Collaborative Research: SHF: Small: Towards Robust Deep Learning Computing on GPUs

图形处理单元(GPU)已经成为科学模拟和深度学习等许多应用领域最有前途的计算引擎之一。借助GPU提供的海量并行处理能力，大多数最先进的服务器和边缘系统都采用GPU作为深度学习模型训练和推理的核心计算引擎。随着深度学习模型的性能成为决定模型创建者市场收益和模型消费者日常生活便利性的最重要分界线之一，实施可靠和健壮的深度学习计算至关重要。该项目旨在探索挑战和机遇，以解决GPU计算作为深度学习加速器带来的可靠性和隐私影响，并设计轻量级保护方案。该项目的技术目标分为三个方面。第一个重点是探索和评估可能的漏洞及其对基于GPU的深度学习计算的影响。第二个推力通过重新设计GPU构建块来解决计算单元级别的漏洞，例如新的调度算法和激活加速逻辑。第三个重点是探索通信通道和存储子系统中的选择性完整性保护机制，以在CPU和GPU之间传输数据，而不会带来显著的性能开销。提出的解决方案将缓解基于GPU的深度学习计算的体系结构和系统漏洞，使深度学习算法开发者能够更多地专注于性能提升和技术进步，使消费者能够使用基于深度学习的认知产品，而不会考虑隐私问题。这项研究的结果将被整合到本科生和研究生课程以及关于K-12教育的各种推广活动中，并通过开放源码库公开分享。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Graphics processing units (GPU) have become one of the most promising computing engines in many application domains such as scientific simulations and deep learning. With the massive parallel processing power provided by GPUs, most of the state-of-the-art server and edge systems employ GPUs as the core computing engines for deep-learning model training and inference. As the performance of deep learning models becomes one of the most important delimiters that determines market revenue of the model creators and the convenience of daily lives of model consumers, it is critical to enforce reliable and robust deep-learning computation. This project aims to explore the challenges and opportunities to address the reliability and privacy implications of GPU computing as a deep-learning accelerator and design lightweight protection schemes.The technical aims of this project are divided into three thrusts. The first thrust explores and evaluates possible vulnerabilities and their impact on GPU-based deep-learning computing. The second thrust tackles the vulnerabilities at the compute-unit level by redesigning GPU building blocks, such as new scheduling algorithms and activation acceleration logic. The third thrust explores selective integrity protection mechanisms in communication channels and memory subsystems to transfer data between the CPU and GPU without imposing significant performance overhead. The proposed solutions will mitigate architectural and system vulnerabilities in GPU-based deep learning computing, which will enable the deep learning algorithm developers to focus more on performance improvement and technological advancement, and the consumers to use deep learning-based cognitive products without privacy concerns. The findings of this research will be integrated into undergraduate and graduate courses as well as various outreach activities on K-12 education, and publicly shared through open-source repositories.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.