What are the types and functions of car memory?

Automotive storage is very important. Tesla Model S used non-automotive grade storage, namely SK hynix's H26M42003GMRA, which is an old product from 2014. Frequent OTA and high temperature caused it to fail, eventually resulting in frequent freezes and black screens. This kind of failure cannot be solved by OTA. The car must be disassembled, the chip removed, and replaced with an automotive grade chip. Needless to say, this solution is extremely costly.

Every SoC and high-end MCU in a car needs memory. The largest usage is in cockpit SoC and smart driving SoC, followed by T-Box and instrument, and then gateway and chassis. Memory is mainly divided into three categories:

One is DRAM, which is used for data and code cache. The mainstream ones are LPDDR4 and LPDDR5.

The second category is storage data, the mainstream ones are eMMC and UFS;

The third category is for storing code, the mainstream of which is NOR Flash.

SAIC Feifan R7 Intelligent Driving Domain Controller

We take the SAIC Feifan R7 intelligent driving domain controller in the above picture as an example to study automotive storage. The large chip in the middle is NVIDIA Orin-X, and there are 4 Micron LPDDR5 memories on the periphery, model MT62F1G64D8EK-031 AAT:B, capacity 8GB, speed 6400MB/s, x64 bit width, which is Micron's new product. A Micron eMMC, codenamed JWD65, model MTFC32GASAQHD-AAT, capacity 32GB, clock frequency 200MHz, relatively slow speed, should mainly store boot root directory files. A Micron UFS, codenamed HSA06, model MTFC256GAVATTC-AAT, is UFS3.1 version, capacity 256GB, should mainly store intelligent driving perception models and decision models and related files, this capacity is twice that of the ideal dual Orin autonomous driving controller. There is a serial flash for booting on the right side of JWD65, model MX25U51279, capacity 512Mb, provided by Macronix in Taiwan, China. There is also a serial flash on the other side, model MX78U64A00FXDR02, capacity 512Mb, also provided by Macronix.

Ideal L8/L9 cockpit domain controller SA8155P module

The above picture shows the cockpit domain controller SA8155P module of Ideal L8/L9, which uses two LPDDR4 chips. The LPDDR4 chips are provided by Micron, model MT53E1536M32D4DT-046 AIT:A, capacity is 6GB, rate is 4266Mbps. UFS is also supplied by Micron, just below SA8155P, codenamed HSA12, model MTFC128GAZAOTD-AAT, capacity is 128GB, standard is UFS2.1.

NIO ET7's SA8155P module

The SA8155P module of NIO ET7 uses two Micron D9XKN chips. D9XKN is Micron's LPDDR4 chip, the official model is MT53E2G32D4DT-046, the capacity is 64Gb, that is, 8GB. NIO's UFS comes from Samsung, the model is KLUEGAJ1ZD-C0CQ, the capacity is 256GB, the version is UFS2.1, the voltage is 1.8/3.3 volts, and the interface is G3 2Lane. There is also a SA8155P Boot Loader serial NOR Flash chip on the back of the PCB board, provided by Micron, the chip identification is RW199, the model is MT25QL128ABA8ESF-0AAT, the capacity is 128Mb, the voltage is 2.7-3.6 volts, and the temperature range is -40℃ to 105℃.

Nissan's latest cockpit chip

The picture above is Nissan's latest cockpit chip. Japanese cars have always been known for their backward electronic architecture. Japanese cars lag behind domestic cars by about 10-20 years. One of the reasons is that the supply chain of Japanese cars is very tight and there is no motivation to upgrade. The second is to save costs. Nissan's cockpit uses extremely rare DDR3 memory, which is the standard in 2002. It has been more than 20 years since then, and basically no manufacturer produces it. But Japanese cars are still using it. The DDR3 memory is supplied by Etron in Taiwan, China. Etron is a rare memory design company that designs its own memory and then entrusts Winbond wafer factory to manufacture it. The model is EM6HE16EWAKG-10H, with a capacity of 4Gb per chip, which is only 0.5GB. The estimated manufacturing process is 350 nanometers, and now 3-nanometer chips are all in mass production. At the same time, Nissan also uses a synchronous SDRAM chip, which is also extremely old and is still provided by Etron. The capacity is only 16Mb and the volume is huge. It may be a 1000-nanometer manufacturing process and has a history of more than 20 years. The eMMC uses a domestically made Longsys with a capacity of 8GB. On the back of the PCB board, there is another storage, which is the Bootloader Flash provided by Taiwan Jinghao Technology Company, model EM29LV320A, with a capacity of only 48Mb and a large size.

Micron has monopolized the high-end market. 90% of the DRAM market in the automotive field with a capacity of more than 1GB is in Micron's pocket. Firstly, this market is relatively small. The market size in the PC field is more than 10 times that of Samsung and SK Hynix. Secondly, SoC companies such as Micron, Qualcomm, Nvidia and even Renesas provide customized services, and these SoCs specify the use of Micron's storage. Only Korean cars use a small amount of Samsung's large-capacity DRAM.

Each time a calculation is performed, the CPU issues an instruction, and the weight model is taken out from the UFS and temporarily stored in the DRAM. If there is video memory, it is placed in the video memory. Usually, the bandwidth of video memory is much higher than that of shared DRAM, so there is no need to take it out from the UFS for each calculation. This is the significance of the existence of DRAM and video memory. Its speed is much faster than UFS.

Source: JEDEC

The standards for the storage industry are set by JEDEC, the Solid State Technology Association, the leading standards body for the microelectronics industry. For more than 50 years, JEDEC has developed standards that have been accepted and adopted by the entire industry. As a global organization, JEDEC's membership is multinational. JEDEC is not affiliated with any one country or government entity. JEDEC's standards-setting process brings together manufacturers and suppliers to develop standards through 50 committees and subcommittees to meet the diverse needs of industry development and technology. JEDEC has nearly 300 member companies, including almost all of the top 100 companies in the industry. In JEDEC's voting process, a company, regardless of its size and influence, has only one vote. Standards can only be adopted if a two-thirds majority of the committee votes in favor. All standards are ultimately approved by a vote of the Board of Directors; approval requires a 75% majority vote in favor.

However, you still need to spend a little money to see the JEDEC standard. For example, the LPDDR4X standard costs $106, and the LPDDR5X standard costs $459. Of course, this is a detailed standard of hundreds of pages, and ordinary people don’t need to know such a detailed standard.

Source: JEDEC

The latest mobile DRAM standard is LPDDR5X, which was released at the end of June 2021. Because LPDDR4/LPDDR4X does not support Bank Grouping, its price is higher than LPDDR5. At the same time, LPDDR4/LPDDR4X is dual-channel and has a higher bandwidth than LPDDR5. The difference between LPDDR4X and LPDDR4 lies in Vddq. Vddq is the voltage of the DRAM I/O buffer, which is basically combined with the core voltage VDD. The lower the voltage, the faster the speed and the lower the power consumption. In order to alleviate the embarrassment that the bandwidth of LPDDR5 is not as good as that of LPDDR4, LPDDR5X appeared. The biggest difference between it and LPDDR5 is that the speed is increased to 8533Mbps, catching up with LPDDR4 (LPDDR4 is dual-channel), and the theoretical maximum bandwidth is increased from 51.2GB/s to 68.26GB/s. At the same time, it is a variable voltage, with a minimum of 0.5 volts and a maximum of 1.1 volts, and power consumption is reduced by about 20%, which is quite suitable for mobile phones. Xiaomi 13 is also the world's first model to use LPDDR5X, but it doesn't matter for the automotive field, as the power consumption is negligible. At present, there are no automotive-grade LPDDR5X products available. In view of the limited improvement of LPDDR5X, some people have already taken the lead. SK Hynix has launched LPDDR5T with a maximum speed of 9600Mbps, which is 13% higher than LPDDR5. LPDDR5T is the future LPDDR6.

GDDR was originally used in graphics cards, but as AI continues to evolve and Transformers become popular in the automotive industry, the demand for storage power is growing. Tesla was the first to use GDDR6, but there are currently no automotive-grade GDDR6 products.

Source: JEDEC

GDDR6X is already available. Given the importance of GDDR, Samsung is unwilling to let Micron continue to lead. In July this year, Samsung announced that it had completed the internal development of GDDR7 and would be able to ship it in early 2024. The actual standard of GDDR7 has not yet been determined.

Source: JEDEC

Samsung also made a minor change to GDDR6, which increased the theoretical maximum bandwidth from 672GB/s to 768GB/s. There is a mistake in the table above, 256bit should be 384bit. GDDR7 reaches 1TB/s, approaching the expensive HBM.