Using dual Orin to create a triple integration domain controller for intelligent driving, cockpit, and parking, the hardware difficulty is not high; the challenge lies in the software. Today, let’s take a look at the triple integration domain controller from a major domestic manufacturer.
Can the integration of intelligent driving, cockpit, and parking be considered a central computing zonal E/E architecture or a software-defined vehicle? Not really. The central computing zonal E/E architecture not only needs to integrate intelligent driving, cockpit, and parking but also consolidate all systems requiring a certain scale of computing resources, such as power chassis, body, seats, etc., into one computing unit. This computing unit does not have to be a single chip; it can be two or four chips, and the chips need to be connected by PCIe switches with a bandwidth of at least 100GB/s. The connection between the central computing unit and the zonal areas should be at least 10Gb/s backbone in-vehicle Ethernet. Each zonal area needs an MCU that can run small virtual machines, and it may also deploy 10Base-T1S in-vehicle Ethernet at the edge, which might save some expensive MCUs.
The zonal architecture is almost an all-Ethernet architecture; however, cameras and display panels may still require MIPI CSI or DP’s SerDes. Given the current iteration speed of car manufacturers, moving towards software-defined vehicles or the zonal E/E era may take another ten years. Currently, cross-domain controllers are already an advanced E/E architecture; Ethernet is merely an embellishment in the E/E architecture, while CAN remains the absolute mainstream. To replace CAN with Ethernet will require at least five more years. A mixed network of CAN and Ethernet, where CAN is signal-oriented and Ethernet is service-oriented, requires multiple gateways, and the software work is also quite troublesome. The most perfect E/E architecture should be fully Ethernet with a small amount of SerDes.
Let’s get back to reality and take a look at this intelligent driving, cockpit, and parking integration architecture. The above image shows the core board part, with peripherals for the dual Orin including DRAM, UFS, eMMC, serial Nor Flash, and secure control Nor Flash. Due to the diverse functionalities, the peripheral components must be as robust as possible. The DRAM chosen is Samsung’s LPDDR5X, specifically K3KL3L30QM, with a rate of 7500Mbps and a capacity of 8GB. Recently, LPDDR5X with a rate of 8533Mbps has been mass-produced and can be upgraded to K3KL3L30CM-BGCU, with a maximum of 8533Mbps, or its capacity can be increased to 12GB, namely K3KL4L40DM-BGCU, also with a rate of 8533Mbps. Orin supports LPDDR5X, while Qualcomm’s SA8255/SA8295 only supports LPDDR4X. Currently, the only mass-produced automotive SoCs that support LPDDR5X are Nvidia’s Orin and Ambarella’s CV3, while NIO’s Adam also supports it. At present, Samsung’s LPDDR5X has a temperature range of -40℃ to +85℃, while Micron’s is -40℃ to +95℃, slightly higher by 10 degrees, with Micron’s model being MT62F1G64D4EK, which can also reach 8533Mbps.
UFS is selected as 3.1, while the mainstream is UFS2.1. The automotive-grade UFS3.1 just started mass production in early 2023, with UFS2.1 having a maximum read speed of 850MB/s, and UFS3.1 being 1700MB/s. The maximum continuous write speed for UFS2.1 is 260MB/s, while UFS3.1 is 1400MB/s. The UFS selected is Samsung’s KLUEG8UHYB-B0EQ051, supporting AEC-Q100 Grade 2, with a maximum continuous read speed of 2000MB/s and a continuous write speed of 700MB/s. For automotive systems, read speed is more important than write speed, which is optimized for automotive chips. The eMMC selected is Micron’s MTFC32GAZAQHD-AAT with a capacity of 32GB. The Nor Flash chosen is Taiwan’s Winbond MX25U51279GXDR00, a new product from June 2023, adopting a multi-I/O design with a maximum speed of over 400Mbps.
Adding a power sequencing system, namely the MPQ79700FS, which is a 12-channel functional safety power sequencer designed for automotive advanced driver assistance systems (ADAS) and autonomous driving platforms, providing necessary control and power sequencing for the entire platform. MPQ79700FS includes a crystal driver, a real-time clock (RTC) with alarm functionality, and a configurable monitor accessible via I2C interface (watchdog), while providing low-level effective system reset and interrupt output. It integrates safety mechanisms such as built-in self-test (BIST) to achieve high diagnostic coverage, allowing the system to meet the target ASIL level. It is developed based on MPS’s advanced MPSafe™ functional safety product development process, which has been independently certified to meet ISO26262 standards.
Dual Orin naturally uses dual safety MCUs, generally choosing Infineon’s TC397. The cockpit system requires many Type-C interfaces, especially for new cockpits, considering AR/VR gaming, external game consoles, and charging needs for rear and even third-row passengers, as well as various laptops or tablets.
Only the DP and CSI ports of Orin can be used for conversion.
There are three LiDAR interfaces, all of which are gigabit-level; currently, 99% are hundred-megabit-level. LiDAR has redundant interfaces, with four 4D millimeter-wave radar interfaces, which are sufficient at the hundred-megabit level. TCAM, or Telematics and Connectivity Antenna Module, also known as T-BOX, considering V2X, requires the highest 2500 megabit level, with a total of three T-BOXes, one of which is for OTA upgrades, only needing gigabit. DTOF refers to direct TOF cameras, which can also be flash LiDAR; the likelihood of flash angular LiDAR is higher. Given the rapid technological changes, sufficient interfaces have been reserved. DSSAD stands for Data Storage System for Automated Driving. Translated, it refers to the data storage system for autonomous vehicles. EDR focuses on accidents (mainly the vehicle’s speed changes triggering EDR device recording), while DSSAD distinguishes responsibility between the autonomous driving system and the driver. ADPU stands for Autonomous Driving Navigation Unit, while ZC/R is the right-side zonal control. However, I suspect that most of the in-vehicle network on the vehicle side is CAN, with a maximum speed of only CAN-FD, and in 99.9% of cases, the speed does not exceed 5Mb/s, making gigabit Ethernet a serious waste. Based on the MCU interfaces below, it should only be reserved and not in use.
The most advanced and powerful Ethernet switch currently selected is Marvell’s 88Q5192, which is Marvell’s third-generation secure automotive Ethernet switch. The 88Q5192 is a 16-port Ethernet switch, currently the Ethernet switch with the most ports, and integrates 1000BASE-T1, 100BASE-T1, and 10BASE-T1S PHY, fully compliant with applicable IEEE 802.3 standards. This 16-port Ethernet switch provides 12 integrated PHYs, with 4 supporting dual-speed 1000/100BASE-T1, 6 supporting dual-speed 100/10BASE-T1(S), and 2 supporting 100BASE-T1/TX. Other supported interfaces include 2 multi-speed 10Gb SerDes (10G/5G/2.5G/1Gbps), 2 multi-speed 2.5Gb SerDes (2.5G/1Gbps), 2 RGMII/MII/RMII, and 2 PCIe Gen3 x1 interfaces. The port interface options provide flexible configurations for connecting external devices, such as 2.5/5/10GBASE-T1 PHY, or uplink to the host SoC. This makes the device an ideal choice for in-vehicle network (IVN) applications, such as advanced driver assistance systems (ADAS), zonal control modules, and central gateways. This switch includes a high-performance dual-core ARM® R52 CPU that operates in a locked mode, with dedicated on-chip memory to support time-sensitive networking (TSN) protocols, such as Precision Time Protocol (PTP) and security firewalls to prevent external malicious attacks. The switch includes many advanced security features, including 802.1AE MACsec, providing link security to prevent man-in-the-middle attacks, denial of service (DoS) engines for deep packet inspection (DPI) TCAM, and trust boot features to ensure the security of the vehicle network. It also includes an embedded hardware security module that enhances device security by supporting secure and encrypted boot and managing security features like MACsec.
The above image shows the Ethernet structure of an MCU, with the Ethernet physical layer using the currently strongest Ethernet physical layer, namely Marvell’s 88Q4364. The 88Q4364 device is a single-pair Ethernet physical layer transceiver (PHY) that supports operation over shielded twisted pairs (STP). The transceiver implements the Ethernet physical layer portion defined by the IEEE 802.3ch standard for 2.5G/5G/10GBASE-T1. The 88Q4364 integrates MACsec to prevent layer 2 in-vehicle network security threats. MACsec protects data exchange on a hop-by-hop basis and prevents attacks such as intrusion, man-in-the-middle attacks, and replay attacks. It supports speeds up to 10Gbps, with 10G/5G/2.5G USXGMII, 10Gbps XFI, 5Gbps XFI/2, 2.5Gbps 2500BASE-X, or 2.5Gbps, with the OCSGMII interface directly connecting to automotive-grade GPU, CPU, Ethernet switches, and electronic control units (ECUs) to support the operations and network speeds required for in-vehicle networks (IVN).
For the radio, Bluetooth, and MCU section, considering exports, DAB is essential. The SAF360X is a DAB digital radio tuner, and there are many microphones in the vehicle, with three rows needing to have them. Naturally, it relies on ADI’s A2B bus transceiver AD2428. The GPS selected is u-blox’s F9K, which can support L1/L2/E5B and L1/L5 frequency bands, greatly enhancing flexibility, satellite signal availability, and security. The ZED-F9K-01A combines multi-band, multi-constellation global navigation satellite system (GNSS) technology with high-precision RTK (real-time kinematic) inertial navigation, achieving decimeter-level accurate positioning. The u-blox ZED-F9K-01A natively supports u-blox PointPerfect GNSS enhancement service, providing multiple GNSS and IMU outputs in parallel to support various possible architectures, including ultra-low latency, 50 Hz inertial position output. For high-precision positioning, the D9S is required, which is an L-band satellite receiver. The D9S receives correction data via satellite L-band channels using the SSR SPARTN data format. This module uses encryption technology to securely provide PPP-RTK GNSS correction data, including data provided by the u-blox PointPerfect service. There is also a D9C, which utilizes the free-to-air centimeter-level enhancement service (CLAS), provided by Japan’s QZSS (Quasi-Zenith Satellite System) constellation through L6 frequency channels, offering centimeter-level positioning services. In eastern Guangdong, China, QZSS is still usable, although the signal is weak, and it is currently free, but will definitely charge in the future.
FlexRay is unnecessary; the future trend should be Ethernet. FlexRay is only used for chassis parts by Mercedes-Benz, BMW, Audi, and Volvo, being expensive and complex in protocol. The main connector is also connected to 12 ultrasonic sensors via SPI interface. TJA1042 connects to the body controller, namely ZC/R and ZC/L, and is also connected to ADPU, body, and power domains.
Video output, USB, and Type-C sections can use Orin’s DP to directly output one, creating a dual-link screen for the vehicle machine and instrument panel, and achieving 4K display without any issues. Additionally, one auxiliary screen can be used through Orin’s eDP output, via ADI’s MAX96589 serializer, which supports two MST outputs, allowing for two displays, one for the passenger seat and one for the central display, or two rear seat displays.
Image source: Internet
Audio, Bluetooth, and WiFi: The audio here is a bit underpowered, possibly considering there is an external audio power amplifier. Bluetooth requires two, as there is a need for Bluetooth calls in the rear seat.
This is the camera board, taking the example of an 8-megapixel camera with high-speed lossless transmission requirements. Generally, it has 3264*2448=7990272 pixels. Based on RGB three colors 24bit and a camera frame rate of 30fps, the data generated by this camera per second is about 7.68Gbps bandwidth. The camera’s data format usually consists of RAW RGB and YUV. There are three common levels of YUV: YUV444, YUV422, and YUV420. Among them, YUV444 is pixels X frame rate X bits X 3, which equals 5.76Gbps; YUV422 is pixels X frame rate X 8 bits X 2, which equals 3.84Gbps; YUV420 is pixels X frame rate X 8 bits X 1.5, which equals 2.88Gbps. ADAS usually does not consider color much, and YUV420 is sufficient; of course, it can be compressed a bit more, but the author believes it is better not to compress, and it may also need to consider color in the future. Therefore, four 8-megapixel cameras require 11.52Gbps. The MAX96712 can also be used, but a sensor like AR0820 with 8.3 megapixels is a bit stretched. Currently, the top-tier automotive camera image sensor is Sony’s IMX735, with effective pixels reaching 17.42 million; one requires 6.27Gbps bandwidth, while the total bandwidth of MAX96712 is only 6Gbps. Additionally, considering high-speed scenarios, a higher frame rate of 40Hz or 45Hz will significantly increase bandwidth, so the best choice here is ADI’s currently top product, MAX96792, which can support 12Gbps bandwidth, but MAX96792 is for two camera inputs, so two are needed. Here, four 360-degree surround view cameras also used 8-megapixels, which is completely unnecessary and could be downgraded to 4-megapixels or 2-megapixels, making it much easier to use MAX96712. Of course, if using 8-megapixels, MAX96712 can still manage.
The integration of intelligent driving, cockpit, and parking with two Orins is somewhat challenging. Most of the time, one Orin is likely responsible for the cockpit while the other handles intelligent driving. Although the most advanced Ethernet switch is used, the maximum bandwidth is only 1.25GB/s. It should add a PCIe switch, such as Microchip’s PM43036B1, supporting 36 lanes. Orin’s PCIe is 4th generation, with a maximum of 2GB/s bandwidth per lane, with 16 available lanes, leading to a maximum of 32GB/s. This bandwidth is insufficient for Orin to achieve cascading doubling effects; after all, the latest NVLink can reach 900GB/s, but it’s still better than Ethernet. Of course, this architectural design is already among the most advanced globally.
Disclaimer: The views and data in this article are for reference only and may differ from actual conditions. This article does not constitute investment advice, and all opinions and data in the text only represent the author’s position, without any guidance, investment, or decision-making opinions.
