CRII: CNS: Design and System Technology Co-optimization Towards Addressing the Memory Bottleneck Problem of Deep Learning Hardware

CRII：CNS：设计和系统技术协同优化解决深度学习硬件的内存瓶颈问题

• 批准号: 2153394
• 负责人: Mehdi Sadi
• 金额: $ 17.49万
• 依托单位: Auburn University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-07-01 至 2025-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2153394&HistoricalAwards=false
• 关键词: CRII; CNS; Design; System; Technology

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Artificial intelligence and deep learning (AI/DL) are influencing a range of areas, including autonomous vehicles, healthcare, cybersecurity, language processing, robotics, gene editing, climate science, and numerous others. Data size is increasing significantly, yielding ever larger data sets and model sizes to achieve desired levels of AI/DL accuracy. Over the last several years, growth in AI compute capability has far exceeded growth in per-accelerator memory capacity, both on-chip and off-chip. Memory has become the key bottleneck in AI/DL hardware, demanding new approaches to resolve this bottleneck. This project includes two key thrusts: (1) Key performance parameters of on-chip and off-chip memory systems will be co-optimized with AI/DL hardware, considering interactions between the Design and Technology (DTCO), and the overall System and Technology (STCO). (2) Emerging Magnetic Random Access Memory (MRAM), chiplets, and packaging interconnect technologies will be utilized to optimally design the hardware.This project will influence novel paradigms for designing high-performance and energy-efficient AI/DL hardware, impacting the development of new AI/DL algorithms – bringing society one step closer to achieving human-level intelligence in machines. With diminishing returns from Moore’s law, STCO and DTCO have recently become emerging paradigms for tuning the technology for the best performance gains in hardware. The outcomes of this work will be instrumental in enriching scientific knowledge in this field and influence future researchers working on other emerging technical domains. Aligned with the goal of establishing United States’ leadership in the AI/DL domain, the efforts of this project are dedicated to achieving excellence in education, workforce development, and outreach through graduate and undergraduate research, mentoring underrepresented and minority students, and promoting AI hardware education at the K-12 level.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.

该奖项的全部或部分资金来源于《2021 年美国救援计划法案》（公法 117-2）。人工智能和深度学习 (AI/DL) 正在影响一系列领域，包括自动驾驶汽车、医疗保健、网络安全、语言处理、机器人、基因编辑、气候科学和许多其他领域的数据规模正在显着增加，产生了越来越大的数据集和模型规模，以达到所需的人工智能/深度学习精度水平。已经远远超过了每个加速器内存容量（片上和片外）的增长已成为 AI/DL 硬件的关键瓶颈，需要新的方法来解决该瓶颈：（1）关键性能参数。考虑到设计和技术 (DTCO) 以及整体系统和技术 (STCO) 之间的相互作用，片上和片外存储系统的设计将与 AI/DL 硬件共同优化 (2) 新兴磁性随机存取存储器。 (MRAM)、小芯片和封装互连技术将用于优化硬件设计。该项目将影响设计高性能和节能 AI/DL 硬件的新范例，影响新 AI/DL 算法的开发 - 带来随着摩尔定律的回报递减，STCO 和 DTCO 最近已成为调整技术以实现硬件最佳性能提升的新兴范例。这项工作的成果将有助于丰富硬件。该领域的科学知识并影响未来在其他新兴技术领域工作的研究人员，与建立美国在人工智能/深度学习领域的领导地位的目标相一致，该项目致力于通过研究生和本科生在教育、劳动力发展和推广方面取得卓越成就。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

System and Design Technology Co-optimization of Chiplet-based AI Accelerator with Machine Learning

基于Chiplet的AI加速器与机器学习的系统和设计技术协同优化

• DOI: 10.1145/3583781.3590233
• 发表时间: 2023
• 期刊: Proceedings of the Great Lakes Symposium on VLSI
• 作者: Mishty, Kaniz;Sadi, Mehdi
• 通讯作者: Sadi, Mehdi

Designing Efficient and High-Performance AI Accelerators With Customized STT-MRAM

• DOI: 10.1109/tvlsi.2021.3105958
• 发表时间: 2021-10-01
• 期刊: IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS
• 影响因子: 2.8
• 作者: Mishty, Kaniz;Sadi, Mehdi
• 通讯作者: Sadi, Mehdi

Analogy-Guided Evolutionary Pretraining of Binary Word Embeddings

类比引导的二进制词嵌入进化预训练

• 发表时间: 2022
• 期刊: Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers
• 作者: R. Alexander Knipper, Md. Mahadi