The latest self-developed self-driving chip, the core of HW 4.0, which Elon Musk revealed in 2020, has been revealed.
Not surprisingly, the new product is a huge improvement over the current FSD chip, and TSMC foundry.
What people didn’t expect was that after Nvidia’s Jen-Hsun Huang threw out the self-driving chip “king bomb”, Elon Musk changed his original plan and chose to follow up.
Tesla’s new brain
Tesla’s latest self-driving chip has at least completed the design verification work.
The relevant news exposure indicates that TSMC has undertaken a huge amount of Tesla self-driving chip orders.
Judging from the chip production routine process, such news indicates that the new generation FSD chip is likely to have been successfully flowed.
There is other hidden information in this news.
As we all know, the current 14nm process of Tesla FSD chip has been produced by Samsung OEM.
Instead, the latest self-driving chip order fell to TSMC, and Samsung’s name did not appear. Most importantly, TSMC took on Tesla’s latest order, using the 5nm process.
There are two very important points.
First of all, basically 100% sure that this order must be Tesla’s latest self-driving chip. Although Samsung also has 4-5nm process mass production capacity, but the yield rate is lower than TSMC, by Tesla abandoned very reasonable.
Secondly, Tesla’s new self-driving chip’s ability to progress may exceed outside expectations.
Because the 2019 release of HW 3.0, Elon Musk has revealed that the next generation of chips will use 7nm process, and now the situation is that Tesla directly use more advanced processes.
You do not see the old yellow next door just dropped a self-driving nuclear bomb – 2000TOPS DRIVE Thor, at least using the 5nm process.
From this perspective, autonomous driving leader Tesla can not lag behind.
Tesla’s new chip, what level?
Can be divided into horizontal and vertical two dimensions of comparison.
Vertical dimension, the current FSD chip computing power of 144TOPS, using Samsung’s 14nm process technology manufacturing, including three quad-core Cortex-A72 cluster, a total of 12 2.2 GHz CPU, a 1 GHz Mali G71 MP12 GPU, two 2 GHz neural processing unit, and a variety of other hardware gas pedals. FSD supports up to 128-bit LPDDR4-4266 memory.
According to the news, the new self-driving chip will be about three times more powerful than the current self-driving chip.
The performance here may refer to the combined energy/calculus parameters, but it is not excluded that it refers to the single-chip computing power. If so, then the new chip is likely to reach 400-500TOPS.
In addition, the new FSD chip will be optimized for AI calculations in response to the task characteristics of autonomous driving.
In a side-by-side comparison, the latest self-driving chip will still be at the world’s leading level if it starts mass production in 2023 and starts getting on board in large quantities in 2024.
NVIDIA’s 2000TOPS nuke will not begin mass production until 2025 at the earliest. At this stage, Orin’s single chip has 256TOPS of arithmetic power, and for OEMs that require more arithmetic power for autonomous driving, they generally have multiple chips to go with it.
Qualcomm’s latest Snapdragon Ride has been on board the Great Wall Wei brand, with an arithmetic power of 360TOPS, and not long ago Qualcomm released the Snapdragon Ride Flex series, which also has an arithmetic power of 2000TOPS, but the production time has not been disclosed.
So Qualcomm’s move is also interpreted as a forced response after being pressured by the Nvidia nuke, in order to maintain market confidence.
Domestic players, Horizon Journey 5 based on TSMC 16nm process to build, AI computing power can reach 128TOPS. 2023 plans to launch Journey 6, computing power 1000TOPS, but mass production on board time may also be until about 2025.
Beyond Horizon, Huawei is another important player.
The MDC 810, with 400TOPS of computing power, is already in mass production and is equipped with a GPU that does not support general-purpose computing, but instead uses Ascend, an AI chip with a “domain-specific architecture” for computing.
The products of Black Sesame Intelligence and CoreTech are still in the stage of catching up with Nvidia Orin.
Another old player in autonomous driving, Mobileye, is far behind in terms of paper parameters, with only 176TOPS planned for mass production in 2025, and its onboard projects are being snapped up by other upstarts.
So, Tesla’s latest self-driving chip, which will start mass production in 2023, with 400-500 TOPS algorithm power level, will lead the world for at least two more years.
Tesla’s strongest brain, what kind of “soul” to fit
The new brain will undoubtedly help take FSD to a new level of capability.
The latest FSD Beta V11 version has just been released, and there are 8 new features.
"1. High-speed scenes can use FSD Beta. which unifies the vision and planning stack on freeways and non-freeways, replacing the traditional freeway stack that has been used for more than four years. The previous freeway stack relied on a few single cameras and a single frame network that could only handle simple lane specific operations. 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, enabling smoother control and making better decisions. 2. Improved Occupancy Network (OPN) retrieves data from close obstacles and accuracy in bad weather conditions, a 4x increase in transformer spatial resolution, a 20% increase in image feature matcher capacity, calibrated side cameras, and an additional 260,000 video training clips (both real and simulated). 3. Improved merging behavior by leveraging lane shapes 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 keep away from general obstacles such as blocked lanes and road debris, while also making switching between lane drift and vehicle lane change smoother. 5、Improved speed-based lane change decisions to better avoid slowing down traffic in fast lanes and to reduce interference with navigation. 6、Reduced speed-based lane change sensitivity in CHILL mode. 7、Improve lane changing to allow for higher sharpness (plus acceleration) when needed to stay on course or stay out of blocked lanes. 8、Reduce the latency of trajectory optimization by an average of 20% by using numerical tricks for more efficient calculations, while maintaining the existing performance."
In addition to the refinement of individual function optimizations, the most important progress is the expansion of FSD to high-speed scenarios, enabling interfacing with urban road scenarios.
This means that the FSDV11 version being tested now has theoretical autopilot capability from P to P. Elon Musk’s words are “you can reach your destination without touching the vehicle controls”.
So, Tesla’s latest Autopilot chip on board will definitely be able to support a mature production version of the FSD software and truly deliver on Elon Musk’s promise of “fully autonomous driving” made N years ago.
One more thing
Tesla’s own chip development history.
Earlier in 2014, Tesla still used the Mobileye assisted driving chip EyeQ3, with less than 1TOPS of computing power.
But since the 2016 Tesla Model S truck crash incident, Tesla and Mobileye parted ways and turned to the embrace of Nvidia Drive PX 2.
With 24TOPS of computing power, there has been a quantum leap. But even the so-called super car computer Drive PX 2 has not become the “perfect self-driving chip” in the eyes of Elon Musk, there are still three problems: high cost, high power consumption, and arithmetic power can not fully meet the demand.
In 2019, Tesla officially broke up with Nvidia and released its own Hardware 3.0 hardware and said “this is the best chip in the world” with an arithmetic power of 144TOPS.
This is also the current Tesla power computing hardware.
In 2020, there is news that Tesla is working with Broadcom to develop Hardware 4.0 chip, is expected to use TSMC 7nm process, and in the fourth quarter of 2021 full mass production.
In 2021, Tesla confirmed that the self-driving chip will continue to be OEM’d by Samsung, but not for HW4.0.
In 2022, the supply chain rumors that the outsourcing of Tesla’s HW4.0 chip will switch to TSMC, using 4nm / 5nm process to build.