More nuggets on BYD ADAS

When BYD was making design choices for its DiPilot-100 platform for the mass market. It had choices of J3/J5 combo, J6M and Orin N. It decided to not go with J3/J5 option, because J5 architecture did not provide enough support for large model technological architecture like transformer. So even though J5 provided 128 TOPS in computation and J6E provided 80 TOPS, it ended up choosing 84 TOPS Orin N and 128 TOPS J6M. This tells me the top of the line TOPS computation number does not tell the full capabilities of ADAS inference chip. It is likely that due to CUDA’s integration with PyTorch library, you can get more out of the full computation power of Nvidia than other chips. There are also other factors like how well the chip support various models you want to run on it. While J3/J5 is said to be enough to support Highway NOA, BYD has larger plans for DiPilot-100. It wants to expand these computation platform to support for more ADAS scenarios including City NOA.

According to BYD, its ADAS division received 72 million km of training data every day. Every 7 days, it could train a new cloud version for OTA into customer end point. As such, it can continually train and improve DiPilot-100 ADAS. At current time, it is likely DiPilot-100 still trails the current Tesla FSD by a year or two. But as DiPilot-100 is equipped with better hardware than FSD, The all important frontal view is supported by 2 120 degree wide angle camera and 1 30 degree long view camera. All supported by 8 Megapixel CIS chip. This allows for 350m detection range. It also comes with 5 MMW Radar, with frontal radar having 300 m range. This hardware configuration is uniform across the entire lineup. That means a 78.8k RMB Seagull has better sensors than most 200k+ RMB EV. China has some of the most busy road system globally. It provides tremendously useful road/driving data for AI development. BYD has China’s largest auto cloud and driving data coming in from a consistent hardware configuration across most of its fleet. It also has the largest engineering team globally with 5000+ ADAS engineers. So even though BYD started this process later than most of its NEV competitors, it is moving the fastest. If we consider that it only announced 100B RMB investment into AI a year ago, their current progress is nothing short of extraordinary.

It has managed to do this while keeping the cost down, which could come down even further once it starts to use self developed ADAS inference chip and scales up production of ADAS domain controller/computation platform. According to 3rd party analysis, DiPilot-100 cost currently is up to 3000 for domain controller. If we add in the 100 RMB average price per camera & radar, BYD can keep the hardware cost to 5000 RMB. Although based on my friend Taylor Ogan’s tweet, the cost is actually going to be even lower than 3000 RMB! Vertical integration + production scaling + lower supplier cost from large orders really make a big difference.

Vertical integration and production scaling are key to BYD cost advantage. Of all the suppliers, chips may represent the largest % of cost that BYD still sends to suppliers, especially expensive Western suppliers. It has invested a lot of money in new fabs and its semiconductor design team. Over the next couple of years, BYD will continue to expand investment into its in house chip design and production to compete against outsider suppliers. Over time, more and more chips for ADAS and other part of the EVs will be produced internally. All of this allows BYD to add more and more functionalities without increasing the cost. BYD is all about delivering more for less.

In the long term, BYD’s goal is to eventually support City NOA for the entire lineup, including the Seagull. I think the leadership already has a vision of how to get there through lowering cost of Lidar to under 1000 RMB and designing more powerful ADAS chips internally. I think we are probably just 1 to 2 years from that point. As such, BYD has a path to complete deployment of smart cars to everyone. That is going to be a huge problem for its competitors.