How to choose autonomous driving domain controller chip?

It is not easy to find a perfect chip in the domain control of intelligent driving. ADLINK is an edge computing solution provider and a partner of NVIDIA, Intel and other manufacturers. As intelligent driving becomes more and more popular, ADLINK aims to make high-performance chips "on the car" in a safer and more reliable way.

Domain Controller CPU Chip Selection: X86 vs ARM

The computer and IT industries we are familiar with are inseparable from the X86 chip architecture, but X86 is rarely used in vehicles. One of the important reasons is the constraints of reliability and security - laptops can freeze and blue screen, but not in vehicles. X86's interrupt processing mechanism, the real-time nature of the operating system that matches it, and its high power consumption characteristics all affect its use in vehicles. However, after careful consideration, ADLINK still chose X86 as the CPU architecture because it saw more advantages in domain control.

First of all, X86 has good ecological compatibility in software. Some software trained by software companies on servers can be directly installed on the X86 architecture without any transplantation. In addition, X86 also performs well in hardware scalability, because X86 has a lot of PCIE, so it can expand almost any function we need.

Of course, the biggest advantage of the X86 architecture is performance. The performance of a chip includes overall performance and single-core performance. In a domain controller, in addition to overall performance, single-core performance is also very important. Although the overall performance of the latest multi-core ARM chips is also very powerful, the single-core performance is only about 1/3 of that of Intel's Xeon series chips. Even the well-known I7 and I9 have reached twice the single-core performance of ARM. Therefore, X86's performance, especially single-core performance, has become a major advantage in domain controllers.

Image source: ADLINK Technology

In recent years, the security of X86 has been continuously evolving, and now there are many security processing mechanisms inside X86.

There are many hardware mechanisms inside X86 to ensure that errors can be corrected. For example, for some correctable errors, the leaky bucket algorithm and setting of some correctable error thresholds can be used to directly implement error correction; for uncorrectable errors, corresponding error handling can be performed through software recovery, positioning reporting, etc. 

Image source: ADLINK Technology

In the current X86 error handling situation, not all errors will cause a crash. A variety of error checking and handling methods can allow the system and equipment to run normally. Even if an unrecoverable error occurs in the X86 system, it can still be safe with the redundancy function of the hardware platform. Therefore, applying the X86 architecture to domain control can undoubtedly ensure the security of the platform.

GPU chip selection: dGPU vs iGPU

In addition to CPU, there are also a variety of chip options for GPU.

Image source: ADLINK Technology

Commonly used GeForce chips are relatively cheap, but due to their high power consumption, they are not suitable for domain control products; Tesla chips are usually used in data centers, and Quadro is mainly used for high-performance computing such as simulation or calculation. Quadro Embedded chips are sold as single chips rather than cards. Currently, only a few companies can get this chip, and ADLINK is one of them because of its cooperation with NVIDIA.

After obtaining this chip, the GPU and CPU can be made into a motherboard, which can meet the reliability requirements such as vehicle vibration. NVIDIA's Pegasus product is based on the Quadro Embedded chip.

With the rapid development of algorithms, the requirements of domain control for computing power are changing every day. If the GPU is made into a hardened board, it will be difficult to iterate and update when the computing power is insufficient. Therefore, ADLINK made the chip into an embedded MXM GPU card with a common interface. While being shockproof and moisture-proof, it also ensures that the board can be flexibly replaced.

ADLINK MXM GPU cards cover a wide range of performance. In the early stages of a product, high computing power can be used to quickly complete product prototype functions. When the algorithm is optimized in the future and a lower-performance GPU needs to be replaced to reduce costs, the GPU chip can be directly replaced, which brings great flexibility to domain control.

Image source: ADLINK Technology

Compared with the standard PEG card, the GPU MXM card is only 1/6 of the size and about 1/10 of the weight. With the same performance, the power consumption is only 2/3 or even 1/2 of the former, and the temperature range is larger, making it more suitable for vehicle-mounted scenarios and domain controllers.

Image source: ADLINK Technology

Compared with SoC chips, the Quadro Embedded mentioned above has more flexible configuration and more versatile software, but it is not without weaknesses. For example, it needs to be used with an additional CPU to run. However, a SoC integrates CPU, GPU, ISP and other IOs, which can meet all functional requirements and has lower power consumption.

So, how should we choose between these two chips, each with its own advantages? In fact, this is not an exclusive choice. We can use the two chips together to maximize the advantages of both.

Domain Controller SoC Chip Selection: NVIDIA Orin vs Horizon J5

At present, in addition to NVIDIA, Horizon Robotics, as a domestic company, has also made some arrangements for SoC chips.

In the domain control field, power consumption is a very important indicator. Compared with NVIDIA Orin, although the overall digital performance of Horizon J5 is slightly insufficient, its functions are very similar. More importantly, its power consumption is only 30W, which is only half of Orin. Therefore, both can be used for L2+/L3+ autonomous driving, and can be selected according to different scenario requirements.

High-performance chips are being installed in vehicles

From all the chips introduced above, it can be seen that ADLINK is not a complete chip manufacturer, but forms solutions based on cost, performance and reliability.

ADLINK is committed to multiple fields such as intelligence, urban measurement, medical, transportation, telecommunications, and military and aerospace, and entered the product design of autonomous driving and domain control three years ago. ADLINK has been deeply involved in embedded components for many years and has rich experience in this field.

Image source: ADLINK Technology

Image source: ADLINK Technology

The difference between ADLINK and other companies in domain control is mainly reflected in the reliability and performance of the overall solution. ADLINK has cooperated with Intel, NVIDIA, Qualcomm and other manufacturers, and has significant advantages in chip resources; in addition, ADLINK also has close cooperation with Huawei in China.

As L2+/L3+ autonomous driving is increasingly used in smart cars, all companies are looking for ideal SoC chips. However, SoC chips that can meet low-operating conditions and high performance have not yet appeared. Based on customers' needs for domain control, ADLINK has developed a solution that combines security and high performance by cooperating with other chips, making domain control chips closer to the "ideal".

The above content comes from the keynote speech "Selection of Autonomous Driving Domain Controller Chips" delivered by Pan Deng, Technical Director of ADLINK Technology (China) Co., Ltd. at the 2022 Second Smart Car Domain Controller Innovation-Cloud Forum hosted by Gasgoo on June 29, 2022.