FSD regulation algorithm may introduce large models

Recently, Musk tweeted that vehicle control is the final hurdle in the FSD AI puzzle, which could reduce over 300,000 lines of C++ control code by about 2 orders of magnitude. Our engineers are making good progress so far, but progress on training calculations is limited.

This seems to signal that the FSD's regulation and control algorithms will be introduced to large models to accelerate the landing.

FSD is Tesla's Autopilot software. In Musk's vision, FSD is going to allow vehicles to be fully self-driving without human intervention and be safer than human drivers.

To formally launch FSD or to realize autonomous driving, it is necessary to overcome the technical hurdles of perception, planning and control. Only after the first step of the perception problem is solved, can we step into the planning and control phase of the attack.

Two years ago, Tesla rewrote the code of the perception module, introduced the big model into the field of automatic driving, and proposed the combination of BEV+Transformer. Subsequently, car companies such as Xiaopeng, Ideal, and Azure, as well as many intelligent driving suppliers, have gradually started to use BEV+transformer to solve the problems at the perception level. The "optimal solution" for the first step of autonomous driving has been recognized by most people.

The introduction of large models has brought obvious improvement to the perception module, and the trend of "focusing on perception but not maps" has become more and more obvious, and the data closed-loop system is gradually taking shape.

Now, Musk once again revealed that Tesla also used AI algorithms to streamline the code in the regulation and control algorithm, which means that the regulation and control code is also being rewritten. After rewriting the code, the training of the system itself requires massive data and arithmetic support, which may also be the reason why Musk said "arithmetic is limited".

Traditional regulation and control algorithms are based on a rule-based system, which is heavy code and complex. If Tesla introduces a large model into the regulation and control algorithm, not only will the code be streamlined, but the training of the regulation and control algorithm will also be further accelerated, and it will be closer to human drivers. However, it should be noted that the safety risk of the regulation and control link is higher than that of the perception link, so it also needs to be backed up by the traditional rule system.

However, as of now, Tesla's introduction of large models in the regulation and control module is still in the early stages, and what needs to be solved is still the problem of arithmetic power. Previously, Tim Zaman, general manager of engineering at Tesla, also tweeted that at present, our computing cluster has only 0.3% free time, of which 84% of the work is high-priority, and we hope that we can have more computing space.

This problem may be expected to be solved with the commissioning of Dojo.

It is understood that in July Tesla's own supercomputer Dojo has been officially put into production. Data show that each Dojo is integrated with 120 training modules, built-in 3,000 D1 chips, with more than 1 million training nodes, the arithmetic power of 1.1EFLOP (trillion floating point operations per second).

Musk has been looking forward to this for a long time, and he believes that the use of Dojo will give Tesla an order of magnitude improvement in training costs. According to Tesla's planning, Dojo will drive Tesla to enter a period of rapid growth in arithmetic power. in February 2024 Tesla's arithmetic power scale will enter the world's top five, and in October 2024 Tesla's total arithmetic power scale will reach 100 Exa-Flops.

With Dojo put into use, Tesla's regulation algorithm will accelerate the reconstruction, and FSD landing in the year may become a reality. And how the progress is after that, I believe Musk will make another preview.