NIO unveils first self-developed smart driving chip
At this annual conference, NIO ET9 was unveiled, self-developed intelligent driving chip, 900V architecture, cylindrical battery and intelligent chassis system, NIO's annual conference released a series of heavyweight new products.
Of particular interest is that this car uses NIO's self-developed intelligent driving chip, which is expected to exceed 1,000 TOPS of arithmetic power.
At the conference site, NIO officially released the first self-developed intelligent driving chip - Shengui NX9031. Li Bin said that previously NIO's intelligent driving system used four Nvidia Orin X chips, the overall arithmetic power of 1,000 + TOPS, and Shengui NX9031 can realize the industry's current four flagship intelligent driving chip performance. The NX9031 can realize the performance of the current four flagship intelligent driving chips in the industry. In other words, the arithmetic power of the Shengui NX9031 will be above 1,000 TOPS.
Shengui NX9031
It is understood that the Shengui NX9031 is manufactured using a 5nm vehicle-specification process, has more than 50 billion transistors, supports 32-core CPU. running more than 6 trillion instructions per second, the performance exceeds the industry's flagship by a factor of 2.5, and is capable of parallel computation, perception, computation, regulation and control, data closure, group intelligence and other ultra-multiple real-time tasks.
The self-developed image, signal processor, and ISP of the Azure Shengui NX9031 chip has a processing delay of less than 5 milliseconds, which is 75% shorter than the industry average; 26bit processing bit width, and 6.5 billion pixels per second processing power. Moreover, Shengui also has the performance of low power consumption and high security, and the NIO eT9 also has dual-chip millisecond backup capability.
In addition, through the self-developed inference acceleration unit NPU (NPU TPP arithmetic, Total Power Performance <4800), flexible and efficient operation of various types of AI algorithms, together with strong security redundancy capabilities, Shengui NX9031 is able to dynamically wake up various subsystems in microseconds, providing a peace of mind and efficient intelligent driving experience.
According to Li Bin, this chip has excellent performance in running BEV+Transformer algorithms.
Shengui NX9031 Specifications
NIO's thoughts on the next generation of smart driving chips
Computing power is the new horsepower, the digital "big displacement". Chips, on the other hand, are the main carriers of this power.
Historically, the first in-vehicle electronic chip is the Tungsram-1 launched in 1968, which is mainly used for engine control, ignition system, and body control. Subsequently, Fairchild/TI/Motorola/Intel/Bosch and other companies began to provide automotive chip products. Early automotive chips, focused on the body domain, in the engine, chassis, seats and other areas, to provide a role. Until today, various types of MCU/power supply and other chips have played an important role in each controller in the body domain.
In 1976, Aston Martin launched the industry's first Lagonda model with a digital instrument panel, followed by more and more models with digital instrumentation, entertainment center control, head-up display (HUD) to the market, which led to the cockpit entertainment domain chip on the stage of history. And the star of the show is the processor.
Intelligent cockpit development to today, there are several obvious trends: 1, more and more screens, the application is more and more complex, more and more high arithmetic requirements. CPU rise after the GPU rise, GPU rise after the NPU rise; 2, entertainment-related functions to the consumer electronics to look at the same pursuit of the processor, rapid iteration. However, at the same time, automobile safety-related functions such as speed and status display are pursuing high reliability, long iteration cycle, and running wildly in the direction of "schizophrenia". Cockpit master control chip and domain control will follow the trend, and there will be some new changes, which we will share with you later.
In the field of intelligent driving, cruise control is the starting point. The first car to support cruise control was the "Chrysler Imperial" launched by Chrysler in 1958, which did not use the Smart Driving chip at that time. Among the batch of L2 assisted driving cars that can support both horizontal and vertical control, the most widely known is the Model S launched by Telsa in 2014, which is equipped with Mobileye's EyeQ3 chip.
Telsa's assisted driving uses Mobileye's EyeQ4H chip, which works in conjunction with the S32V chip, with the EyeQ4H providing perception and the S32V being responsible for regulation and control. All perception algorithms are still provided by Mobileye, and the Azalea algorithm team is responsible for the development of regulation and control. In order to realize the sense of open, closed-loop data, improve the development efficiency and iteration speed, from NT2, NIO from the sensor, domain control hardware, AD algorithms full-stack self-research! 1000TOPS computing power of the self-driving domain control Adam and Aquila supersensory system (including a number of 8M cameras, ultra-long-range LIDAR, etc.) together, so that NIO's full range of navigational aids NOP + open the speed and quality of rapid growth. grow by leaps and bounds. The tweet posted by Shaoqing last week mentioned that NIO has accumulated nearly 200,000 kilometers of validated routes in a short period of time in more than 130 cities, the importance of an intelligent computing platform with sufficient performance, you know.
From Telsa's Model S to NIO's ET7, the computing power of the Smart Driving system has increased by thousands of times. Which direction will Smart Driving chips develop in the future? What kind of chip is a good Smart Driving chip? Is it "all you can eat"? No, it's not! Calculation power is only enough to solve the problem of subsistence, in order to "run a well-off", AD chip needs to be born for security, data, evolution!
Crashworthiness is a good car, not crash safer. Some time ago, Shaoqing mentioned that big data shows that the accident rate of our users using assisted driving and manual driving has been lower. I predict that after the continuous upgrading of intelligent driving, it will be at least 1 order of magnitude safer than human driving. As the brain of the smart driving system, the smart driving chip should be: 1. safe and reliable enough to pass the AECQ100 standard is a piece of cake. At the same time, we also need to achieve system-level functional safety, FIT (faults per billion hours) to one hundred, DPPM (per million pieces of the repair rate) single-digit; 2, the various modules should have ultra-low latency for the algorithms to leave enough time to deal with and respond to; 3, reliable and fast back-up, the system is easy to not go wrong. And once the error, the backup system can instantly take over completely.
Data is an important fuel for Smart Driving. The speed at which the NIO Urban Pilot now opens roads and covers cities proves the importance of the 4 Orin and Aquila super sensory systems that come standard in the entire lineup. The Smart Driving chip needs to "feel the fish" while driving the car: pay attention to extreme scenarios in real time, record special traffic information, verify the availability of road sections, and realize the rapid closure of the loop. Therefore: 1. MSP (Mixed Signal Processor) has to receive data from various sensors and feed them to CPU and NPU after pre-processing; this module has to be powerful enough to take care of not only the camera but also the LIDAR and all sensors; 2. System bandwidth and car-cloud link are powerful enough to allow the data to "ride wild" at will. "The system bandwidth and car cloud link should be strong enough to allow data to go wild; 3. The processing power should be strong enough and compatible with various mainstream smart driving algorithms, CPU power is not too much, but NPU power is only too little.
Intelligent electric vehicles are "high-tech products" that are continuously iterated over a long period of time, and NIO's current 4 Orin domain-controlled Adam will be domain-controlled for the next 5 years due to sufficient computing power. The algorithm team can upgrade the algorithms on Adam, fix long-tailed bugs, and continue to improve the smart driving experience to stay ahead of the industry. So a Smart Driving chip, at least 5 years of algorithm evolution needs to be met.
If you like this article, may wish to continue to pay attention to our website Oh, later will bring more exciting content. If you have product needs, please contact us.
