Using the LM Studio software, we tested the performance of the Deepseek-R1 LLM on the SER9 Pro HX370 and the SER9 Pro AI 365 across different parameter scales and quantization levels to provide a comprehensive performance reference.
Test Overview:
This test was conducted on two PCs: the SER9 Pro HX370 (with 64GB of RAM) and the SER9 Pro AI 365 (with 32GB of RAM). The Deepseek-R1 distilled model was used, with parameter scales ranging from 1.5B to 14B and quantization levels from Q2 to Q8.
The test results showed that both PC performed excellently when deploying LLMs locally. The SER9 Pro is not only able to utilize the CPU for computation but also prioritizes the use of integrated GPU.
Performance Comparison:
Performance in GPU Mode:
-When processing the 1.5B and 7B models, the SER9 Pro HX370 demonstrated superior performance due to its more powerful Radeon 890M iGPU.
-For the 14B model, the performance difference between the two PCs was minimal, with both running stably.
-GPU mode offers advantages in terms of VRAM usage and response speed, making it suitable for scenarios with higher performance demands.
Performance in CPU Mode:
In CPU mode, the performance of both PCs when running LLMs was similar.
Memory and VRAM Configuration:
-The SER9 Pro with 64GB of RAM can allocate up to 48GB as dedicated VRAM, while the 32GB version supports up to 24GB of dedicated VRAM.
-This design allows running larger parameter-scale models on the GPU.
Recommendations:
-Daily Use: The 8B model is the best choice, offering a perfect balance between performance and hardware usage, suitable for most scenarios.
-High-Performance Needs: The 14B model is capable of handling more complex tasks, ideal for scenarios requiring higher performance.
