Architecture Analysis of Automotive Ethernet Protocol

Whether it is software-defined cars, distributed ECUs or adaptive Autosar, they are inseparable from the basic technology of the smart car era, automotive Ethernet. For hardware engineers, automotive Ethernet physical layer and switches are the chips they pay the most attention to. This is also a field with a profit margin far exceeding that of high-computing chips, and it is also a field that is basically monopolized by European and American companies.


The above picture is a 7-layer OSI model and standard distribution diagram of the entire automotive Ethernet. We most often mention TSN or EAVB, and the physical layer standard is rarely mentioned. Because most engineers don’t deal with the physical layer.



Many people are talking about domain controllers, service-oriented architectures, distributed computing or software-defined cars, but they don’t know that the key 1G automotive Ethernet physical layer chip will only be SOP in 2020, and the multi-G bandwidth automotive Ethernet physical layer standard was just completed last year, and SOP will take about one year or two to three years. Without this chip, domain controllers, service-oriented architectures, distributed computing or software-defined cars are all castles in the air.

The comprehensive use of domain controllers, service-oriented architecture, distributed computing or software-defined cars all require Ethernet as the backbone network, that is, multi-G Ethernet. The first chip that supports the multi-G automotive Ethernet physical layer was officially launched in November 2020, and mass production is estimated to be at the end of 2021 or early 2022, or possibly at the end of 2022. The use of a single domain controller is also inseparable from the 1G automotive Ethernet physical layer chip, which was mass-produced in 2019.

First of all, why do we need to develop an independent physical layer standard for automotive Ethernet? Wouldn't it be better to use the traditional Ethernet physical layer?



The biggest difference between traditional Ethernet and automotive Ethernet is that traditional Ethernet requires 2-4 pairs of wires, while automotive Ethernet only requires one pair, and it is unshielded. This alone can reduce connector costs by 70-80% and weight by 30%. This is the main reason for the birth of automotive Ethernet. It is also to meet the EMC electromagnetic interference in the car.



List of common automotive transmission interface characteristics. There are



four physical layer standards for automotive Ethernet. Automotive Ethernet is ambitious. 10Base-T1S is an attempt to replace the traditional CAN network. The first physical layer chip of the 1000Base-T1 standard is Marvell's 88Q2112. Although it was launched in October 2015, it was not mass-produced until 2019. Typical applications include Nvidia's flagship box Pegasus.

Tesla's latest HW3.0 uses Marvell's 88EA1512, which is a traditional Ethernet physical layer standard from 20 years ago. In October 2020, Marvell launched the third-generation product. In fact, there are two generations of 88Q2112. The first generation was not mass-produced, but only took the first place. So Marvell's third-generation model is 88Q222xM.

The third generation specifically adds support for Open Alliance TC10 for sleep mode and wake-up. It can meet the most stringent ASIL-D standard and reach the level 1 standard in AEC-Q100 temperature, which means it can withstand a maximum temperature of 125°. Usually, physical layer chips are level 2 standards, that is, 105° high temperature. In addition to Marvell, Texas Instruments and Broadcom can also provide 1000Base-T1 standard physical layer chips. Taiwan Realtek has switch chips that support the 1000Base-T1 standard physical layer.

Broadcom has taken the lead in the field of NGBase-T1 physical layer chips. In November 2020, Broadcom announced the launch of BCM8989X and BCM8957X. BCM8989X is the industry's first MACsec physical layer chip corresponding to the NGBase-T1 (ie IEEE 802.3ch) standard. So far, Broadcom is the only manufacturer that can provide NGBase-T1 chips. BCM8957X is the industry's first L2/L3 automotive Ethernet switch chip that supports rates of 10Mbps to 10Gbps. Tesla and Broadcom have cooperated on the next-generation FSD chip, but it is unlikely to embed the multi-G physical layer into the FSD because the physical layer chips are generally independent. There is also an IEEE 802.3cy, which is the physical layer standard for automotive Ethernet above 10G, supporting 25, 50, and 100G respectively.

The IEEE's in-vehicle Ethernet physical layer is basically a copy of the OPEN Alliance's standards. The OPEN Alliance is a non-profit automotive industry and technology alliance that aims to encourage the large-scale use of Ethernet as a standard for the Internet of Vehicles. OPEN Alliance members use the scalability and flexibility of Ethernet to achieve low-cost communication in the Internet of Vehicles and reduce communication complexity. Ethernet communication networks are also an important part of realizing future autonomous driving and intelligent connected car functions. Since its establishment, the OPEN Alliance has grown to nearly 400 members.

10BASE-T1S is IEEE 802.3cg, which is TC14 of OPEN Alliance; 100BASE-T1 is IEEE802.3bw, which is TC1 of OPEN Alliance; 100/1000BASE-T1 ECU test standard is TC8 of OPEN Alliance; 1000BASE-T1 is IEEE802.3bp, which is TC12 of OPEN Alliance; 2.5/5/10GBASE-T1 is IEEE802.3ch, which is TC15 standard of OPEN Alliance.

OPEN Alliance has 18 Promoter members. In November 2011, BMW, Broadcom and NXP established OPEN Alliance. Later, Marvell (originally an Adopter member, later promoted), GM, Toyota, Volvo, Volkswagen, Jaguar Land Rover (once a member, now withdrawn), Hyundai, Bosch, Renesas, Volkswagen, Renault (once a member, now withdrawn), Continental Automotive, Mercedes-Benz (once a member, now withdrawn), Samsung Harman (once a member, now withdrawn), Hyundai, Taiwan Realtek (once a member, now withdrawn). Promoter members can only become Promoter members if they are invited by the original three companies.

In addition, you can become an Adopter member by paying $1,500 per year, including 20 Chinese companies (there may be omissions), these 20 Chinese companies are: BAIC, Beijing Automotive Research Institute, BAIC Foton, Brilliance Auto, FAW Group, Hangsheng Electronics, Taiwan HTC, Huizhou Huayang, Haowen, Hirain Technology, Neusoft, Ningbo Kabei, Pan Asia Automotive, Taiwan Pegatron, Shenyang Eastcom, Shenzhen Pengyi Industrial, Taiwan Sunplus Innovation, Taiwan Sunplus Technology, Truly Optoelectronics, Truly Semiconductor. Although IEEE has set the standard, the automotive Ethernet ECU test standard is set by OPEN Alliance. This covers all influential companies in the industry.

The physical layer standard is further divided into three layers, namely PCS, PMA and PMD.



PMA is more critical and has more changes. The PCS sublayer is responsible for 3B2T (Gigabit Automotive Ethernet) encoding, which can convert 8-bit parallel data received from the GMII port into 10-bit parallel data output. Because 10-bit data can effectively reduce the DC component and reduce the bit error rate, and the use of 3B2T encoding facilitates the extraction of clocks in the data and the first synchronization. The

two ends of the PCS can be regarded as the GMII interface and the TBI interface. The PMA sublayer further transmits the encoding results of the PCS sublayer to various physical media, mainly responsible for completing the serial-to-parallel conversion. The PCS layer transmits 10-bit code in parallel at a rate of 125M to the PMA layer, which is converted by the PMA layer into a 1.25Gbps serial data stream for transmission, so that the actual Gigabit Ethernet transmission rate of 1Gbps can be obtained. The two ends of the PMA sublayer can be regarded as the TBI interface and the SGMII interface respectively. The PMD sublayer will complete the interface for various actual physical media and complete the real physical connection.

Because it is an unshielded single-pair cable, there is a concern about RF leakage, so the US FCC will inspect such products. Testing of in-vehicle Ethernet is a big problem. It was not until February 2021 that Tyco was the first to launch an in-vehicle Ethernet test system that meets the IEEE 802.3ch MultiGBASE-T1 specification.

Full-duplex communication and PAM3 signaling add complexity to verifying ECUs under real-world conditions. Most serial standards work in simplex mode, with only one device communicating at a time. Some communication standards use a separate link for sending and receiving, while in automotive Ethernet, the master and slave devices can communicate simultaneously over the same link.

Therefore, the signal from the master is superimposed on the signal from the slave. The master knows which data it is sending, and it can determine the signal from the slave from the superimposed signal, and vice versa. Although the transceiver is designed to handle this situation, it is almost impossible to isolate the signal on an oscilloscope and perform signal integrity testing or protocol decoding.

Testing requires expensive hardware and software, and a strong learning ability. Almost no one is familiar with this field. This requires frequent testing and adjustment in the late stage of development, and it is almost impossible to rent a set of equipment for testing. Hardware is expensive, such as a high-bandwidth oscilloscope is very expensive. A 13G oscilloscope costs about 2 million RMB. Hardware engineers must have one, and it is not enough to just buy one. The price of a 110G oscilloscope is one million US dollars. Spectrum analyzers, RF signal sources, and network analyzers all cost millions of yuan once they have G bandwidth.



The main test contents are shown in the table above. Physical layer chips are a field with very high technical barriers. Only NXP, Broadcom, Marvell, Realtek, Microchip, and Texas Instruments can complete them in the world because physical layer chips are analog fields. When the physical layer chip sends data, it receives data from MAC (for the physical layer chip, there is no concept of frames. For it, all data is data regardless of the address, data, or CRC). It adds 1 bit of error detection code for every 4 bits, and then converts the parallel data into serial stream data, and then encodes the data according to the encoding rules of the physical layer (NRZ encoding of 10Based-T or Manchester encoding of 100Based-T), and then converts it into an analog signal to send the data out.