The impact of autonomous driving on energy consumption
The automotive industry constantly strives to improve safety. Safety has been a key driver to develop driver assistance and - in future - autonomous driving. The main challenge facing a (semi) self-driving car is to have a good perception of the environment. To achieve this, the industry has been vigorously focusing on engineering better sensors and higher-performance computer systems. It is expected that the future will consist of cars with an increased number of data-intensive sensors such as LiDARs and cameras. Processing and managing data on- and off-board will be a key challenge as a large portion of the overall energy consumption will be associated with this. The real-time computing requirements for these data streams will drive energy consumption to unsustainable levels, unless we find better ways to process these vast amounts of sensor data. It is a key realization that quality in sensor data cannot be defined by the sheer number of pixels in a specific sensor design, but by the metrics used to warrant the quality of the data generated in real-time, departing from the traditional definition of number of pixels or quantization per pixel.
Sensor data processing requires a lot of energy
Of all the elements of ADAS and of L4-L5 Autonomous Driving, the sensors and sensor processors are by far the highest contributors of power consumption. Efficient sensor data processing and pre-processing is the key to lower this power consumed. Optimized pre-processing and filtering of sensor data in the edge can lead to a substantial reduction in overall energy consumption. This will lead to a far higher efficiency for connectivity, AD(AS), sensor fusion, object detection, etc. For instance, it will enable to use CPU’s instead of ‘energy-heavy’ GPU’s. Data reduction and pre-processing algorithms compatible with AI or machine learning (ML) tasks show a potential of up to 10-30% savings in battery range for a L4-L5 electric vehicles by requiring lower power hardware that consumes less.
Where can power be saved?
The autonomous driving software stack typically solves different tasks. The calculations associated with it are typically very complex and require power-hungry CPUs and GPUs. This energy is taken either from the battery of an electric car or from the fuel of a current car. The more energy must go towards data processing, the lower a car’s mileage reach becomes. Edge pre-processing software can directly reduce this amount of data processing with factor 10X or more. Such software is available for all the three sensor families that are used in a car: telematic, video and 3D point cloud.
Understanding the data flow could be helpful in the data processing requirements and the effectiveness of the algorithms to save energy. To collect sensor data for remote operations, the car must transmit the data to the cloud. The raw data generated by L4-L5 sensors are too heavy in size, and to have them transmit and receive data to and from the cloud requires energy guzzling 4G/5G standards. Edge techniques such as sensor data fusion helps to couple sensors and to assist compressing their data to ease the load of energy consuming transmissions. The upload/download cycle is also dependent on Sensor accuracy (about >95%) translating to the number of data exchange iterations required.
The additional computing power demanded by Autonomous driving technology, can reduce an electric vehicle’s range by 10-30%. Achieving performance improvements in this domain depends on the right combination of hardware and software. The vehicle safety standards demand real-time reactions as well as a deterministic instead of a “Neural Network black box” approach. Autonomous driving algorithms can work efficiently with pre-processed sensor data. In fact, the computing requirements for a L4-L5 vehicle could be reduced by as much as 10X by focusing processing resources on a smaller but “higher quality” dataset. Quality is defined by the ability to accurately predetermine objects and situations of interest within the raw data. In the end, it is estimated that the off-board processing would just be a fraction of the time compared to the total processing time.
2X - 4X energy reduction through embedded software
One path towards a low energy consumption in a car is an optimal reduction in computing resources by making use of dedicated hardware acceleration. Another - complementary - path is to deploy embedded pre-processing of sensor data at the sensor level. By identifying objects and by reducing raw data one saves power consumption by a factor 2-4x. Optimizing power consumption in such a way at the sensor-level and at the sensor fusion process adds significantly to the power efficiency. This process refines the data collected and subsequently lowers the energy impact of updating of the sensor-driven AI-models in the cloud. These benefits compound together for an overall lower energy consumption in the car.
Latency and accuracy should not suffer
A typical data reduction of 10x only degrades the AI-performance in a negligible way. In parallel the car must deliver immediate driving decisions to steer. Obtaining an almost perfect result in real time in the car is more important than a perfect result coming too late. The edge-processing technology takes about 20ms to process the data. Therefore, pushing for additional - but AI-compatible - data compression is crucial to realize a safe and reliable AD.
We must carefully plan for the on-board processors to be as accurately as possible for the needs of computing without negatively impacting the mileage reach of the car too much. Optimum performance, efficiency and energy consumption of onboard processor demands an efficient pre-processed data at the edge. The edge pre-processed data must be AI compatible with pre-identified regions of importance, segmented and with reduced size to improve the time and power consumption in processing. The data is pre-processing and refinement at the edge level ranges around 20ms. While the high-performance processors work with the pre-processed data to make the real-time driving decisions as quickly as possible.
Difference between low- and high- power hardware
Perception models for AD must be continuously synchronized with cloud-models that are been fed by extensive data collection campaigns. “Going autonomous” demands therefore more power on board. Power that will be taken from battery power for its operation. This is even without considering the 15 to 25 kgs of additional equipment AD brings. A production car of today with advanced safety systems can generate up to 6 gigabytes of data every 30 seconds with cameras and radars. Even more would be generated in a self-driving car with additional LIDARs sensors on board. The generated data needs to be sorted and combined efficiently for an AI-friendly perception of the world while enabling a more granular data collection campaign for ongoing training of models. These model actualization efforts take large computing power and in turn large energy consumption. Teraki’s pre-processing support on minimizing the processing power and help saving energy as a first step, due to overall improved quality further computational requirements related to the inference process of perception models can be further reduced by more than 10x.
To understand the difference in energy consumption, let’s hypothetically construct 2 prototype cars: one with high spec that uses a powerful computer with CPU + GPU units, along with high resolution sensors and another with low spec of the sensors with limited processing power dependent on refined algorithms and marginally lower resolution sensors that are required for autonomous driving.
By listing the required number of sensors, processors and necessary connection equipment along with its power consumption specification available today, we can get an idea of the total required power needed for autonomous driving. The low spec hardware reaches a total power requirement of around 200 watts, while the high spec hardware demands 1160 watts, almost 950 watts more from low spec for its operation.
A traditional car operates of 50,000 kms covered a year, a low spec car using Teraki Software, produces 70kgs of CO2 while the energy consuming high spec car produces 430kgs of CO2 per year for autonomous driving alone.
For every 3 Autonomous vehicle’s, Teraki’s software saves 1 Ton of CO2 emissions per year!
Three cars running autonomously for a year in high spec hardware produce about 1.3 Tons of CO2. With low spec hardware three cars produces only produces 0.2 Tons per year.
Bringing Teraki software to the equation, would therefore save 1 Ton of CO2 for every 3 cars. Each year. This significant saving in CO2 emissions is the result of running embedded Teraki software that enable data-intensive applications to run on low powered hardware. Intelligent pre-processing solution like Teraki’s will have a huge impact on the overall energy consumption of the car. In an industry such as automotive, an efficiency delivered at any level of a car will scale to massive results when in mass production. Teraki has developed embedded AI-edge processing software that improves energy efficiency while warranting less degradation for the training of machine learning models than standard technology, enabling low powered hardware to do the job.
This results in a significant lower carbon footprint in the world. 1 ton of CO2 per year for any 3 cars. Making the world a bit greener. With software only.