DIGITIMES Research believes that the neural processing unit (NPU) will bring new choices for AI accelerator chips in edge devices. Compared to general-purpose graphics processing units (GPGPU), NPU has power consumption and price/performance ratio advantage, making it suitable to handle AI computing tasks in edge computing, according to DIGITIMES Research's findings in the new AIoT report.
With the emergence of edge computing, tasks are shifting towards decentralized edge servers and terminal devices to meet demands for low latency, energy efficiency, and information security across diverse scenarios. This trend anticipates the introduction of more specialized edge AI accelerator chips into the market, with NPUs expected to be a key direction. NPUs offer high price/performance ratios due to their low power consumption and enhanced computing capabilities, enabling flexible adjustments in different edge computing environments.
However, due to the high specialization of NPU, it can only execute AI algorithms under specific frameworks, and the software development environment is relatively closed, posing a significant challenge to the development of the NPU industry, the report's study found.
The focus of short-term development of related NPU developers, such as Kneron and Hailo, has been aligning with machine vision. In addition to providing hardware, they also offer downstream customers software solutions or software development kits (SDK).
Despite the advantages of NPU in power consumption and computational efficiency, market entry still depends on chip costs and market size in specific areas. Major industrial players focus on edge machine vision and actively collaborate with industrial PC (IPC) and IP camera makers to develop comprehensive solutions.
NPU is a computing component specifically designed for AI applications. Compared to GPGPU, which processes tasks using many parallel computing units, NPU emphasizes the computational principles of synaptic weights, aiming at simulating the message transmission mechanism in human neurons to allow the chip to complete specific tasks efficiently. NPUs are designed for more customized AI computing tasks.
