Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Managing Bandwidth: The Key to Cloud-Assisted Autonomous Driving

Alexander Krentsel, Peter Schafhalter, Joseph E. Gonzalez, Sylvia Ratnasamy, Scott Shenker, Ion Stoica

TL;DR

This work identifies an opportunity to offload parts of time-sensitive and latency-critical compute to the cloud, which requires carefully allocating bandwidth to meet strict latency SLOs, while maximizing benefit to the car.

Abstract

Prevailing wisdom asserts that one cannot rely on the cloud for critical real-time control systems like self-driving cars. We argue that we can, and must. Following the trends of increasing model sizes, improvements in hardware, and evolving mobile networks, we identify an opportunity to offload parts of time-sensitive and latency-critical compute to the cloud. Doing so requires carefully allocating bandwidth to meet strict latency SLOs, while maximizing benefit to the car.

Managing Bandwidth: The Key to Cloud-Assisted Autonomous Driving

TL;DR

This work identifies an opportunity to offload parts of time-sensitive and latency-critical compute to the cloud, which requires carefully allocating bandwidth to meet strict latency SLOs, while maximizing benefit to the car.

Abstract

Prevailing wisdom asserts that one cannot rely on the cloud for critical real-time control systems like self-driving cars. We argue that we can, and must. Following the trends of increasing model sizes, improvements in hardware, and evolving mobile networks, we identify an opportunity to offload parts of time-sensitive and latency-critical compute to the cloud. Doing so requires carefully allocating bandwidth to meet strict latency SLOs, while maximizing benefit to the car.

Paper Structure

This paper contains 8 sections, 2 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. How Autonomous Driving Works
  3. A Pre-emptive Q&A
  4. Feasibility: A Path Forward
  5. Performance
  6. Cost
  7. Utility-Informed Method
  8. Discussion

Figures (3)

  • Figure 1: Cloud accelerators can execute more accurate models with lower runtime than hardware designed for autonomous driving. We plot the runtimes of different models in the the EfficientDet tan20efficientdet family of object detection models when executing on the NVIDIA's edge-focused Jetson Orin and cloud-focused H100 GPU. Prior work has shown that more accurate object detectors are able to improve driving safety by identifying obstacles at greater distances erdosschafhalter2023leveraging.
  • Figure 2: User-experienced cellular network speeds are improving according to Ookla's SpeedTest.net dataset ookla-network-performance. We find improvements in upload, download, and ping speeds across the 10th, 50th, and 90th percentiles between 2019 and 2024, capturing the expansion of 5G in 2019 and the shut-down of 3G in 2022. As the data aggregates connections ranging from 2G to 5G, we are unable to report performance by network generation.
  • Figure 3: Utility curves for a service with 3 object detection models.