Domestic SoC vision algorithm adaptation has no unified industry standard for pricing, but evaluations generally revolve around a few core dimensions. Many clients ask the same question on first contact: "How exactly is this priced?" When you break it down, the logic isn't that complicated.
The most common model is overall project-based quoting. This works well when requirements are clear and scenarios are fixed, such as developing a face recognition access control solution on a Rockchip or Horizon chip. The service provider evaluates the complexity of algorithm migration, the workload for optimizing the original model, the hardware interfaces that need to be adapted, plus the time for joint debugging and testing, and packages all of that into a lump-sum price. The advantage is that the client knows exactly where they stand; the downside is that if requirements change frequently later on, it can lead to extra costs.
Another common approach is per-person-day billing. This suits projects with vague requirement boundaries and long-term iteration needs, such as repeatedly porting algorithms across multiple domestic platforms, or continuously optimizing detection rates under different lighting conditions. Usually, a daily rate is quoted based on the engineer's level (junior, mid-level, or senior), and the total cost is determined by the actual person-days invested. This model offers more flexibility for the client, but it also requires the team to have strong project management capabilities to avoid delays.
There is also a royalty or licensing model. If the hardware product is going to be mass-produced after the algorithm adaptation is complete, the service provider may charge a one-time adaptation development fee plus a licensing fee calculated by shipment volume. For example, a few yuan per device. This model lowers the upfront threshold, but future costs are tied to sales volume, making it suitable for clients who are confident about market expectations.
In addition, for some standardized algorithms, such as common object detection or pose estimation, some service providers have already optimized them in advance on mainstream domestic SoCs. In such cases, the fee is relatively low, perhaps just a simple porting and adaptation charge. But if the chip is very obscure or the vision requirement is extremely specialized, the upfront research and debugging costs will increase significantly, and the quote will naturally be higher.
Overall, the pricing for domestic SoC vision algorithm adaptation mainly depends on the complexity of the algorithm itself, the maturity of the chip platform, and your requirements for delivery time. When communicating your needs, try to clearly specify key information such as the application scenario, operating frame rate, power consumption limits, and mass production expectations. This way, the service provider can give a more reasonable quote and avoid later disputes. After all, when it comes to adaptation, saving time is often more important than saving costs.