Tesla launches DOJO chip of artificial intelligence training computer

On August 20th, Tesla officially launched his artificial intelligence training computer DOJO chip.

It is understood that Musk first mentioned DOJO, and called it Tesla's major project in 2019; on August 15 last year, Musk issued a recruitment notice, planning to take one year to develop and launch a neural network training computer. In February, Musk said again that a large number of programmers and engineers were needed to join Tesla's team to do vehicle control and other vehicle-related work.

According to Musk, DOJO chips can perform trillions of floating-point operations per second (1000 petaFLOPS).

As we all know, Tesla has always adhered to the visual scheme in the field of autopilot, and the visual scheme is highly dependent on strong computing power. From 2014 to 2016, Tesla first selected Mobileye's EyeQ3 chip and developed an AutoPilot HW1.0 autopilot platform based on this chip. From 2016 to 2019, Tesla abandoned Mobileye's chip scheme to use Nvidia's DRIVE PX 2 AI computing platform, and upgraded the AutoPilot HW platform to 2.0,2.5 based on this.

While choosing the Nvidia chip scheme, Tesla began to develop his own autopilot chip in 2017 and launched an AutoPilot HW3.0 computing platform based on the first Tesla FSD chip in April 2019. According to public data, the chip has 6 billion transistors, a computing power of 36 TOPS, and a video stream processing speed of 2100fps per second.

In order to solve the problem of self-driving performance, on the one hand, Tesla constantly updates the software version of the platform and launched the new version of FSD Beta V9.1 in July; on the other hand, he is also planning a new generation of FSD chips. It is understood that Tesla is working with Broadcom to develop a new autopilot chip, and with the new HW 4.0 computing platform, is expected to mass production in the fourth quarter of 2022, the new generation of chips will use 7nm technology. It is estimated that the computing power of HW4.0 is expected to reach more than 432 TOPS.