Efficient Computing for AI and Robotics

About this Video

• Video Title: Efficient Computing for AI and Robotics
• Channel: MIT EEMS Group - PI: Vivienne Sze
• Speakers: Vivienne Sze
• Duration: 00:50:51

Introduction

This video lecture discusses energy-efficient computing for AI and robotics, focusing on the shift from cloud processing to edge processing for improved privacy, latency, and accessibility. The core challenge addressed is the high power consumption of current processors, particularly in applications like self-driving cars. The lecture explores cross-layer design strategies to improve energy efficiency in deep neural networks (DNNs) and other AI algorithms.

Key Takeaways

1. Edge Computing Advantages: Processing AI tasks at the edge (on the device) offers advantages over cloud processing, including enhanced privacy, reduced latency, and increased accessibility.
2. Data Movement Dominates Power: The majority of energy consumption in AI processing stems from data movement, not computation. Minimizing data movement is crucial for energy efficiency.
3. DNN Properties for Optimization: Deep neural networks exhibit properties like data reuse and parallelism that can be exploited for optimization. Strategies like row-stationary dataflow effectively manage data movement.
4. Energy-Aware Algorithm Design: Incorporating energy and latency metrics directly into the design of AI algorithms (e.g., energy-aware pruning) leads to greater efficiency gains than focusing solely on the number of operations or weights.
5. Flexible Hardware Design: Given the diversity of AI algorithms, hardware needs to be flexible to adapt to various network architectures, data sparsity levels, and reuse patterns. A hierarchical mesh network-on-chip (NoC) architecture offers flexibility and scalability.
6. Data Compression Techniques: Compressing data based on its properties (dense/sparse, structured/unstructured) is essential to minimize memory requirements and improve efficiency. Different compression methods are suitable for different data types.