Environment and self-driving cars — is there balance?

How Much Does ADS Cost?

Environmental Cost-Benefit of Self-Driving Cars

“Roads filled with automated vehicles could also cooperate to smooth traffic flow and reduce traffic congestion. Americans spent an estimated 6.9 billion hours in traffic delays in 2014, cutting into time at work or with family, increasing fuel costs and vehicle emission.”

I came across this quote while performing research on the alignment, or lack thereof, between the areas of research AI, is heavily funding and those that provide the greatest societal benefit. Governments all over the world are investing heavily in self-driving cars, which are also referred to as Automated Driving Systems (ADS). According to the NHTSA, there are three main areas ADS are meant to improve:

  • Safety
  • Mobility
  • Congestion

Essentially, improving reducing, or eliminating human error-caused crashes would save lives, and protect the quality of life and worker productivity. Improved mobility could improve and increase employment opportunities for the otherwise less mobile. And relieving congestion could stabilize or decrease fuel costs and vehicle emissions. The Federal Highway Administration (FHA) is also testing ADS to address congestion on critical transportation corridors. Washington State and Minnesota consider autonomous vehicles to be “an important tool in our efforts to combat climate change.”

If I came to you and told you that if you buy my device, I can save you $50 a year on your electric bill, what would be your first question? How much does the device cost.

My immediate question after reading through these sites was — What are the carbon reduction targets and how much carbon does it take to generate those savings?

Carbon saving v. carbon cost?

Costs v. Savings

Carbon savings

There are no carbon reduction targets on the NHSTA, FHA, or state websites.

According to a Harvard study, Level 4 self-driving cars would introduce features that could improve fuel economy by 83.5%. Most studies rely on a mixture of direct and indirect environmental factors in their carbon reduction assessments and the results of the myriad factor combinations vary wildly — from 7 to 94% estimated carbon reduction. Although studies have been commissioned, seemingly to directly address this question, currently there is no publicly available estimate of the levels of carbon reduction that have been achieved by currently deployed ADS. Nor is there any information on how much carbon reduction governments are targeting for the technology to achieve over the next 5 to 10 years in order to assist them in meeting their carbon emissions goals.

Carbon costs

Computing is not carbon-neutral technology. ADS are trained on data that is collected from test vehicles. The test vehicle data must be collected, labeled, and preprocessed before it can be used for training. Using NVIDIA’s conservative estimates for the length of time required to train an ADS, the most energy-efficient training configuration (AlexNet on a single Tesla P100), would require 2,630 kWh of electricity over 50 cycles of training. Which is the equivalent of 210 gallons of gas:

Estimated carbon emissions over a single 50 epoch training cycle (Data sources listed below)

While these numbers are significant, they don’t provide a complete picture of how much CO2 is emitted by a fleet of self-driving cars. In order to generate the massive amounts of data that are required to train an ADS, test vehicles are loaded with sensors and computing systems. This can increase the vehicle’s energy use by 3–20%. Also, the estimate only considers the carbon emitted training a single model for 50 epochs. 50 epochs is not the standard number required for model convergence, often models require significantly more epochs to reach convergence.

ADS are not set-it-and-forget-it systems.

Training is only one stage of the continuous development process. Every stage in the automotive AI process requires compute time and produces CO2 emissions:

Automotive AI Engineering Workflow (NVIDIA)

New models are constantly built and trained so that there is a reliable stream of improved models to be optimized and deployed.

Publicly available system costs

There is no information on ADS carbon emissions on the NHSTA, FHA, or state websites. Research papers on the topic tend to be theoretical. Government studies can be highly speculative, listing a myriad of external factors in their impact assessments. There are studies on the potential upstream effects on fuel supply markets, but there are no studies that focus on the current energy impacts of ADS on the environment.

I reached out to my network, no one had any direct information. A snippet from a presentation by Tesla’s Director of AI that gave specs on their current data center setup, an article on Volvo’s new ADS data center that will be able to store 225 gigabytes of data, an estimate of how much data a Level 5 ADS system would generate per hour, a research-focused impact tracker and CO2 estimator. Breadcrumbs that we are forced to cobble together to try to glean enough information to answer a question that the industry has simple answers for.

Transparency

AI companies know how much carbon their ADS emit and how much carbon reduction they’re likely to create.

Self-driving cars have been on roadways since 2014. The major companies in the industry have significant data on the emissions generated by their systems and cars, along with data on how their systems have affected traffic patterns and congestion in test areas. Tech companies also know how much carbon reduction their systems are likely to produce in the near term, given test car specs, the current state of relevant external factors, and near-term estimated rates of adoption. Both of these issues must be well understood for tech companies to manage their businesses, responsibly advise their clients and partners, and keep their climate change commitments. So, if they know, why don’t we?

Governments know how much carbon reduction they need ADS to generate to meet their carbon emissions goals.

Governments have emissions targets, and they seem to believe that ADS can help them meet those targets. In order to assess whether the targeted reductions are possible, governments must ensure that ADS can meet some specified carbon reduction threshold. Those thresholds should be made public to allow for effective oversight and accountability.

Effective oversight requires data — realistic estimates based on current facts.

Currently governments seem to be settling for industry estimates of potential savings that are calculated utilizing a panoply of factors that may or may not be likely to occur. That is irresponsible and unnecessary. Governments have a variety of means for requiring industry to provide necessary data. For instance, governments can include the following statements in any contract, RFQ, RFP, or funding application:

Please provide an estimate of the amount of carbon dioxide reduction that is expected in [preferred terms] over [set time frame(s)].

Please provide an estimate of the amount of carbon dioxide that will be expended in achieving that reduction in [preferred terms] over [set time frame(s)].

Transparency is non-negotiable.

Daily we are presented with evidence that tech companies — by default and intentionally — act in their own interest, regardless of the societal implications and human costs. If we are to have a future as a species, we cannot rely on the goodwill of individuals or industries. We need to make informed decisions for the best interest of our planet and our citizens. That starts with transparency. Simple answers to simple questions, using current data and current technological capabilities.

Data Sources — Carbon Emissions Chart

Data for GPU configuration, network type, and training time found here: https://developer.nvidia.com/blog/training-self-driving-vehicles-challenge-scale/

Data for estimated power consumption for GPU configurations found here: https://images.nvidia.com/content/tesla/pdf/nvidia-tesla-p100-PCIe-datasheet.pdf; https://developer.nvidia.com/blog/dgx-1-fastest-deep-learning-system/

Data for greenhouse gas equivalence is found here:
https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Masheika Allgood

Masheika Allgood

24 Followers

AI Ethicist — Making AI accessible by using plain language to discuss complex topics. Advocating for an AI that is for all.