Is There a New Super Cycle in the Making for Nvidia, Courtesy Of Tesla?

Author(s): Purankhoeval

Originally published on Towards AI.

In the whirlwind of technological evolution, AI stands as a beacon of imagination and Artificial General Intelligence (AGI) is not an elusive dream anymore. While AGI has a long way to go on the path to finding what seems possible in the near future is autonomous driving of vehicles, and AI has a big role to play in solving this equation. There are few prominent players out there who are actively working on the solution and Tesla is a shining star amongst them and is very unique in it’s approach. Let’s unpack this to understand the nitty-gritty of Tesla’s approach and progress made so far and how it will give yet another huge lift-off to NVIDIA's compute cycle.

Tesla’s approach to solve the autonomous driving

Tesla relies on computer vision technology to solve this problem. Tesla vehicles are equipped with 8 external cameras that capture information about the surrounding environment, including other vehicles, pedestrians, objects, traffic lights, signs, and lanes.

Almost all other players aiming to solve the car autonomy problem are relying on the LIDAR technology. Lidar uses laser pulses to create a 3D map of the environment, providing precise distance and object detection. It works well in various lighting conditions and has a longer detection range than cameras. With all the technical promise, it comes at a tremendous cost and, in principle, would need models that can act on LIDAR data for the inference to work, and the law of averages would not be on their side. This gives Tesla the unique value proposition of solving the autonomy using computer vision and end to end deep learning model. Tesla recently launched FSD (Full self driving) version 12, we will unpack how Tesla had a major pivot in it’s approach of imitation model in the version 12 but for now let’s understand what Tesla does with it’s 8 cameras.

Tesla captures the video data from the cars that have chosen to participate in the program. Some drivers in the Reddit community who are part of the program have shared data, and it turns out the daily upload is somewhere between 5–7 GB and monthly well over 100 GB per user. Tesla has touted earlier that they have collected over 1 billion miles of driving data. Dojo is the supercomputer built by Tesla, and it has been trained on millions of miles of data already. Going forward with the rollout of FSD 12 Tesla claims to collect 1 billion miles per month 🤯. Tesla is very strategic in driving this push for the FSD and have rolled out FSD free for the first 30 days and have adjusted the price of FSD subscription very competitively at $99/month, this pricing is genius to accelerate the mass adoption of FSD. (**Consumers are less likely to subscribe to anything over $100)

Now let’s understand what changed between FSD version 11 vs version 12

FSD 12 : Deterministic model to a non-deterministic, end-to-end imitation learning model

1. It was almost impossible to code for all the edge cases in a deterministic way per se and achieve self-driving using classification and a label-based code approach. Tesla has pivoted from this approach to a much less complex strategy and kept it simple to imitate learning through the best drivers. Video clips of the 5-star driver as an input to the model enable the model to learn from the best behavior and situation handling. This reminds me of my driving lessons from a Arkansas high school teacher, he had noticed that I was applying some bad techniques of crossing my hands while turning the steering wheel and he said I have started to imitate based on the wrong techniques I have subconsciously observed from the people around me. He corrected me at very early stage of learning process so I was able to correct it, this emphasizes the quality of training data and draws a parallel to how we humans learn in the real world. The model training should be no different, and thus, Tesla’s supercomputer is aptly named “Dojo” 😎.
2. FSD Beta v12 is a major overhaul of Tesla’s autonomous driving software, featuring a “near-complete rewrite” of the codebase 1.
3. The new version switches the software from controlling the vehicle based on 300,000 lines of human-written code to a system that relies completely on the vehicle’s neural network, trained on millions of real-time video clips 1.
4. FSD Beta v12 includes a major update to the vehicle control functions, transitioning to a “network-path-based” approach, which Elon Musk claims will ensure the car “will never get into a collision if you turn this thing on, even in unstructured environments”.

Earlier this week two democratic Senators have called on the National Highway Traffic Safety Administration (NHTSA) to prohibit Tesla from enabling its Autopilot driver assistance system on roads where it was not intended to be used. With a more prudent approach, the regulators in the US and Europe are going to be laggards, and other markets will drive the adoption first, such as Dubai, Singapore, Mexico, Latin America, and China. This was confirmed earlier this week when a surprise visit of Elon Musk to China secured the regulatory approval of FSD in China, this is a watershed moment and clearly paves the way for Tesla FSD in a very giant auto market. Refer to the table below for some numbers to emphasize the size of the country-wise market share.

Now let’s take a deeper look into what is enabling these models training, the underlying infrastructure is provided by Nvidia. Tesla has bought around 30,000 Nvidia H100 GPUs in total and they are at around 40,000 H100 equivalents for their AI compute capacity.

Traditional auto players are on the sidelines and watching the tech giants to solve this problem and once it is evident which approach works and the economics around it, the race to autonomy is going to heat up in no time. Traditional automakers are not structured with the Software capabilities, and they can not change the economic moat and operations of the company overnight, so sooner or later, software companies like Tesla, Nvidia, and Google, and emerging start-ups will play a part in the vertically integrated supply chain of all the automakers.

Since the inception of GPU compute technology Nvidia has gone through several super cycles. It began with gaming, followed by Bitcoin mining, and the recent massive enterprise boost of Gen AI. The open source community’s commitment to improving the model inference capability at a reduced size will drive more adoption of the models into production-grade solutions and will allow the pairing of multiple models to design more complete AI systems. Inference is where the major cost is going to be in the future and, consequently, the monetization pillar for many tech startups. Edge devices would demand high-performance on-premise inference to make split-second decisions, and as it becomes more accessible, it will drive the growth in the space of edge AI and on-premise inference.

The success of Computer vision based model in achieving autonomy will inspire so many other industry use cases throughout the manufacturing and operation life cycle. Nvidia is primed to be the beneficiary of this monumental shift and it’s evident in their vision with offerings such as Omniverse and digital twins of Manufacturing facilities.

This is my analysis and opinion piece, it should not be used to make any investment decisions. Do your own due diligence and ask yourself what Bobby Axelrod would ask “ What’s your level of certainty ?“

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI