The confrontation of Tesla’s “FSD Smart Driver Support” with Chinese solutions comes out devastatingly for the latter. “The world’s most advanced systems” have fundamental problems with handling cars, meanwhile Tesla’s FSD manages even in difficult situations. How is this possible? The key to the solution is supposed to be the computing power of the chips Tesla uses and the level of sophistication of the algorithms.
FSD 1.5 years ahead of China in software, with more computing power
A year ago, in March 2024, Tesla began distributing FSD/Autopilot end-to-end, version 12 (v12). End-to-end (E2E) means that artificial intelligence is applied at every possible stage of training, from collecting data from the world (cameras in the cars -> FSD computer), to labeling it (Tesla’s computing center), to software training (also computing center), to decisions made on the road (back: the car). A human makes sure the process doesn’t get derailed, but all the learning happens at a gigantic speed, because it’s within specialized processors – you can see them here:
Tesla’s Dojo Processor (c) Tesla
The Exapod, a set of servers for teaching artificial driving intelligence (c) Tesla
A Chinese think-tank assessed that this alone gives Tesla an advantage, as end-to-end solutions were not planned for use in China until 2025 at the earliest (source). The organization estimated that in the first quarter of 2023, Tesla had computing power of 35 EFLOPS, 35 x 1018 floating-point operations per second. It planned to expand to 100 EFLOPS by the end of 2024. For comparison: Huawei had 3.3 EFLOPS at its disposal, Baidu 2.2 EFLOPS, Geely 0.81 EFLOPS and Xpeng 0.6 EFLOPS. The artificial intelligence in Tesla’s computing center was able to function (=analyze + learn) an order of magnitude faster than its Chinese competitors.
In addition, Tesla is not subject to export restrictions on the latest GPUs, so it is free to expand its supercomputers. The Chinese must rely on those solutions they have access to (up to 30 billion transistors), or work on their own.
Dojo processor cores used to exercise artificial intelligence (c) Tesla
It doesn’t stop there: in April 2024 Tesla sold 6 million cars, and all of them collected data and learned driving from their human owners. FSD alone drove 2 billion kilometers then, while Xpeng boasted 700 million kilometers and Huawei 220 million kilometers. A larger batch plus more computing power on the back end yields better skills on the output. The way FSD is organized at the algorithm level puts Tesla a year and a half ahead of Chinese solutions.
The disadvantage of the solution created by Musk’s company is supposed to be that it has trained for the US market. Its adaptation to other roads (people on the streets etc.) is expected to take 1-1.5 years. After that time, FSD is expected to become a commercializable product, the think-tank concludes.
Editor’s note: the opening photo is from a video in which the FSD pulls over to let a truck pass.
FSD encountered a large truck on a narrow road and, as if it had its own brain, made the exact same decision as a human driver would. pic.twitter.com/QhgQYvZiqA
— 𝐶ℎ𝑎𝑟𝑙𝑒𝑠 (@CharlesWBoy) March 9, 2025
