Tesla's most powerful self-driving chip exposed

Tesla's self-driving next-generation brain. The latest self-developed self-driving chip, which is the core of HW 4.0 revealed by Musk in 2020, has the latest progress. Unsurprisingly, the new product has a huge improvement over the current FSD chip, and it is manufactured by TSMC. Unexpectedly, after Nvidia Huang Renxun threw out the self-driving chip King Zombie, Musk changed his original plan and chose to follow up.

Tesla's latest self-driving chip has at least completed design verification. Relevant news exposure shows that TSMC has undertaken a huge order for Tesla's self-driving chips. Judging from the conventional process of chip production, such news shows that the new generation of FSD chips is likely to be successfully taped out. There are other hidden messages in this message. As we all know, the current 14nm process Tesla FSD chip has been produced by Samsung.

However, the latest self-driving chip order fell to TSMC, and Samsung's name did not appear. Most importantly, the latest order from Tesla undertaken by TSMC uses a 5nm process. There are two important points. First of all, it is basically 100% sure that these orders must be Tesla's latest self-driving chip. Because although Samsung also has a mass production capacity of 4-5nm process, the yield rate is lower than that of TSMC, so it is reasonable to be abandoned by Tesla. Second, the advancement of Tesla's new self-driving chip capabilities may exceed outside expectations. Because when HW 3.0 was released in 2019, Musk once revealed that the next generation of chips will use a 7nm process, but now the situation is that Tesla directly adopts a more advanced process. why?

Don’t you see that the old Huang next door just dropped the self-driving nuclear bomb the 2000TOPS DRIVE Thor, which uses at least a 5nm process. From any point of view, Tesla, the leader in autonomous driving, cannot lag behind. It can be divided into horizontal and vertical two-dimensional comparisons.

Vertically, at this stage, the FSD chip has a computing power of 144TOPS and is manufactured using Samsung’s 14nm process technology. It includes 3 quad-core Cortex-A72 clusters, a total of 12 2.2 GHz CPUs, a 1 GHz Mali G71 MP12 GPU, and 2 2 GHz Neural Processing Units, and various other hardware accelerators. FSD supports up to 128-bit LPDDR4-4266 memory.

The performance of the new self-driving chip will be about three times that of the current self-driving chip. The performance here may refer to the comprehensive energy consumption/computing power parameters, but it does not rule out the single-chip computing power. If so then the new chip is likely to reach 400-500TOPS.

In addition, for the task characteristics of autonomous driving, the new FSD chip will also be optimized for AI computing. In horizontal comparison, if the latest self-driving chips start mass production in 2023 and start mass production in 2024, they will still be at the leading level in the world. Nvidia's 2000TOPS nuclear bomb will not start mass production until 2025 at the earliest. At this stage, Orin has a single-chip computing power of 256TOPS. For OEMs with high computing power requirements for autonomous driving, generally use multiple chips.

Qualcomm's latest Snapdragon Ride has been launched on the Great Wall Wei brand, with a computing power of 360TOPS. Not long ago, Qualcomm released the Snapdragon Ride Flex series with a computing power of up to 2000TOPS, but the mass production time was not disclosed. Therefore, Qualcomm's move was also interpreted as a forced response to the pressure of Nvidia's nuclear bomb, with the purpose of maintaining market confidence.

For domestic players, Horizon Journey 5 is based on TSMC’s 16nm process, and its AI computing power can reach 128TOPS. In 2023, it is planned to launch Journey 6, with a computing power of 1000TOPS, but the mass production time may also be around 2025. Beyond the horizon, Huawei is another important player.

MDC 810, with a computing power of 400TOPS, has been mass-produced. The MDC 810 is not equipped with a GPU that does not support general-purpose computing but uses the "domain-specific architecture" AI chip Ascend to take care of computing. The products of Black Sesame Smart and Xinchi Technology are still in the stage of catching up with Nvidia Orin.

Another veteran player in autonomous driving, Mobileye, is far behind in terms of paper parameters. The mass-produced products in 2025 are only planned to reach 176TOPS. The car project was also robbed by other rising stars.

Therefore, Tesla's latest self-driving chip, which will start mass production in 2023, will have a computing power of 400-500TOPS, and it will lead the world in at least two years.

The new brain will undoubtedly help FSD to a higher level. The latest FSD Beta V11 version has just been released, and there are mainly 8 new features:

  1. FSD Beta can be used in high-speed scenarios. Unified vision and planning stacks on highways and off-highways, replacing the traditional highway stack that was used for more than four years. Previous highway stacks rely on several single-camera and single-frame networks and can only handle simple lane-specific manipulations. The new FSD Beta multi-camera video network and next-generation planner allow for more complex agent interactions while reducing lane dependencies, making way for adding more intelligent behavior for smoother control and better decision-making.
  2. Improved the occupancy network (Occupancy Network) to recall the data of close-range obstacles and the accuracy under severe weather conditions, the spatial resolution of the transformer has been increased by 4 times, the capacity of the image feature matcher has been increased by 20%, and the side camera has been calibrated, and an additional 260,000 video training clips (real and simulated).
  3. Improve vehicle merging behavior by utilizing lane shape and lane boundaries, correlation with approximate map information, and better gap selection algorithms to provide a smoother and safer experience.
  4. Added highway behavior to move away from common obstacles like blocked lanes and road debris, while also making transitions between lane departures and vehicle lane changes smoother.
  5. Improved speed-based lane change decisions to better avoid slowing down traffic in the fast lane and reduce disruption to navigation.
  6. Reduce the speed-based lane change sensitivity in CHILL mode.
  7. Improved lane changes for higher jerk (jerk) maneuvers when required to maintain course or move away from blocked lanes.
  8. By using numerical techniques for more efficient calculations, the delay of trajectory optimization is reduced by an average of 20% on the basis of ensuring the existing performance.

In addition to fine-grained optimization of individual functions, the most important progress is to expand FSD to high-speed scenarios and realize the connection with urban road scenarios. This also shows that the version of FSDV11 currently being tested theoretically already has the ability to automatically drive from P to P. Musk's original words as you can reach your destination without touching the vehicle controls. Therefore, Tesla's latest self-driving chip will definitely be able to support the mature mass-produced version of the FSD software, and truly fulfill the "full self-driving" promised by Musk N years ago.

The whole story of Tesla's self-developed chip 

Earlier in 2014, Tesla still used the Mobileye-assisted driving chip EyeQ3, with a computing power of less than 1TOPS. But since the Tesla Model S crashed into a truck in 2016, Tesla parted ways with Mobileye and turned to Nvidia Drive PX 2. With a computing power of 24TOPS, there is a leap. But even the Drive PX 2, known as a supercar computer, failed to become the "perfect self-driving chip" in Musk's eyes. There are still three problems: high cost, high power consumption, and computing power that cannot fully meet the demand.

In 2019, Tesla officially broke up with Nvidia, released the self-developed Hardware 3.0 hardware, and said that this is the best chip in the world, with a computing power of 144TOPS. This is also the current Tesla power computing hardware.

In 2020, it was reported that Tesla is cooperating with Broadcom to develop the Hardware 4.0 chip, which is expected to adopt TSMC's 7nm process and be fully mass-produced in the fourth quarter of 2021.

In 2021, Tesla confirmed that the self-driving chip will continue to be manufactured by Samsung, but not HW4.0. In 2022, it is reported in the supply chain that the outsourcing of Tesla's HW4.0 chip will be transferred to TSMC, which will be built with a 4nm / 5nm process.

Post a Comment