Tesla has announced the launch of its self-developed Full Self-Driving (FSD) chip, which is designed with dual processor chips and manufactured on Samsung's 14nm FinFET process. At the same time as the release of the new self-driving chip, Tesla also compared its products with NVIDIA, emphasizing its own design of the chip, compared to the computing power of NVIDIA products has made significant progress; in this regard, NVIDIA recently responded, thinking Tesla's comparison is not "accurate" and emphasizes that NVIDIA's self-driving platform not only has high-performance AI, but also has "openness."
Rob Csongor, vice president of NVIDIA Autonomous Machines, said that it is undeniable that Tesla is raising the production standards for self-driving cars. First, Tesla uses dual AI processing chips, each containing CPU, GPU, and Deep-learning Accelerators, with a computing power of 144 TOPS, and can collect data from a variety of surround camera, radar, ultrasonic, and deep neural network algorithms. . In addition, Tesla also stressed that they will continue to develop the next generation of self-driving chips to achieve higher calculation performance than 144 TOPS.
Csongor pointed out that Tesla continues to reaffirm that autonomous vehicles are the key to improving driving safety, efficiency and convenience, as well as the future of the automotive industry, and therefore require extraordinary computing performance. This is the same as NVIDIA's point of view, which is why we designed and built the NVIDIA Xavier SoC. Xavier features programmable CPUs, GPUs and deep learning accelerators with data processing speeds of 30 TOPS.
At the same time, in response to the high computing needs of self-driving, NVIDIA also has a two-chip solution "DRIVE AGX Pegasus", which uses Xavier with GPU to increase the performance of single-processor computing to 160 TOPS; and integrates two processors into the same platform, data Processing efficiency is as high as 320 TOPS.
Csongor believes that high-speed computing is the key to developing autonomous vehicles. Both Tesla and NVIDIA are moving in this direction; however, Tesla is inaccurate in comparing its FSD self-driving chips with Xavier. Because the FSD self-driving chip is a two-chip solution, and Xavier is a single chip, the benchmarks are different. To compare, the DRIVE AGX Pegasus should be compared to the DRIVE AGX Pegasus, and the 320 TOPS performance of the DRIVE AGX Pegasus is much higher than the 144 TOPS of the Tesla FSD self-driving chip.
Csongor further stated that, in addition, the NVIDIA Xavier SoC is mainly used as Level 2 (or Level 2+) Autopilot assist system (ADAS), rather than full autopilot, because autopilot requires more powerful computing performance. At the same time, Tesla also misplaced the Xavier SoC data in the briefing. Xavier can provide 30 TOPS calculations, but Tesla's data shows 21 TOPS.
Csongor stressed that it is indisputable that Tesla develops high-performance chips and improves the level of self-driving computing. This is the future direction. However, in addition to high computing performance, NVIDIA's platform is more "open" for the industry, allowing automakers to provide high-performance self-driving cars; Tesla processors are currently only used by themselves.