时间: 2023
• 期刊: Thirty-seventh Conference on Neural Information Processing Systems
• 作者: Meng, Jian;Yang, Li;Lee, Kyungmin;Shin, Jinwoo;Fan, Deliang;Seo, Jae-sun
• 通讯作者: Seo, Jae-sun