Recently, the popular DeepSeek has been a topic of research both within and outside the industry. After all, various sectors are eager to leverage AI for their services, and the field of image processing is no exception. Consequently, many have inquired whether RK3588 + DeepSeek is a viable option.
To answer this: relying solely on the RK3588 is unrealistic! However, it is possible to attach a high-performance computing module.
The primary role of DeepSeek is to serve as a front-end decision-making brain, assisting in intelligent decision-making for the data processed by the image processing board, which is one of the steps towards achieving full automation.
To achieve this goal, the required data is immense; therefore, the larger the parameter version of the DeepSeek model, the better. For instance, the full version of the DeepSeek model has 671 billion parameters, which, given the RK3588’s current performance of 6.0 TOPS, is clearly unfeasible.
Recently, an engineer successfully deployed DeepSeek on the RK3588, but it was only able to run the DeepSeek R1 (Qwen2 1.5B) model. What does 1.5B mean? It can only handle lightweight text tasks and cannot achieve rapid real-time responses. Besides the computing power issue, the model’s porting is also a challenge, as many network layers are not supported by the Rockchip platform.
However, this at least indicates that running DeepSeek on the RK3588 is feasible, albeit limited by its hardware.
If a high-performance computing module is attached to the RK3588 image processing board, theoretically, it could enable the application of high-parameter models of DeepSeek.
Currently, our entire team is conducting in-depth research in this direction, and we believe results will emerge soon.
●Is achieving 100Hz tracking difficult? Huishi custom RK3588 easily breaks through●Huishi Viztra-HE030 image processing board – Customized industrial-grade RK3588 board●Say no to illegal flights! Huishi RK3588 image processing board helps create no-fly zones in cities
Click, like, and follow us!
