The Architectural Implications of Autonomous Driving

Slides by:Sukrit Kalra sukrit.kalra@berkeley.edu

�1

Challenges

“Correct” Decisions

“Real-Time” Decisions

Power Budgets

!2

Challenges

“Correct” Decisions

“Real-Time” Decisions

Power Budgets

!3

Computational Pipeline

!4

Camera

LIDAR

Radar

GPS

Sensing

Object Detection

Lane Detection

Traffic Light Detection

Traffic Sign Detection

Localization

Perception

Route Planning

Motion Prediction

Behavior Planning

Trajectory Planning

Planning

Control

Computational Pipeline

!5

Camera

LIDAR

Radar

GPS

Sensing

Object Detection

Lane Detection

Traffic Light Detection

Traffic Sign Detection

Localization

Perception

Route Planning

Motion Prediction

Behavior Planning

Trajectory Planning

Planning

Control

Computational Pipeline

!6

Camera

LIDAR

Radar

GPS

Sensing

Object Detection

Lane Detection

Traffic Light Detection

Traffic Sign Detection

Localization

Perception

Route Planning

Motion Prediction

Behavior Planning

Trajectory Planning

Planning

Control

Computational Pipeline

!7

Camera

LIDAR

Radar

GPS

Sensing

Object Detection

Lane Detection

Traffic Light Detection

Traffic Sign Detection

Localization

Perception

Route Planning

Motion Prediction

Behavior Planning

Trajectory Planning

Planning

Control

Design Constraint: Performance

!8

Camera

LIDAR

Radar

GPS

Sensing

Sensors generate 1-2GB of data per second. Autonomous vehicle system should provide:

High Throughput Low Latency

!9

Camera

LIDAR

Radar

GPS

Sensing

Object Detection

Lane Detection

Traffic Light Detection

Traffic Sign Detection

Localization

Perception

Route Planning

Motion Prediction

Behavior Planning

Trajectory Planning

Planning

Control

Tail latency <= 100ms

Design Constraint: Predictability

!10

Localization

Design Constraint: Storage

41 TB!

4TB of logging data generated by a car each day.

!11

Design Constraint: Thermal and Power

“Datacenter on wheels.”

Power requirements of up to 3kW.

!12

Key Idea

FPGAsGPUs ASICs

Exploit different accelerator platforms to achieve predictability and performance while reducing the power requirements.

!13

Implementation

Detector Tracker

Localization

!14

Metrics of Success

!15

DiscussionAre DNNs the major computational part of the perception pipeline?

!16

DiscussionWhat is the reason for zero variability in both ASICs and FPGAs?

Mean Latency 99.99th Percentile Latency

What is the source of runtime variability for the GPU?

!17

DiscussionWhat about the cost of inter-component communication across the devices?