Navigating Autonomous Roads
Amidst a bustling wave of optimism surrounding AI with frequent product launches, new billion-dollar investments and startup activities, the autonomous driving industry, which once bathed in similar excitement is now navigating through a series of unfortunate events. Very recently, Cruise has had its self-driving permit revoked in San Francisco after an incident where its robotaxi was involved in dragging a person. Additionally, it lost its CEO and faced funding cuts from GM, which plans to reduce its investments significantly. A number of Lidar companies have either experienced a steep decline in their share prices or have serious cash flow problems. Waymo’s valuation has plummeted to less than 30% of its previous peak of over $170 billion. Ford and Volkswagen-backed Argo AI ceased operations all together.
Two weeks ago, Tesla quietly rolled out its FSD v12 update to employees. To understand the significance of this step toward achieving full self-driving capabilities, for both Tesla and the entire autonomous driving industry, it is instrumental to revisit Tesla’s history and efforts in developing its own ADAS capabilities.
In 2014, Tesla launched Autopilot 1.0, relying on the Mobileye EyeQ3 chip. However, the opaque black-box solution from Mobileye was deemed unfriendly for ambitious auto OEMs like Tesla. What further pushed Tesla to abandon Mobileye was the first ever fatality caused by ADAS in human history. In 2016, a Tesla Model S driver tragically died when the vehicle mistakenly recognized a white truck against a bright sky as a road sign and the Automatic Emergency Braking system did not activate and respond. This incident exposed the limitations of the Mobileye EyeQ3’s computing power. Knowing that the next generation of Mobileye’s chips being two years away, Tesla made the decision to switch to Nvidia while at the same time doubling down on designing its own ADAS chip.
Nvidia entered the ADAS field in 2015 with its initial Drive Platform, utilizing the Tegra X1 chip. Just 5 months after the Florida accident, Tesla unveiled its second-generation hardware platform, featuring 8 cameras and 13 radars and powered by Nvidia’s Tegra X2 chip. Nvidia enabled Tesla’s Autopilot 2.0 to deliver 10 TOPS, 40 times of that of Autopilot 1.0 enabled by Mobileye. Moreover, the openness of Nvidia’s Drive platform empowered Tesla to develop its own algorithms.
Despite Nvidia’s industry leading performance in hardware, Tesla’s commitment to developing its chips is non-stop and evident through the recruitment of prominent figures to its autonomous driving team, including Jim Keller (architect of AMD microarchitecture), Andrej Karpath (founding member of OpenAI), Chris Lattner (founder of SWIFT) and many others. Harnessing its talented workforce, Tesla dedicated efforts to all four crucial pillars of ADAS: data, algorithms, chips, and computing platform.
In 2019, Tesla unveiled HW3.0, featuring a groundbreaking 14nm Full Self Driving (FSD) chip with a highly customized CPU+NPU+GPU architecture, achieving an impressive 144 TOPS where the GPU only contributes less than 1 TOPS and plays only a supporting role. By 2021, Tesla showcased its utilization of Neural Network Transformers for FSD, employing the same Transformer model that underpins the development of OpenAI. The computing platform, Dojo, served as a neural network training supercomputer on the cloud.
Remarkably, Tesla achieved significant advancements in all four critical pillars of autonomous driving, a feat never matched by any automotive OEM, tier-1 or any technology giant. As of 2021, more than 400,000 drivers in the US actively contributed real-time data to Tesla’s FSD. Despite Elon Musk’s penchant for exaggeration, his claim that “even some of the best AI software engineers in the world don’t realize how advanced Tesla AI has become” holds some truth. Tesla’s AI power is not fully released only because of the industry in which it is deployed — the automotive and autonomous driving industry. The automotive industry in its core is still manufacturing industry that places quality and safety at the most important place. Heightened scrutiny from regulatory bodies and the public underscores any ADAS-related accidents, despite human driving being statistically more perilous.
Nevertheless, if we learn anything from the OpenAI story, it is the non-linearity of software and AI development. A reevaluation of the autonomous driving industry is warranted today. The industry has experienced hype akin to today’s GenAI, followed by disappointment and the collapse of major companies. What does the future hold?
Let’s refocus our attention on Tesla’s FSD 12 upgrade. This update signifies a significant shift towards neural networks assuming control of vehicle functions, moving away from hardcoded engineering controls of rigid “if else”. Over 300,000 lines of code previously governing FSD functions that controlled the vehicle are eliminated. FSD 12 now uses neutral networks to control steering, acceleration, and braking, some of the most important functions, for the first time. Previously, the neutral network was only used to detect objects and determine their attributes. This is indeed a leap in technology. Musk calls it “end-to-end” AI employing a “photons in, controls out” approach akin to human optical processing.
As we approach this pivotal moment in technology, our confidence grows in the increased consumer adoption propelled by decreasing hardware costs and advancing software capabilities. On the regulatory front, over 40 state governments, along with the District of Columbia, have enacted policies or legislation to facilitate the use of autonomous vehicles on public roads, including commercial service offerings.
In the ever-evolving landscape of autonomous driving, the recent tribulations faced by various players highlight the challenging journey toward achieving widespread adoption. But positive shifts are occurring, whether discussing technological advancements, regulatory developments, or increased willingness for consumer adoption. The road ahead, while challenging, holds the promise of safer, more efficient, and revolutionary transportation.
