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-based processing to edge computing for improved privacy, latency, and global 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, including algorithmic and hardware innovations, to achieve significant energy savings.

Key Takeaways

1. Edge Computing Advantages: Processing data at the edge (on the device) offers benefits in privacy, latency, and accessibility compared to cloud-based processing.
2. Data Movement Dominates Power: The majority of energy consumption in computing stems from data movement, not computation itself.
3. DNN Optimization Strategies: Deep Neural Networks (DNNs) offer high accuracy but demand significant power. Optimizations focus on exploiting data reuse in low-cost memories, using techniques like row-stationary dataflow.
4. Energy-Aware Algorithm Design: Incorporating energy and latency directly into algorithm design (e.g., energy-aware pruning, NetAdapt) yields better efficiency than focusing solely on reducing the number of operations or weights.
5. Flexible Hardware Design: Hardware should be flexible enough to accommodate diverse DNN algorithms and varying levels of sparsity. A hierarchical mesh Network-on-Chip (NoC) architecture enhances flexibility and efficiency.
6. Beyond DNNs: The principles of energy-efficient computing extend beyond DNNs, as demonstrated by efficient visual-inertial odometry for robot navigation. This requires adaptable data storage and processing techniques to handle diverse data characteristics.

The row-stationary dataflow approach improves energy efficiency by balancing data movement for all data types (weights, partial sums, inputs) involved in DNN processing. Unlike other approaches that prioritize keeping one data type stationary (e.g., weights or partial sums), row-stationary dataflow processes one row of filters and one row of input feature maps together. This maximizes convolutional reuse within processing elements, minimizing overall data movement and thus reducing energy consumption. The lecture shows that this method is 1.7x to 2.5x more energy-efficient than other dataflow methods.