There has been great interest in industry and academia in characterizing the emissions from data centers [1], [2], especially with respect to increased workloads expected from deep neural networks (DNNs) [3]. In 2018, data centers collectively consumed an estimated 205 TWh or 1% of the world's electricity [4] and contributed to about 0.3% of the world's emissions [1], with demand expected to grow [5]. However, less attention has been paid to the carbon footprint of computing in the emerging field of autonomous vehicles (AVs).
There is reason to expect the amount of computing will be significant onboard Level 4 or Level 5 AVs, where a human back-up driver is unnecessary [6]. For a global fleet of AVs, the overall computing workload is comparable and may even exceed current data centers' workloads if AVs are widely adopted. For example, Facebook runs trillions of DNN inferences per day across its data centers [2]; an AV that drives for an hour per day computing 10 DNN inferences at 60 Hz on each of the inputs of 10 cameras would make 21.6 million inferences per day, and one billion AVs would make 21.6 quadrillion inferences per day! Due to the computing capability onboard, AVs have even been referred to as "data centers or supercomputers on wheels" [7]. While previous works have explored the carbon impact of AVs due to changes in driving strategies [8], [9], [10], this work is the first to characterize the carbon emissions from the computers onboard the AV itself. The contributions of this paper are threefold. First, we introduce an open-source framework to probabilistically model emissions from computing onboard AVsfoot_0 . Second, based on our literature survey to estimate parameters in the model, we find that emissions from computing onboard AVs have the potential to be comparable to that of all data centers today. We recommend targets for computer power and rate of hardware energy efficiency improvement for various scenarios. Finally, we discuss several avenues of future research unique to AVs to better characterize and reduce emissions from computing onboard AVs.
Generating the electricity needed to run the computers onboard a global fleet of AVs introduces a source of carbon emissions. The carbon dioxide equivalent (CO 2 e) tons emitted per year from computing onboard a fleet of N AVs is given by G = αN P QI,
where G is the CO 2 e tons emitted per year from computing, P is the average computer power for each AV, Q is the average hours per day driven by each AV, I is the average carbon intensity of the electricity used by the AVs or the grams of CO 2 e emitted to produce 1 kWh of electricity, and α = 3.65×10 -7 is a constant that captures unit conversions. Note, Eq. 2 . We consider the following constant baselines for emissions: 1) 0.14 Gt CO 2 e, or all GHG emissions from data centers in 2018 [4], [11], 2) 0.52 Gt CO 2 e, or 1% of all GHG emissions (not including land use change) in 2019 [12], and find it would take 335 million AVs for AV computing emissions to equal 2018 data center emissions and 1.25 billion AVs for AV computing emissions to equal 1% of 2019 emissions.
Next, we sweep over the computing power for the autonomy stack and plot the emissions from computing onboard AVs in Fig. 1. With one billion AVs, less than the number of cars today [A1], the computer power must be less than 0.84 kW to have computing onboard AVs contribute less emissions than data centers. The variables in Eq. 1 are not exactly known, and there is large uncertainty with respect to future trends. We now probabilistically model each variable in Eq. 1 based on current trends (summarized in Table I) and model different scenarios based on future trends (summarized in Table II
1) Current trends: The average computer power consumed by each AV depends on the workload of the autonomy stack and the hardware energy efficiency of the computer. The autonomy workload consists of perception, localization, planning, and control [C13]. This workload is challenging to model since solving Level 4 or Level 5 autonomy is still an active area of research [C1, C2] and is proprietary for industry. For perception tasks such as object tracking and semantic segmentation, DNNs are the leading approach in computer vision [C13, C14]. For other sensors and tasks such as planning, there is ongoing debate on whether these components will remain DNN-based, non-DNN based, or a hybrid version of both [C13]. Since recreating an entire stateof-the-art autonomy stack is beyond the scope of this paper, we choose to model only the DNN portion workload which likely plays a substantial role in the autonomy stack.
We select a multitask DNN architecture with a shared encoder and separate decoders for each task based on its popularity in academia and industry [C3, C4]. We use EfficientNet-B0 as the encoder [C5] and DeepLabV3 heads as the decoders [C6]. We consider the number of tasks T to be the number of decoders on the autoencoder and the number of cameras C to be the number of times we run the autoencoder. We measure the power P meas (T ) and latency L meas (T ) of the autoencoder at various values of T for 1344 × 1344 resolution inputs, a resolution close to that found in AV benchmarks [C15, C16]. Based on the desired throughput F of the full autonomy stack, we model multiple computers needed to achieve the desired throughput. Finally, we scale the hardware energy efficiency by multiplying by η, the ratio of the tera operations per second (TOPS) per Watt of the measured hardware and the TOPS per Watt of the target hardware, as seen in
While TOPS per Watt is known to not be a holistic measure [D1], we use it as an approximation since it is expensive and difficult to get access to state-of-the-art AV hardware and hardware not yet in production. We measure P meas and L meas on an Nvidia RTX 2080 Ti and scale the hardware energy efficiency for the Nvidia DRIVE Orin system to be the target platform, such that η = 0.344 [D2, D3]. When we substitute Eq. 2 for P in Eq. 1, we obtain
We model C, 2) Future trends: There is uncertainty around how autonomy workloads will change over time since Level 4 or 5 autonomy remains unsolved. If there is a paradigm shift due to a breakthrough technology, DNNs may not make up the majority of the autonomy workload. However, given significant investment in DNNs from industry and academia [I6, I7, H10, C14] and difficulty predicting breakthrough technologies, we assume DNNs will likely remain a large component of the autonomy workload.
In general, DNNs have gotten larger over time in domains such as NLP and recommendations [I1]. An exponential scaling of DNN parameters may be required for a linear gain in accuracy [3]; for such a safety-critical system, a slightly more accurate DNN may be preferable even if it is much larger. Moreover, higher resolution cameras processed at higher frame rates allow AVs to see farther and drive faster [I2, I3], and there likely will need to be uncertainty estimation and redundancy built into the hardware and algorithms [I4, I8]. On the other hand, the growth of workload size may be slowed down due to methods such as pruning and network architecture search [13] if they can maintain metrics important for safety (e.g., accuracy, robustness, uncertainty quality) while decreasing latency. We model the workload increasing by multiplying L meas by a factor a and we sweep over values for a 2025-2050 with doubling times equal to 3, 5, and 10 years.
To model how η will change over time, we model the rate of increase in TOPS per Watt of AV hardware. Based on historical patterns until 2009, Koomey's law states that TOPS per Watt doubles on average every 1.6 years [H1]. However, the slowdown of Dennard scaling and Moore's law has made keeping up with this rate of improvement challenging [H2]. We plot the natural log of the reported TOPS per Watt for current and announced AV hardware platforms and their production dates in Fig. 2. We fit a linear model to find the average doubling rate of TOPS per Watt for AV hardware to be 2.8 years. We model η 2025-2050 at various rates of hardware energy efficiency doubling including at the current pace.
C. Average Time Driven (Q) 1) Current trends: American vehicles were driven on average 0.79 hours per day according the 2017 National Household Travel Survey (NHTS) [E1]. There is uncertainty in how driving behavior may change in response to the widespread adoption of AVs. Projections for changes range from -35% to 40% [E2] due to an increase in driving due to multitasking [E3] and expansion of the transportation userbase to individuals who currently face limited mobility, or a decrease in driving due to increased car-sharing [E2, E4, E5]. We represent the range of -40% to 40% as a 95% confidence interval to obtain Q ∼ N (0.79, 0.03).
2) Future trends: We model a scenario where the average hours driven by AVs increases by 14% due to expansion of driving to under-served populations [J1] and stays constant over time such that Q 2025-2050 = 0.90.
1) Current trends: Generating electricity to power the computers onboard the AVs generates carbon emissions, whether generated from gasoline for gasoline-powered vehicles or generated by the mix of energy sources that power the electric grid used to charge electric vehicles. We use the global average carbon intensity of electricity generation to capture the average carbon intensity across all AVs. Due to changes in economic activity due to the COVID-19 pandemic, we use the average global carbon intensity estimate for 2019, select a variance that captures differences in carbon intensity estimates between years [F4], and set I ∼ N (471, 25).
2) Future trends: We model four different scenarios for carbon intensity I 2025-2050 using different annual decarbonization rates from 471 g CO 2 e/kWh: 1) 1.5% (business as usual, 2019 rate), 2) 2.5% (2020), 3) 8.1% (consistent with 2 degrees of warming), 4) 12.9% (consistent with 1.5 degrees of warming) [K1].
In this section, we present results from applying the carbon emissions framework using the parameters discussed in Section II. We first look at modeling emissions based on current trends and varying the adoption rate of AVs. We use a Monte Carlo simulation with one million samples to estimate the distribution of emissions G, seen in Fig. 3a for the workload with mean 8 cameras and 10 tasks at 60 Hz. The expected value of the distribution is approximately twice 2018 data center emissions when adoption rate p n is 0.95; in 90% of the scenarios, emissions were greater than 88% of 2018 data center emissions. In Fig. 3b, we set p n = 0.95 and vary the average number of cameras and tasks; the larger workloads lead to surpassing 2018 data center emissions and 1% of 2019 emissions with high probability. We also estimate the probability distribution of the computer power P given emissions G is equal to the baselines, as seen in Fig. 3c and Fig. 3d. The computer power must be less than 1.2 kW in order for the emissions to be less than 2018 data center emissions for 90% of the simulated scenarios at p n = 0.95.
Next, we compute the emissions over time in various scenarios with future trends. In Fig. 4, assuming a businessas-usual decarbonization rate and the workload doubling every three years, we sweep over various values for the halflife of η, or equivalently, the doubling time of the hardware energy efficiency, for the high adoption scenario (Fig. 4a) and medium adoption scenario (Fig. 4b). Maintaining the current average rate of hardware energy efficiency increase of 2.8 years in both scenarios leads to large emissions by 2050. In order to keep the emissions from computing onboard AVs in 2050 under 2018 data center emissions or under 1% of 2019 total emissions in the high adoption scenario, hardware energy efficiency must double faster than 1.1 years or 1.4 years respectively.
Next, we assume hardware energy efficiency doubles every 2.8 years and the high adoption scenario. In Fig. 4c, assuming a business-as-usual decarbonization rate, we sweep over the doubling time of the workload. We see a slower rate of increase of the workload yields lower emissions. Finally, in Fig. 4d, assuming the workload doubles every 3 years, we sweep over different decarbonization rates. An aggressive decarbonization rate lowers emissions, but even that scenario cannot keep AV computing emissions below that of 2018 data center emissions. A business-as-usual decarbonization rate yields high emissions in this scenario.
Business-as-usual trends alone are not enough to contain the operational carbon emissions from computing onboard AVs in various scenarios presented in Section III. We highlight several future research directions unique to AVs to help better characterize and potentially decrease the carbon footprint from computing onboard AVs.
1) Characterize emissions from sensing: Unlike data center servers, AVs also must sense their environment, and the power consumption of the sensors can be non-negligible for sensors such as LiDAR. Characterizing emissions for current and future trends in sensor suites will help capture the complete operational carbon footprint of AVs.
2) Characterize embodied vs. operational carbon: With the current average lifespan of a car ranging from 10 to 20 years [G1], AVs will likely have much longer lifespans than that of data center servers and mobile devices. An analysis of the embodied carbon emissions from manufacturing the computers and sensors onboard AVs would not only help cap- ture the total carbon footprint, but also enable comparisons between operational carbon and embodied carbon over the lifespan of an AV. For example, it was found that operational carbon dominates the carbon footprint of mobile devices at device lifespans over 4 years [14]; characterization of AV embodied carbon will shed light on whether operational carbon likewise dominates AVs' total carbon footprint over its longer lifespan. In addition, understanding embodied carbon can suggest the impact of usage patterns such as car-sharing. For example, one billion privately-owned AVs driving one hour per day results in the same amount of operational carbon emissions as 100 million shared AVs driving 10 hours per day. However, the second case would result in lower embodied carbon emissions since fewer components need to be manufactured.
3) Explore trade-off between hardware specialization and generalization: Unlike data center servers, the computers onboard AVs handle constant workloads that are known ahead of time, presenting an opportunity for hardware specialization. The design of accelerators specific for autonomy tasks can deliver large reductions in energy consumption and help maintain a high rate of increase in hardware energy efficiency despite the slowdown of Dennard scaling and Moore's law [15] for both DNN and non-DNN workloads [13], [16], [17]. However, since AVs will have longer lifespans [G1], hardware will still need to maintain some ability to generalize to future workloads. 4) Explore algorithmic efficiency improvements without sacrificing safety: AV workloads are safety-critical and cannot tolerate a decrease in performance in metrics relevant to human safety (e.g., accuracy, latency, robustness, uncertainty quality). While waiting to run a workload until renewable energy is available in a data center or running a smaller DNN with lower accuracy on a mobile device may be viable strategies for those domains, they do not transfer directly to AVs due to safety concerns. Research into algorithmic changes such as compact DNN architectures [13] and efficient non-DNN algorithms that modify the algorithm to reduce the computing energy needed [18], [19] are worth exploring to understand the design space for algorithmic efficiency improvements without sacrificing safety. Moreover, Jevon's paradox may manifest when it comes to safety [20]; for example, pruning a DNN so it has half the original latency may result in AV autonomy stacks running it twice as often to increase safety. 5) Encourage an industry standard to release computer power for autonomy stack on AV hardware: Much of the difficulty in assessing the carbon impact of computing onboard AVs is due to a lack of visibility of the workloads and hardware efficiency of current AV companies' autonomy stacks. Ideally, industry would release a set of holistic metrics in order to conduct fair comparisons between different designs [D1]. However, due to concerns about intellectual property, we encourage industry to at least release computer power since that enables the community to assess the carbon impact of the autonomy stack while keeping the system details proprietary.
In this paper, we highlight the potential for significant emissions from computing onboard AVs, comparable to all data centers today. Our framework to estimate emissions is adaptable as the community gains more information or for industry to use based on internal numbers. We hope this work encourages research in several exciting directions unique to AVs that can help better characterize and hopefully, reduce the carbon footprint of computing onboard AVs.
https://github.com/mit-lean/carbon-computing-avs
In this section, the alphanumeric references point to references in TableIand II.