As the marathon for autonomous vehicles charges onward, software is often correctly at the center of the conversation. However, before autonomous vehicles (AVs) can safely operate on public roads, regulators would like to see the average frequency of failures be approximately 1,000× less than that of an average human driver. This will require further development of the underlying technology, which includes not only software model improvements but also improvement and scaling of the hardware that drives increasingly complex software.
My colleague, Cole McCollum, previously covered the performance improvements seen in the AI chip industry. In this blog, I'll summarize some of the approaches to applying those chips in autonomous vehicle applications.
Tesla, Mobileye, and Nvidia are all good examples of the range of strategies for AI chips in the automotive space. Tesla's vertical integration provides the greatest level of customization for its specific solution, though it will be more difficult to reach the economies of scale available to Mobileye and Nvidia. Mobileye has its own AV aspirations but lacks the broad network of Nvidia. Nvidia is able to leverage its existing GPU businesses to enhance its AI systems.
In order to bring its automation capabilities – called Full Self-Driving (FSD) – to market faster, Tesla decided to build its own computer, featuring a custom-designed neural network chip. The FSD computer consists of dual systems on a chip (SoCs) developed internally by Tesla. It runs at 144 TOPS, with a power consumption of 72 W. In 2019, the company claimed that the FSD computer cost 80% of what Tesla had paid for the previous Nvidia-based system (not including the design costs).
While it isn't an auto OEM like Tesla, Mobileye does develop its own ADAS and AV software. (Click here for our analysis of Mobileye's CES 2021 AV strategy). In order to bring these capabilities to its customers, Mobileye has developed its own custom EyeQ SoC devices, which are used by 27 vehicle manufacturers. Mobileye claims that the most recent generation, the EyeQ5, is capable of Level 4 to 5 autonomy. The chip runs at 24 TOPS at a power consumption of 10 W. We can expect for the chip to see use in Q4 2021 when Geely plans to launch Lynk & Co models with Mobileye's SuperVision ADAS product, which uses two EyeQ5 chips. It's currently unknown how much power that 2× chip ECU will draw.
While Nvidia does the least in terms of providing ADAS and AV capabilities on its chips, it has a very broad reach across the industry. Through its hardware and software solutions, Nvidia partners with OEMs, Tier 1s, and both major AV developers (Argo AI, Aurora, Zoox) and smaller ones (Nuro, Momenta, Navya). Nvidia's automation-specific chip, the Drive AGX Xavier (30 TOPS at 30 W), can be paired with its Turing GPUs to support higher levels of computing but with a much higher power draw. Nvidia's AV computer, the Pegasus, runs at 320 TOPS at 500 W. The company is targeting a 2022 generation system called Orin, which will be able to scale across different levels of automation (Levels 2 to 5).
While Mobileye has the best compute-to-power ratio (2.4 TOPS/W) and Nvidia can provide the largest bulk on a chip (30 TOPS), the competition in vehicle automation relies less on these types of hardware specs. Ultimately, the AV competition does not follow who has the best chips. Rather, the chips follow the solutions. Tesla boldly believes that it has the best solution, so it makes its own chip.
Those who trust Mobileye's ADAS and AV solutions are going to use Mobileye's chips. Developers who trust their own software but want to scale across multiple OEMs (Argo AI, Aurora) use partnerships with Nvidia. While AI chip development is an important aspect of bringing automated features to vehicles, it is not the make-or-break part of the solution.
While these three companies are suppliers in the space, other companies in the space are being proactive about chip development as well. In 2020, Toyota and Denso launched a joint venture called Mirise Technologies, which focuses on R&D in semiconductors, including SoCs for autonomy, electrification, connectivity, etc. Others like Nio and Geely are also investigating custom SoCs. Similarly, clients would do well to not consider AV chips in a vacuum but rather as part of ADAS/AV strategy or even a vehicle ECU portfolio.