High-Level ADAS Hits “Storage Bottleneck”: Who Seizes the New Opportunity to Break Through?

On one hand, with the rapid evolution from CNN to BEV+Transformer and then to the “VLA world model”, the number of parameters in AI large models has surged, which not only leads to a sharp increase in demand for SoC computing power but also a dramatic growth in storage bandwidth requirements. Traditional storage products are no longer sufficient to meet the needs of high-level ADAS and the integration of AI large models into vehicles. On the other hand, along with the explosive growth in storage demand driven by automotive intelligence and factors such as the shortage of upstream storage chips, the demand for some in-vehicle storage chips far exceeds the supply, putting many intelligent driving enterprises in a crisis of storage chip shortage. It is reported that since the second half of last year, the 8GB storage chips supporting intelligent driving basic chips have entered a “shortage wave”.

Currently, a new round of larger-scale high-level ADAS battles has begun, and urban NOA with higher thresholds and better user experience has become the real breakthrough point in the competition of ADAS. Data shows that from January to May 2025, the number of new passenger vehicles delivered in the Chinese market (excluding imports and exports) with pre-installed urban NOA (including software options) reached 902,200, a year-on-year increase of 152.5%.

Faced with the dual pressures of “soaring storage demand” brought by the popularization of high-level ADAS such as urban NOA, and the “performance shortcomings” and “supply gaps” of storage chips, how can in-vehicle storage break through the dual constraints of technology and production capacity? Who is striving to break through the bottleneck of in-vehicle storage?

01 AI Large Models in Vehicles Encounter “Stumbling Blocks”

When BEV+Transformer large models become the standard configuration for high-level ADAS and VLA end-to-end large models penetrate into more complex scenarios, the bandwidth, latency, and power consumption of traditional storage systems start to be “inadequate” and even become the “invisible threshold” restricting the implementation efficiency of AI large models.

“For AI large models to be integrated into vehicles, the real bottleneck is not the SoC computing power itself, but the storage and processing of data,” said an industry insider. To ensure the smooth operation of ADAS based on the Transformer architecture, the storage bandwidth needs to reach 400GB/s or even 600GB/s or more. For VLA models with larger parameters, some bandwidth requirements have even exceeded 1TB/s.

Another industry insider pointed out that 90% of the power consumption and delay in AI computing come from the storage system, and the performance of the storage system directly determines the rate of data transfer. “Even with the current best LPDDR5/GDDR6 automotive-grade storage chips, it is still impossible to achieve the smooth operation of AI large models at the 10-billion-parameter level.”

Take Li Auto as an example. The VLM model deployed on Orin X has a parameter scale of about 2B, but the VL (Vision + Language) part of the VLA model has 32B parameters, which is more than ten times that of the VLM model. Correspondingly, the mainstream automotive-grade LPDDR5 in the current market has a bandwidth of about 84GB/s, while the GDDR6 used in Tesla HW4.0 has a single-chip bandwidth of about 56Gb/s, supporting a maximum of 12 32-bit GDDR6 in parallel, with a total bus width of 384 bits.

In addition, the input of AI large models is no longer limited to text but has expanded to multi-modal, multi-granular unstructured data such as images, audio, video, and code. Such data is not only large in volume but also complex in format and structure, with diverse access modes, making traditional storage solutions difficult to handle.

Overall, the mainstream in-vehicle storage products in the current market can no longer keep up with the iteration pace of AI large models. As the number of parameters in AI large models grows rapidly and data increments increase geometrically, the “transmission efficiency” and “processing capacity” of storage systems have become key variables restricting the advancement of high-level ADAS.

Duan Zhifei, Senior Director of XPeng Motors’ Embedded Platform & Chief Architect of Vehicle Electronics and Electrical Systems, said in a public report that under the general trend of AI + automobiles, both intelligent cockpits and intelligent driving have put forward new demands for in-vehicle storage. On one hand, the realization of AI cockpit applications and ecological functions requires large-capacity, high-bandwidth, and high-reliability storage; on the other hand, the storage, management, and use of XPeng’s Turing AI intelligent driving end-to-end large models and data require high read-write speed and large-capacity storage.

Currently, XPeng Motors is developing a super-large-scale autonomous driving model with 72 billion parameters, namely the “XPeng World Base Model”, whose parameter count is about 35 times that of mainstream VLA models. To this end, XPeng Motors has independently developed an underlying data infrastructure (Data Infra), which has increased the data upload scale by 22 times and the data bandwidth during training by 15 times. At the same time, through the joint optimization of GPU/CPU and network I/O, the model training speed has been increased by 5 times.

It can be seen that with the advancement of advanced ADAS, as well as the application of new technologies such as AI intelligent cockpit systems and V2X, the data volume of intelligent vehicles has shown an explosive growth trend. Data shows that currently, the flash memory capacity of an intelligent connected vehicle ranges from 64GB to 256GB. With the upgrading of end-side large models, entertainment systems, and sensor accuracy, the storage capacity of intelligent vehicles will be upgraded to the TB level.

In this context, the market demand for higher-performance, larger-capacity, and lower-cost in-vehicle storage products is increasing, and the requirements for storage capacity, bandwidth, latency, and other performance indicators are rising sharply. Especially against the backdrop of the accelerated integration of AI large models into vehicles, in-vehicle storage chips have become a key support.

02 Who is Striving to Seize New Opportunities?

With the continuous evolution and upgrading of AI + automobiles, the in-vehicle storage chip market has spawned huge innovation and market opportunities. It is predicted that by 2030, automotive-grade storage products will account for more than 15% of the hardware cost of the vehicle electronic architecture, nearly doubling compared with 2020.

As we all know, due to the high technical threshold and long certification cycle of in-vehicle storage chips, overseas storage giants such as Micron, Samsung, and SK Hynix have long occupied a monopoly position in China’s in-vehicle storage market. However, in recent years, with the gradual development of the “new four modernizations” of automobiles, the in-vehicle storage market has entered a critical period of comprehensive upgrading, which also provides a rare opportunity for emerging enterprises to enter the in-vehicle storage field and reshapes the pattern of the in-vehicle storage industry.

In the past few years, more than 20 enterprises in the Chinese market have continued to 发力 the in-vehicle storage market, and manufacturers such as GigaDevice, CXMT, and Yangtze Memory Technologies Co. (YMTC) have achieved domestic breakthroughs. For example, in the NOR Flash field, GigaDevice’s GD25/55 full series of SPI NOR Flash have passed ISO 26262 ASIL D certification and AEC-Q100 certification, with global cumulative shipments exceeding 200 million units.

Specifically, in-vehicle storage is mainly divided into three types:

• DRAM: It is mainly responsible for internal memory that directly exchanges data with the CPU. Currently, the mainstream in the market is LPDDR4, LPDDR5, or GDDR6, etc.
• NAND Flash for data storage: The mainstream in the market is eMMC and UFS. High-level ADAS vehicles have adopted UFS3.1 storage as the mainstream choice, and will gradually iterate to UFS4.0, UFS5.0, and PCIe SSD. Currently, SanDisk, Kioxia, etc. have launched automotive-grade storage products complying with the UFS4.1 standard, while international giants such as Micron and Samsung have also launched automotive-grade SSD products.
• NOR Flash for code storage: It is mainly used in devices with high requirements for boot speed, such as display systems and ADAS systems.

Among them, GigaDevice’s GD55LX02GE series of SPI NOR Flash have been widely used in the in-vehicle field. It is understood that this 1.8V 2Gb SPI NOR Flash has a maximum clock frequency of 200MHz, a data throughput rate of up to 400MB/s, an 8-channel DTR SPI interface, compatible with single-channel and 8-channel SPI instruction sets, and supports XIP (Execute-In-Place) for efficient code data reading.

In addition, GigaDevice’s SPI NOR Flash has a built-in ECC algorithm and CRC check function, which improves reliability while extending product service life; on the basis of being compatible with existing SPI interface specifications and operation methods, it provides guarantees for high-speed system design through DQS and DLP functions; it supports standard TFBGA24ball packaging and can support a maximum temperature specification of 125℃. It is reported that this series of products from GigaDevice are suitable for automotive applications with strict requirements for high performance, such as intelligent networking, intelligent cockpits, and autonomous driving.

It is understood that NVIDIA and more and more automakers have chosen LPDDR5X as the core, while Micron Technology, SK Hynix, Samsung, and other companies are also vigorously promoting the large-scale integration of next-generation GDDR7 and HBM2E into vehicles. For example, SK Hynix has launched industry-leading GDDR7 with a data transmission rate of 32-40Gbps and a data processing speed of up to 1.5TB/S. At the same time, SK Hynix has cooperated with NVIDIA, Tesla, and other enterprises to expand the application of HBM from AI data centers to the intelligent automotive market. Currently, SK Hynix’s HBM2E has been applied in Waymo’s L4-level Robotaxi, with a capacity of up to 8GB, a transmission speed of 3.2Gbps, and a bandwidth of up to 410GB/s.

Overall, with the continuous improvement of the intelligence level of automobiles, in-vehicle storage has entered an important product upgrading period. DRAM is gradually evolving from LPDDR4/4X to LPDDR5/5X, LPDDR6, and even GDDR7 and HBM2E, and flash memory is also upgrading from eMMC to UFS and continuously developing towards SSD. At the same time, the pace of domestic substitution is accelerating, and domestic storage enterprises such as CXMT, YMTC, and GigaDevice are gradually expanding their market share.

However, it is undeniable that the demand for the intelligent automotive market is volatile, with great uncertainty in the demand quantity and types of storage chips. Coupled with the risk of insufficient supply brought by the cyclical changes in the industry, domestic players still face many challenges.