Samsung BM1773: Ultra-High-Density Storage Solution for AI Data Centers

Recently, the AI industry has been rapidly evolving beyond simple command-and-answer levels toward an agentic AI environment, where multiple AI agents interact organically to maintain complex contexts and generate step-by-step outputs. In this process, the demand for high-capacity storage to efficiently store and manage the vast amounts of context that AI must reference is increasing, requiring a fundamental structural change in data center infrastructure.

The evolving requirements of AI workloads

As the parameter scale of large language models (LLMs) expands and inference services become more complex, the amount of data to be processed is increasing exponentially. Furthermore, the introduction of advanced services such as retrieval-augmented generation (RAG) has drastically improved the accuracy of AI responses, but has simultaneously increased the data processing burden of searching and referencing vast amounts of external data in real time.

This creates demanding I/O requirements for low-latency and high-frequency random access, yet the physical space and power resources of data centers to accommodate this are limited. Consequently, customers require ultra-high-density, high-performance, and high-reliability SSDs that secure maximum capacity within limited server slots without performance degradation, and it is also essential to secure a security architecture that guarantees thorough hardware-level protection.

Unveiling the features of BM1773

Samsung's BM1773 is the world's first product to implement a storage capacity of up to 245.76TB in the E3.S form factor, maximizing space utilization efficiency inside the server. By securing capacity with a single or a few drives—which previously required the parallel installation of numerous low-capacity SSDs or HDDs—physical complexity is drastically reduced inside the server and cabling and power connection structures are simplified, thereby increasing infrastructure management efficiency.

In particular, BM1773 serves as a high-performance, ultra-high-capacity cache layer that resolves bottlenecks in existing HDD-based object storage, enabling rapid access to massive multimodal data, which is the core of AI workloads. Through this, BM1773 implements both the high-capacity characteristics of HDDs and the high-speed access performance of SSDs, proving exceptional efficiency in object data caching.

With its high-capacity storage and data access capabilities, BM1773 efficiently accommodates long-lived cold KV cache data generated during LLM inference and enables fast access to data needed for AI workloads. This helps support scalable storage infrastructure as AI workloads continue to grow.

Looking at the technical specifications, Samsung's BM1773 is equipped with 9th-generation high-density NAND flash (V9Q 2Tb) and the latest high-performance controller to deliver optimal performance even in data-intensive workloads.

• Sequential write performance: Through a maximum sequential write performance of 4.5GB/s, it quickly stores large-scale context data and rapidly secures GPU memory space, thereby improving overall AI response speed.


• Random write performance: It provides a maximum random write performance of 50K IOPS, minimizing system bottlenecks that occur when frequently updating and replacing different large-scale data in multi-tenant environments where multiple users send requests simultaneously.


• Interface and architecture: By applying a 16-channel PCIe 5.0 architecture to maximize data access speed, it supports the acceleration of AI applications and the improvement of overall system efficiency.


• Security grade: By supporting the Commercial National Security Algorithm (CNSA 2.0) and undergoing FIPS Sub-chip certification, a recognized safety certification, it meets the highest security requirements.

Reliability at scale

Samsung's BM1773 has drastically improved durability, which has been a commonly known vulnerability of high-density QLC SSDs. It supports a durability rating of 0.6 DWPD (drive writes per day). This is a twofold increase in TBW (total bytes written) compared to the previous generation of QLC products, ensuring consistent data reliability and long-term operational stability in AI and enterprise storage environments where large amounts of data are continuously generated.

Additionally, by applying die failure recovery technology, it resolves the existing risk where the entire SSD switched to read-only mode upon a single die failure, maintaining normal operation even when a failure occurs, thereby reducing the risk of service interruption and maximizing availability.

Setting a new standard for QLC SSDs

There was a technical prejudice that high-capacity storage devices like QLC SSDs have large capacities but low speed and durability, but Samsung's BM1773 has overcome these traditional limitations, simultaneously achieving the core values of capacity and performance.

In the future, the scale of data processing will further expand due to the continuous growth of LLMs and the popularization of RAG technology. Accordingly, high-capacity QLC SSDs will play a key role in rapidly processing vast vector databases while supplementing the capacity limits of the TLC layer to efficiently accommodate large-scale contexts. As a result, high-capacity QLC SSDs are expected to evolve beyond simple storage to become standard components of AI servers and intelligent memory layers, and Samsung's BM1773 will be a core solution leading this change.