As the software required to make a vehicle safely drive itself becomes ever more complex, it takes significantly more computing power to process the data pouring out of sensors and make decisions. Until recently, Nvidia and Intel (including its Mobileye subsidiary) have been the primary players in this competition to power automated vehicles (AVs). But as the industry inches closer to deployment of automated driving, other vendors are stepping up to try to grab a slice of this potentially lucrative market segment. The latest to step into the high-performance automotive compute market is Qualcomm, and GM will be deploying the new chips from 2023.
At CES 2020, Qualcomm, which is best known for its smartphone chips is announcing its SnapDragon Ride platform as a direct challenger to Nvidia’s recently announced Orin chip. At the 2019 CES, Qualcomm demonstrated its first iteration of an AV compute platform in a Lincoln MKZ hybrid sedan. Using a cluster of six SnapDragon 820A chips, the Qualcomm test car performed on-ramp to off-ramp partially automated driving similar to what Tesla currently offers with AutoPilot.
However, compared to the systems-on-a-chip (SoC) offered by Nvidia or the field programmable gate array (FPGA) chips from companies like Xilinx, the 820A is limited in its performance. The 820A is based on the 820 smartphone SoC that debuted in 2015. In mid-2019, Qualcomm switched its development platform over to the newer SnapDragon 8155, but CES brings the new Snapdragon Ride platform. However, the 820A is finding a place in vehicles starting with the Land Rover Defender where it will power the digital cockpit system and feature dual LTE modems to enable simultaneous media streaming and OTA updates.
The big news is the Ride platform which actually consists of two separate chips. The starting point is ADAS application processor. This is the primary SoC that the whole platform is built on. As with Snapdragons for mobile devices, the ADAS processor is based around a ARM CPU architecture combined with numerous other components to handle data input/output and optimized neural network accelerator cores to handle the artificial intelligence software that now used for much of the perception part of the AV system. This is the part of the stack that tries to make sense of the sensor data and recognize the objects around the vehicle.
The second chip is the Autonomous Driving Accelerator that essentially just adds more of these same neural network cores to handle the required workload. Qualcomm is promoting 3 basic configurations. A single ADAS processor is more than adequate to handle all of the current and upcoming advanced driver assist systems. This includes features such as adaptive cruise control, blindspot monitoring, forward collision avoidance, pedestrian detection and more. These are the features that are quickly becoming standard equipment on almost all cars and will soon be required in order to get top scores on new car assessment programs.
Adding a second ADAS processor provides the compute performance to handle so-called level 2 and level 3 partially automated systems. These L2 systems include Tesla AutoPilot and Cadillac’s Super Cruise that combine lane centering, adaptive cruise and features like auto lane change or in the case of the Cadillac, hands-off driving. L3 systems allow the driver to look away from the road, but they must still be ready to take over under certain conditions. Using two of the ADAS processors provides the necessary computing power and redundancy in the event that one fails.
According to Qualcomm, adding the accelerator chip takes the platform to a whole new level of performance. While Qualcomm isn’t providing all the performance details yet, the company claims that a three chip setup with two ADAS processors and the AD accelerator is capable of delivering 400 trillion operations per second (TOPS). For comparison, Nvidia’s current four chip Drive Pegasus system can deliver 320 TOPS, the new Orin chip can do 200 TOPS and the Xavier that will hit production applications in 2020 can deliver 30 TOPS.
Almost as important for AV applications however, is power efficiency. The Xavier delivers its 30 TOPS while consuming about 30W of power, While the new Orin will consume about 65-70W for its 200 TOPS. Qualcomm is claiming 33-50% better efficiency with about the same 60-70W for 400 TOPS. Where more performance or added redundancy is needed, developers can just add more chips, scaling up to 700 TOPS and beyond with as little as 130W power consumption.
It is also claimed that this will be delivered with air-cooling. Many of the compute platforms in development for AVs are relying on liquid cooling to keep temperatures under control. Tesla’s V3 “Full Self-Driving” Computer is claimed to deliver 144 TOPS from its two in-house developed SoCs but it is liquid cooled. That adds complexity, cost, and a potential failure mode.
Like Nvidia and Intel, Qualcomm is also offering its customers a full software stack for automated driving. This includes perception, localization, and path planning. The full system also incorporates Blackberry’s QNX operating system and Synopsys middleware layers.
Qualcomm is already sampling the accelerator chip and will start sampling the ADAS processor by mid-2020. Both chips should be in full production by 2023 with General Motors already planning to use Snapdragon Ride for more advanced next generation ADAS at that time. If the platform lives up to expectations, it could also find a home in some of the first second-generation AVs.
