When you click on links to various merchants on this site and make a purchase, this can result in this site earning a commission. Affiliate programs and affiliations include, but are not limited to, the eBay Partner Network.
how can some of the chinese ev makers have fsd type systems that seem way ahead of tesla?
Good question. FSD is now operating in China but it's not as good at least from what I've seen. But consider, if you took one of those Chinese systems and put it in say Los Angeles how would it do? Very badly I suspect. I would also be careful in evaluating full self driving in China you are not going to see the bad. With Tesla literally anyone can go out and make a video.
Another aspect is regulation, in China there is much more tolerance for risk and failure vs. in the United States.
Good question. FSD is now operating in China but it's not as good at least from what I've seen. But consider, if you took one of those Chinese systems and put it in say Los Angeles how would it do? Very badly I suspect. I would also be careful in evaluating full self driving in China you are not going to see the bad. With Tesla literally anyone can go out and make a video.
this i posted in another thread is quite impressive.
All Li Auto vehicles come standard with the electric automaker’s advanced driver assistance system, Li AD Max. To achieve surround perception, the system uses one forward-facing lidar, 11 cameras, one radar and 12 ultrasonic sensors, as well as DRIVE Orin SoCs.
Can you find a drive that isn't so good? Or are they all perfect meaning they've completely solved self driving.
why the snark? because it might be (gasp) better in some ways than tesla? i certainly never claimed it was perfect or completely solved, but did you actually watch the video? all i said is it's impressive in that horrendous traffic. i haven't sought out problematic drive videos on that car or system, but feel free.
why the snark? because it might be (gasp) better in some ways than tesla? i certainly never claimed it was perfect or completely solved, but did you actually watch the video? all i said is it's impressive in that horrendous traffic. i haven't sought out problematic drive videos on that car or system, but feel free.
I watched that vid when it came out, it is very impressive. No snark, it's one video need a far larger sample size. Maybe it's far better than FSD maybe not. And "gasp" maybe it's not as incredible as you think.
Now talk about the technical aspects that's what this thread is for.
All Li Auto vehicles come standard with the electric automaker’s advanced driver assistance system, Li AD Max. To achieve surround perception, the system uses one forward-facing lidar, 11 cameras, one radar and 12 ultrasonic sensors, as well as DRIVE Orin SoCs.
I guess the bolded part might be helpful?
Dang, that car is packing.
Some serious hardware right there.
The Li Auto L9, a flagship full-size SUV from the Chinese automaker Li Auto, is equipped with an advanced self-driving system known as Li AD Max. This system is designed to provide robust intelligent driving capabilities, and there’s quite a bit of information available about its hardware, onboard computing power, and some hints about its development approach. Here’s what I can tell you based on the latest details: Hardware The Li AD Max system in the L9 is built around a comprehensive sensor suite that enables high-level perception for autonomous driving. The hardware includes:
LiDAR: A forward-facing 128-line LiDAR unit with a resolution of 1200x128 and a point cloud generation rate of 1.53 million points per second. This is one of the most advanced LiDAR systems in mass-produced vehicles, offering precise detection of objects and environments, especially in challenging conditions like low light or bright glare where cameras might struggle.
Cameras: The system uses 11 cameras in total—six 8-megapixel cameras and five 2-megapixel cameras. These provide 360-degree coverage around the vehicle. Notably, the two 8-megapixel cameras facing forward have a wide 120-degree field of view and can detect objects (vehicles, pedestrians, cones) up to 550 meters away.
Radar: One forward millimeter-wave radar enhances distance and speed detection.
Ultrasonic Sensors: Twelve ultrasonic sensors assist with close-range detection, such as for parking or navigating tight spaces.
3D ToF Sensors: Inside the cabin, these time-of-flight sensors are used for features like sentry mode (monitoring the interior and exterior) and gesture-based interactions with the vehicle’s screens.
This combination of sensors is designed to give the L9 a detailed understanding of its surroundings, supporting full-scenario assisted driving capabilities. Onboard Computer The computing power behind Li AD Max is impressive and tailored for real-time processing of the massive data influx from its sensors:
Processors: The system relies on two NVIDIA DRIVE Orin processors (Orin-X chips), delivering a combined 508 TOPS (trillion operations per second). These are systems-on-chip (SoCs) optimized for AI workloads, particularly deep neural networks used in autonomous driving.
Redundancy: The dual-processor setup isn’t just for performance—it’s also about safety. The two chips provide redundancy, meaning if one fails, the other can take over, ensuring the system remains operational. This is a critical design choice for reliability in autonomous driving.
Real-Time Processing: With 508 TOPS, the L9 can process fusion signals from all its sensors simultaneously, enabling features like navigation on advanced driver-assistance systems (ADAS), lane change control, automated parking, and automatic emergency braking.
Li Auto has since announced plans to transition to the NVIDIA DRIVE Thor platform for future fleets, which promises even higher performance (up to 1,000 TOPS), but as of the L9’s current configuration, it’s running on the Orin-based setup. Training Compute While specific details about the training compute for Li AD Max aren’t fully public—Li Auto doesn’t openly disclose the backend infrastructure like some competitors (e.g., Tesla)—there are some insights:
In-House Development: Li Auto emphasizes that its AD Max system uses a full-stack, self-developed software solution for perception, decision-making, planning, and control. This suggests significant investment in training neural networks, likely using large-scale AI models similar to those mentioned in their AD Max 3.0 upgrade, which features an end-to-end algorithmic architecture with occupancy networks and spatio-temporal trajectory planning.
Data-Driven Approach: The system’s ability to handle China-specific road conditions (e.g., recognizing crossing pedestrians and two-wheelers for automatic emergency braking) implies it’s been trained on extensive datasets from Chinese driving environments. This would require substantial computational resources, possibly leveraging cloud-based or on-premises GPU clusters, though exact specs aren’t specified.
Partnership with NVIDIA: Given their use of NVIDIA’s DRIVE platforms, it’s reasonable to assume Li Auto collaborates with NVIDIA’s ecosystem for training, potentially using NVIDIA DGX systems or similar AI training platforms. These are common in the industry for developing autonomous driving models, offering the kind of compute power needed to process billions of miles of driving data.
Additional Notes
Software Features: The L9’s AD Max supports full-scenario assisted driving, including highway navigation (NOA, or Navigation on Autopilot), lane changes, and urban road operation in some cases. However, full functionality, especially in cities, depends on regulatory approval and is being rolled out gradually.
Safety and Redundancy: Beyond the dual processors, the L9 has redundant power, steering, and braking systems to ensure reliability, aligning with high safety standards for autonomous driving.
Future Upgrades: Li Auto has hinted at over-the-air updates to enhance capabilities, suggesting the hardware is future-proofed to handle more advanced models as training and software improve.
What’s Missing? There’s no explicit mention of the exact training compute scale (e.g., number of GPUs or FLOPS used in development), which companies often keep proprietary. Unlike Tesla, which has publicized its Dojo supercomputer, Li Auto’s focus seems more on the in-vehicle hardware and less on showcasing its training infrastructure. That said, their partnership with NVIDIA and the complexity of AD Max suggest a robust backend, likely involving cutting-edge AI training techniques tailored for their hybrid-electric vehicle platform. If you’re curious about something specific—like comparisons to Tesla’s FSD or more on the software side—let me know, and I can dig deeper!
The Li Auto L9, a flagship full-size SUV from the Chinese automaker Li Auto, is equipped with an advanced self-driving system known as Li AD Max. This system is designed to provide robust intelligent driving capabilities, and there’s quite a bit of information available about its hardware, onboard computing power, and some hints about its development approach. Here’s what I can tell you based on the latest details: Hardware The Li AD Max system in the L9 is built around a comprehensive sensor suite that enables high-level perception for autonomous driving. The hardware includes:
LiDAR: A forward-facing 128-line LiDAR unit with a resolution of 1200x128 and a point cloud generation rate of 1.53 million points per second. This is one of the most advanced LiDAR systems in mass-produced vehicles, offering precise detection of objects and environments, especially in challenging conditions like low light or bright glare where cameras might struggle.
Cameras: The system uses 11 cameras in total—six 8-megapixel cameras and five 2-megapixel cameras. These provide 360-degree coverage around the vehicle. Notably, the two 8-megapixel cameras facing forward have a wide 120-degree field of view and can detect objects (vehicles, pedestrians, cones) up to 550 meters away.
Radar: One forward millimeter-wave radar enhances distance and speed detection.
Ultrasonic Sensors: Twelve ultrasonic sensors assist with close-range detection, such as for parking or navigating tight spaces.
3D ToF Sensors: Inside the cabin, these time-of-flight sensors are used for features like sentry mode (monitoring the interior and exterior) and gesture-based interactions with the vehicle’s screens.
This combination of sensors is designed to give the L9 a detailed understanding of its surroundings, supporting full-scenario assisted driving capabilities. Onboard Computer The computing power behind Li AD Max is impressive and tailored for real-time processing of the massive data influx from its sensors:
Processors: The system relies on two NVIDIA DRIVE Orin processors (Orin-X chips), delivering a combined 508 TOPS (trillion operations per second). These are systems-on-chip (SoCs) optimized for AI workloads, particularly deep neural networks used in autonomous driving.
Redundancy: The dual-processor setup isn’t just for performance—it’s also about safety. The two chips provide redundancy, meaning if one fails, the other can take over, ensuring the system remains operational. This is a critical design choice for reliability in autonomous driving.
Real-Time Processing: With 508 TOPS, the L9 can process fusion signals from all its sensors simultaneously, enabling features like navigation on advanced driver-assistance systems (ADAS), lane change control, automated parking, and automatic emergency braking.
Li Auto has since announced plans to transition to the NVIDIA DRIVE Thor platform for future fleets, which promises even higher performance (up to 1,000 TOPS), but as of the L9’s current configuration, it’s running on the Orin-based setup. Training Compute While specific details about the training compute for Li AD Max aren’t fully public—Li Auto doesn’t openly disclose the backend infrastructure like some competitors (e.g., Tesla)—there are some insights:
In-House Development: Li Auto emphasizes that its AD Max system uses a full-stack, self-developed software solution for perception, decision-making, planning, and control. This suggests significant investment in training neural networks, likely using large-scale AI models similar to those mentioned in their AD Max 3.0 upgrade, which features an end-to-end algorithmic architecture with occupancy networks and spatio-temporal trajectory planning.
Data-Driven Approach: The system’s ability to handle China-specific road conditions (e.g., recognizing crossing pedestrians and two-wheelers for automatic emergency braking) implies it’s been trained on extensive datasets from Chinese driving environments. This would require substantial computational resources, possibly leveraging cloud-based or on-premises GPU clusters, though exact specs aren’t specified.
Partnership with NVIDIA: Given their use of NVIDIA’s DRIVE platforms, it’s reasonable to assume Li Auto collaborates with NVIDIA’s ecosystem for training, potentially using NVIDIA DGX systems or similar AI training platforms. These are common in the industry for developing autonomous driving models, offering the kind of compute power needed to process billions of miles of driving data.
Additional Notes
Software Features: The L9’s AD Max supports full-scenario assisted driving, including highway navigation (NOA, or Navigation on Autopilot), lane changes, and urban road operation in some cases. However, full functionality, especially in cities, depends on regulatory approval and is being rolled out gradually.
Safety and Redundancy: Beyond the dual processors, the L9 has redundant power, steering, and braking systems to ensure reliability, aligning with high safety standards for autonomous driving.
Future Upgrades: Li Auto has hinted at over-the-air updates to enhance capabilities, suggesting the hardware is future-proofed to handle more advanced models as training and software improve.
What’s Missing? There’s no explicit mention of the exact training compute scale (e.g., number of GPUs or FLOPS used in development), which companies often keep proprietary. Unlike Tesla, which has publicized its Dojo supercomputer, Li Auto’s focus seems more on the in-vehicle hardware and less on showcasing its training infrastructure. That said, their partnership with NVIDIA and the complexity of AD Max suggest a robust backend, likely involving cutting-edge AI training techniques tailored for their hybrid-electric vehicle platform. If you’re curious about something specific—like comparisons to Tesla’s FSD or more on the software side—let me know, and I can dig deeper!
Insane. Lol
I work with asics everyday, just basically means a custom "system on a chip" for anyone who wants to know.
They can be built from bare bones simple to very complex and are always made for a custom application.
These are teraflop asics, good golly.
I have no idea how they are made.
I know an asic can be designed like a microcontroller with onboard memory but these speeds its like FPGAs built in them too, crazy.
I know an asic can be designed like a microcontroller with onboard memory but these speeds its like FPGAs built in them too, crazy.
These designs have become so complex that it's likely no one person understands the entire processor, or perhaps only a few people. And it's going to get very scary when AI is used for some portions (or maybe all) of the chip, that takes knowing control away from the designers. I have trust issues here lol.
Toyota and Lexus Join Mille Miglia For The First Time
Slideshow: A five-car lineup spanning more than five decades of Toyota performance and engineering will tackle one of Italy's most celebrated automotive routes.
