Currently, the open-source LLM DeepSeek, with its powerful language understanding and generation capabilities, has become a popular. Whether it is for copy writing or coding, simply click "DeepSeek" and you can easily solve it. However, with the sharp increase in user visits, issues such as server unresponsiveness and long waiting times are also common incidents. Suddenly, DeepSeek's on-device deployment, which can run offline and is more private, has become a new trend.

Although with many advantages, local deployment has certain requirements for the hardware configuration of computers. LLMs contain a large number of parameters, and even distilled small models have tens or even hundreds of gigabytes in size. Besides requiring CPUs and GPUs to run efficiently, a high-performance SSD is also indispensable for computers. Union Memory AM541 is equipped with the new generation Jaguar6020 main controller, which has built-in high-capacity SRAM and IO acceleration modules. The sequential read speed can reach up to 7000 MB/s, which can easily handle high-load scenarios such as loading DeepSeek, providing users with a smooth experience and enabling users to "DeepSeek freely".

100% Compatibility: Smooth Initial Loading for LLMs

During the local loading and running process of DeepSeek, SSD is the first component of the entire data flow. After DeepSeek completes the local deployment, the model files are saved in the SSD. When the users loads the model, the system will first read the LLM file from the SSD into the system memory, and then transfer it from the memory to the GPU memory for inference and processing by the GPU. Therefore, the better the performance of the SSD, the faster the data can be transferred to the GPU for computing, which in practical application, is reflected by the shorter loading time of LLMs.

LLM loading flowchart

In order to verify the performance and scenario compatibility of Union Memory AM541, we deployed the Deepseek-R1 8B model using the Ollama model framework locally. We used Union Memory AM541 1TB SSD and a 1TB A product from a domestic competitor, along with a GeForce RTX 3060 Ti graphics card, and conducted model loading tests under the same conditions.

SSD state during LLM loading (as shown in the right image)

The test results show that the PC equipped with AM541 performs excellently when loading LLMs. The initial loading time (the fastest) is only 2.486 seconds, outperforming comparable products from leading domestic SSD manufacturers by approximately 9%. This achievement fully demonstrates that AM541 is 100% adaptable to high-load applications such as DeepSeek, and can provide users with a more smooth usage experience. 

Sequential reading speed hits 7GB/s, powerful performance enables users to "DeepSeek freely"

Shorter loading times reflect the SSD's superior performance and enhanced scenario adaptability. Thanks to the acceleration module built into the SoC, the nominal sequential reading/writing speeds of AM541 reach 7000 MB/s and 5600 MB/s, respectively, while the 4KB random reading/writing speeds can reach 800K IOPS and 800K IOPS, respectively. From the pattern analysis during the DeepSeek loading process, the performance peak of AM541 is close to 7GB/s, which is highly aligned with its nominal performance.

Why is AM541 faster? The testers further conducted trace analysis on the loading process and discovered that the model loading process mainly consists of large-size low-QD sequential reads. However, the built-in Big SRAM strategy and latency control mechanism of AM541 had significant advantages in handling such commands, resulting in a substantial reduction in the loading time for LLMs, outperforming that of the competitors.

In addition, after testing, it was found that when the LLM completed its initial loading on the computer equipped with AM541, the model file was immediately cached in DRAM. Therefore, when reloading after device Idle, the model file could directly be transferred from DRAM to VRAM, resulting in a loading time that was faster than the initial loading. This truly turned DeepSeek into a "private tool" for users, allowing them to "DeepSeek Freely" at any time.

Embracing LLMs, Union Memory continues to provide support for the AI ecosystem

With the acceleration of the trend of AI localization, terminal devices such as DeepSeek AIO are gradually becoming more popular. In the field of consumer electronics, many mainstream PC manufacturers have already embedded DeepSeek into AIPC, and many mobile phones have also begun to focus on AI LLMs. In the future, personal computers, mobile phones, and other terminal devices are very likely to be equipped with multiple LLMs, and various industrial applications will also be connected to LLMs. Facing AI models of different parameter sizes and the frequent high-frequency reading/writing processes caused by the frequent switching among different models, SSDs not only need to have sufficient capacity, but also it poses the challenges to the comprehensive performance and stability of SSDs.

Relying on hardware acceleration and advanced software algorithms, AM541 not only has an advantage in low QD Latency, but also performs excellently in various mixed reading/writing operations, and can easily cope with multiple application scenarios.

In the PCMark10 test, the AM541 running score exceeded 3700 points, outperforming the competitor's similar product A by approximately 300 points. It outperformed the competitors in multiple scenarios such as office work, gaming, and content creation. Among them, in the creation of common digital contents, AM541 performs on average 10.5% faster than the competitor's similar products.

The AI wave is surging continuously, and technological innovation is advancing with each passing day. As the underlying hardware support, SSDs play a crucial role in the process of AI localization. Union Memory AM541, with its robust performance, not only provides an efficient solution for the local deployment of DeepSeek, but also brings users an unprecedented usage experience. In the future, Union Memory will keep up with industry trends, focus on the technical challenges in end-user LLMs, promote technological innovation and product iteration, and provide sustained momentum for the prosperous development of the AI ecosystem.