Many autonomous systems such as self-driving cars rely on machine learning inference to perform their tasks. These computations are extremely demanding especially when a lot of data from multiple cameras and other sensors need to be processed very fast. To process all this data, special VLSI circuits, called "AI Accelerators" or "Neural Engines", are used. The first part of this lecture will introduce the basic design principles and common architectures of such AI Accelerators.
To meet the performance demand, such accelerators are usually fabricated with modern, cutting-edge VLSI technology that is not as robust against EM noise, soft-errors or wear-out than older technologies. In the second part of the lecture, we will analyze reliability and safety threats that may arise and introduce some key techniques to ensure safe and secure operation of such systems.
自動運転等を支える組み込みシステムの中核である効率的かつ高信頼なDNN推論を実現するためのハードウェアアーキテクチャに関する教育と研究。特に、マトリックス操作の高速化、最先端VLSI技術における信頼性分析、及び、システムの安全かつ確実な動作保証に貢献する重要技術を紹介し、その実現に欠かせない高度なハードウェアアーキテクチャを解説する。
- 教師: 教職員 ホルスト シュテファン