Tesla's parallel processing patent can reduce the risk of hardware failure in the fully autonomous driving kit FSD

According to foreign media reports, Tesla recently applied for a patent called "Parallel Processing System Runtime State Reload", which includes a system with three or more processors working together to effectively eliminate possible hardware failures in Tesla's Autopilot or Full Self-Driving (FSD) system. The patent involves a powerful parallel processor system that can keep running when one of the processors fails or runs in an erroneous state. The patent reads: "If one of the parallel processors fails, at least one other processor will be able to continue running the autopilot function."

Tesla's fully autonomous driving (Image source: teslarati.com)

Tesla applied for and published the patent on August 26, local time, just one week after the company held its Artificial Intelligence Day event last Thursday (August 19, local time). The system described in the patent has at least three processors running in parallel, monitored by circuits, and if one of them has an operating state error, it can locate and identify the faulty processor. Then, the circuit can identify and switch to the second processor when the operating error occurs, access the operating state of the second processor, and load the operating state of the second processor into the first processor with the operating state error.

Tesla describes the patent in detail: "A system on a chip (SoC) includes multiple processing systems arranged on a single integrated circuit, and each individual such processing system generally performs a set of processing functions accordingly, and is generally interconnected by one or more communication bus structures, and such communication structures include an N-bit wide data bus (N is an integer greater than 1). Some SoCs are deployed in high-availability systems such as financial processing systems, autonomous driving systems, medical processing systems, and air traffic control systems. Such parallel processing systems generally operate based on the same input data and include substantially identical processing elements such as pipeline structures so that when each parallel processing system operates normally, it can produce substantially identical output results. Therefore, if one of the parallel processors fails, at least one processor can continue to perform the autonomous driving function."

Tesla patent diagram (Image source: teslarati.com)

Technically, a self-driving car only needs one processor to perform the functions described. However, when loaded with neural networks, such processors may be overloaded with data and may experience short-term and non-permanent operating errors. When this happens, the system can switch to one of the other processors for normal operation. The patent includes at least two backup processors, as it repeatedly mentions three backup processors.

After that, the second processor will be activated and load the operating state into the first processor to allow the main processor chip to operate again. The patent reads: "Therefore, in order to overcome the above disadvantages, among other disadvantages, the parallel processing system of the present invention includes at least three processors operating in parallel, a state monitoring circuit, and a state reloading circuit. Among them, the state reloading circuit is coupled to the at least three parallel processors and is configured to select a second processor of the at least three parallel processors to reload the state, access the operating state of the second processor, and load the operating state of the second processor into the first processor."

In addition, it seems that the patent is related to the Dojo supercomputer that Tesla CEO Elon Musk previously described in detail at AI Day. In order to improve accuracy and allow the processors to operate in parallel, the system will use clock inputs to calibrate the two processors.

Tesla has been focused on achieving precise FSD operation and has revised its strategy several times. Earlier this year, Tesla switched to a camera-only approach for both the Model 3 and Model Y, coordinating processing from eight external cameras to achieve more precise FSD operation. The internal processors responsible for compiling, compressing, and sending data to the neural network may temporarily fail to operate, so a backup processor is a good idea in order to continue to understand the autonomous driving data.