Self-driving cars or driverless cars use on-board sensors such as cameras, lidars, millimeter-wave radars, ultrasonic radars, navigation & IMU inertial navigation, and V2X modules to perceive the surrounding environment, and make decisions and judgments based on the acquired information. The working model is used to formulate corresponding strategies, such as predicting the motion state of the vehicle, other vehicles, pedestrians, etc. in a period of time in the future, and planning the collision avoidance path. After planning the route, the next step is to control the vehicle to drive along the desired trajectory. This involves sensor environment perception, high-precision map/GPS precise positioning, V2X car networking communication, multi-sensor data fusion, control decision-making and planning calculation, electronic control and execution of calculation results, etc. In this process, a strong The "brain" is used to perform real-time analysis, process massive amounts of data, and perform complex logical operations, which requires very high computing capabilities.
Based on the perception and computing power requirements of autonomous driving, the core controllers of autonomous vehicles need to have strong computing performance. It is generally believed that the computing power required for L2 is less than 10TOPS, the computing power required for L3 is 30~60TOPS, and the computing power required for L4 Power>100TOPS, the computing power required for L5 is currently not clearly defined (it is predicted that at least 1000 TOPS is required), and the current computing platform can only meet the requirements of some L3 and L4 autonomous driving.
Huawei believes that in the future, every car will be a Mobile Data Center (MDC) mounted on wheels. In response to the demand for computing platforms for autonomous driving, Huawei launched the MDC intelligent driving computing platform solution, which integrates Huawei’s self-developed Host CPU chip, AI chip, ISP chip and SSD control chip, and is optimized through the integration of the underlying software and hardware. Leading the industry in terms of time synchronization, accurate processing of sensor data, multi-node real-time communication, minimum noise floor, low power management, and fast and safe startup.
After years of development, Huawei has established a complete chip system, including the Barong series of chips that support 5G, the Shengteng series of AI chips developed based on the new DaVinci architecture, and the CPU processor chips on mobile phones. Kirin series and Kunpeng series of server-level processor chips. Among them, the Shengteng series of AI chips mainly focus on the demand for AI computing power. Shengteng 310 uses Huawei’s self-developed efficient and flexible CISC instruction set. Each AI core can complete 4096 MAC calculations in one cycle. It integrates tensor, vector, scalar and other arithmetic units, and supports multiple mixed precisions. Calculate, support data accuracy calculations in training and inference scenarios.
As an NPU, Shengteng 310 integrates the advantages of FPGA and ASIC chips, including the low power consumption of ASIC and the programmable and flexible features of FPGA, so that its unified architecture can be adapted to a variety of scenarios, with power consumption ranging from several Ten milliwatts to hundreds of watts, flexible multi-core stacking, can provide the best energy consumption ratio in a variety of scenarios.
In comparison, the GPU used by Nvidia is a general-purpose AI chip with high computing power but high power consumption. Specifically, Nvidia’s Xavier has a hashrate of 30 TOPS, power consumption is 30W, and energy efficiency is 1 TOPS/W. In contrast, Huawei Shengteng 310 has a hashrate of 16 TOPS, power consumption is only 8W, and energy efficiency is 2. TOPS/W.
In addition, in the field of autonomous driving chips, Mobileye's EyeQ series is representative of typical ASIC chips. However, in terms of computing power, Mobileye EyeQ4 has a computing power of 2.5 TOPS, a power consumption of 3W, and an energy efficiency of 0.83 TOPS/W. Under the circumstances, the advantages of Huawei Shengteng 310 are also particularly obvious. Moreover, the Shengteng 310 chip will continue to iteratively optimize at a rate of once every two years.
Huawei's MDC 600, a computing platform that can support L4 level autonomous driving, is based on 8 Shengteng 310 AI chips, and it also integrates a CPU and corresponding ISP modules. The computing power of MDC 600 is as high as 352TOPS (TOPS: trillion times per second), and the power consumption and computing power ratio of the overall system is 1TOPS/W. Although the MDC 600 is aimed at L4 level autonomous driving, the market is not currently in strong demand for L4 level autonomous driving. Therefore, Huawei introduced the MDC 300. MDC 300 is mainly aimed at L3 level automatic driving, that is, application scenarios such as congested car following, high-speed cruise, and automatic parking. After the demand in the advanced autonomous driving market rises, Huawei will make more series of new products such as MDC 800/900.
In summary, compared with other autonomous driving computing platforms in the industry, Huawei MDC has the technical advantages of "three highs and one low":
High performance: Equipped with multiple Huawei's latest artificial intelligence Ascend chips, MDC600 can provide up to 352TOPS of computing power, meeting the needs of L4 level autonomous driving. High computing power means that more external sensor data streams (such as cameras, millimeter wave radar, lidar, GPS, etc.) can be accessed and processed in real time, providing safer and more reliable computing power support for autonomous driving, and can cope with more complex processing Road conditions.
High safety & reliability: Integrating Huawei's 30 years of ICT equipment R&D, design, and manufacturing experience, end-to-end redundant backup design to avoid single points of failure; support -40°C~85°C ambient temperature to cope with harsh external environments ; Comply with the industry's vehicle-level reliability and functional safety requirements (such as the ASIL D of ISO 26262).
High energy efficiency: industry-leading end-to-end 1TOPS/W high energy efficiency (industry generally 0.6TOPS/W). The main value of high energy efficiency lies not only in energy saving and extended cruising range, but also in lower power consumption and temperature under the same computing power, improving the reliability of electronic components, and without the need to configure vulnerable parts such as fan heat dissipation or water cooling heat dissipation, and reduce The volume reduces the impact on the existing structure of the vehicle.
Deterministic and low latency: The underlying hardware platform needs to be equipped with a real-time operating system, and efficient integration of the underlying software and hardware is optimized and integrated. The kernel scheduling delay is less than 10us, and the internal node communication delay is less than 1ms, which is an end-to-end autonomous driving belt for customers. Comes with a low latency of less than 200ms (typically 400~500ms in the industry) to improve safety in the process of autonomous driving.
Huawei MDC is also an open platform with the characteristics of component service, interface standardization, and development tooling. Based on this platform, autonomous driving algorithms and functions can be quickly developed, debugged, and run. For different levels of autonomous driving algorithms, a set of software architecture and different hardware configurations are adopted to support the smooth evolution and upgrade of L3~L5 autonomous driving algorithms. For example, in an architecture that can achieve L4 level autonomous driving, the demand for Robo-taxi, mid-to-high-end models, and ordinary models can be achieved by increasing or reducing the number of sensors such as lidars and cameras and computing power configurations. In terms of computing power, the number of MDC intelligent driving computing platform Shengteng 310 chips can be increased or decreased according to actual needs to meet the different needs of autonomous driving for computing power.
In addition, Huawei's MDC intelligent driving computing platform is also compatible with AUTOSAR and ROS. Combined with the tool chain and HIL simulation platform provided, car companies can flexibly and quickly develop intelligent driving applications of different levels.
Based on the MDC solution, Huawei and Audi launched a joint innovation of L4 autonomous driving. The test results of Audi Q7 cars equipped with Huawei MDC in a certain area in China showed that on a dimly lit road in a certain urban-rural junction in the evening, it is aimed at complex traffic conditions, such as unclear lanes, pedestrians crossing the road, bicycles, and electric vehicles And so on, successfully completed the automatic driving functions in scenarios such as high-speed cruise, congestion following, traffic light recognition, pedestrian recognition, and automatic parking in underground garages. The autonomous driving-related algorithms developed jointly and innovatively have also achieved good results in the industry's authoritative KITTI's 2D, 3D, BEV and other test items. In addition to Audi, companies such as FAW, Volvo (passenger cars), Dongfeng, Suzhou Jinlong, Shandong Haorui Intelligent, and Neolithic have also cooperated with Huawei in autonomous driving.
Facing the trend of the traditional auto industry upgrading and transformation, consumer market upgrading, saving lives, and universalizing equal travel rights, Huawei MDC will become an enabler through chip-level innovation capabilities, platform-level engineering capabilities, and a complete development, commissioning, and diagnostic tool chain. The "car brain" capable of autonomous driving has installed an intelligent AI engine for the traditional automobile industry, and together with automobile manufacturers, it will lead autonomous driving into the fast lane.