The Evolution of Storage Technology Driven by Large Models

Disassembling servers in data center cabinets typically reveals several solid-state drives (SSDs). These SSDs bear the heavy responsibility of data storage, with their controller chips acting as the brain, efficiently and accurately managing data flow in and out of storage units.

Storage is one of the key infrastructures in the era of large models. As long as there is data, storage is required. When data becomes the core resource in the AI field, storage technology determines the efficiency of data processing for large models and affects training and inference speeds. With the exponential growth of training dataset sizes, balancing storage costs and performance is crucial.

“Storage is a definite large market,” said CFO of storage chip design company InnoGrit, Zhong Xiaohui. She noted that computing power, storage capacity, and transmission capability often promote and develop symbiotically. The current wave of large models is driving the evolution of storage technology, raising the bar for differentiated competition and technological iteration among storage chip companies and SSD manufacturers.

The Importance of Computing Power and Storage

The main hardware components of SSDs include NAND flash memory chips, DRAM cache, and controller chips. If we compare the data that needs to be stored to cars, then SSDs are like a giant parking lot, where storage units on flash chips are parking spaces, and the controller chip is the “administrator” of this giant parking lot, directing each vehicle to enter and exit its space accurately, quickly, and reasonably.

The controller chip can be seen as the brain of the SSD, executing complex operations such as data reading, writing, and encryption through corresponding firmware code. InnoGrit is a developer of these core storage components, with its main products including complete SSD solutions and internal storage controller chips.

Globally, the enterprise SSD market has long been dominated by South Korea’s Samsung Electronics and SK Hynix, which together hold over 70% market share. Domestic enterprises in the enterprise SSD supply chain are still in a rapid catch-up phase. When InnoGrit was founded in 2017, mainstream storage technology was shifting from mechanical hard drives to SSDs, and data transmission interfaces were transitioning from SATA to the faster PCIe, providing opportunities for domestic startups.

InnoGrit SSD controller chips and SSD module reference images for PCs and servers.

“From 2017 to 2021, domestic manufacturers focused on whether technology could be transformed into products. From 2021 to 2024, after solving product issues, the focus shifted to whether there were customers,” Zhong Xiaohui stated. The objective push for domestic replacements has accelerated the development of China’s storage chip industry, with the AI boom providing even greater support.

At the “Wukong Intelligent Computing” 6876P computing center in Lianyungang, Jiangsu, servers are lined up in cabinets performing 68.7 quintillion floating-point operations per second. After optimizing the DeepSeek full parameter version through hardware and software collaboration, it achieves an ultra-high throughput of over 6900 tokens per second, enabling enterprises to quickly launch AI applications in just three minutes.

In the current wave of large models, data volumes are increasing. Cold data is becoming scarce, with more data turning into warm or even hot data. Previously, data in financial systems or traditional data centers would be stored after five years of inactivity, but now the situation has changed. Once models are running, they need to process data in real-time, transforming previously cold and warm data into hot data.

Jiangsu Zhonghuan Cloud Control IoT Technology Co., Ltd. is developing a large model for sanitation services based on “Wukong Intelligent Computing,” exploring applications for intelligent agents. Sanitation workers equipped with smart wristbands can transmit their vital signs, locations, and task progress in real-time, allowing the system to automatically adjust work routes. Unmanned cleaning vehicles and drones share real-time data on road conditions and waste distribution, refreshing operational strategies at a second-level frequency. Through virtual-physical mapping, coordinated scheduling, and autonomous collaboration, traditional sanitation operations are transforming into a new intelligent agent model. “In the past, we referred to intelligent sanitation as information management; now we call it embodied intelligent agents. The difference lies in the system no longer just observing data but allowing every device, worker, and operational link to become a thinking, communicative, and self-evolving digital entity,” said Xu Lei, executive director of Zhonghuan Cloud Control.

Meanwhile, applications like DeepSeek have opened doors to inference and edge computing. Lightweight model design, hardware adaptation optimization, and reduced model deployment costs are shifting computing demands from the training side to the inference side, concentrating training tasks in the cloud while pushing inference tasks down to edge devices. The rising temperatures of massive data and the evolving demands for computing power are continuously upgrading people’s pursuit of inference experiences, striving for extreme low latency, which further raises the requirements for storage capacity.

Large Models Driving Storage Technology Upgrades

In the surface polishing industry, excellent craftsmanship is an insurmountable technical barrier, while AI’s value lies in continuously accumulating process data to develop smarter robotic brains, further optimizing processes.

Founded in 2018, Sophis Intelligent Technology (Shanghai) Co., Ltd. has transitioned from robot agency to product self-research, focusing its industrial robots on applications in polishing, cutting, drilling, and deburring in manufacturing. Founder Du Ling stated that only through AI can robots become smarter. The team has developed intelligent polishing machines that can display data on smartphones and computers, ensuring that employees stay away from harmful dust and noise while recording key parameters such as pressure, temperature, speed, and materials during the polishing process. The goal is to develop a polishing large model to meet the polishing needs of different products and enhance craftsmanship.

This underscores the urgent demand for computing power and storage. According to InnoGrit, the collection of raw data and inference logs generates a massive amount of data, requiring storage that supports high-volume writing and high-speed reading. Data cleaning and model training require storage that can handle high-concurrency mixed read and write operations, with even higher requirements for random performance. Different data application scenarios have begun to show differentiated requirements for storage chips.

The demand for SSDs in traditional data centers ranges from 4TB to 8TB, but with the emergence of DeepSeek, flash memory capacity requirements have risen to 32TB, 64TB, or even 128TB. The larger the flash memory chip capacity, the greater the development difficulty. This is akin to building a taller building, which requires higher structural standards. InnoGrit’s products of the same generation have already spawned various niche applications, raising the bar for differentiated competition and technological iteration among storage chip companies and SSD manufacturers.

In fact, AI is driving the evolution of storage technology. “In the past, many domestic data centers were still using mechanical hard drives. Over the past two years, due to speed requirements, they have switched to SSDs, moving from SATA to PCIe 4.0, and now entering the PCIe 5.0 era,” Zhong Xiaohui explained. After the launch of ChatGPT in 2022, the application market represented by AIGC began to impose higher performance and capacity demands on storage. The emergence of DeepSeek has facilitated the application of large model inference, and the new generation of PCIe 6.0 SSDs and storage-class memory solutions based on CXL interfaces are also gaining attention. These technologies will support large model data center cloud services and local deployment integrated machines in new ways, accelerating the landing of open-source large models like DeepSeek. “The introduction of SSDs into the market accelerated after AI emerged; we managed to onboard standard server manufacturers in about a year, and by the first half of 2024, shipments surged over tenfold.”

Computing power, storage capacity, and transmission capability often promote and develop symbiotically. Domestic AI chip companies are exploring more open architectures like RISC-V, aiming to establish a layout from edge servers to cloud servers. Meng Jianyi, CEO of Zhihe Computing, stated that breakthroughs in high-performance computing using RISC-V not only require entry into high-performance domains at the general computing level but also need to integrate AI-enhanced computing at the architectural level for AI-native implementations.

“Storage has always followed computing power and transmission capability. Whenever either end rises, you must keep up,” Zhong Xiaohui believes. It is essential to match differentiated storage solutions to various application scenarios to support computing power demands. This year, the team’s focus is on developing storage controller chips and solutions that meet future AI requirements. “There are still many manufacturers of global storage controllers; to carve out a niche in this market and establish a solid foothold, we must excel in the iterative upgrade process, offering what others do not have.” In the future, domestic manufacturers must not only focus on meeting domestic replacement needs and sustainable product iteration capabilities but also prioritize the ability to export products overseas, which will be a necessary phase for domestic storage companies in the next 3-5 years, or even 5-10 years.