Core Viewpoint - The article discusses the transition of a major company from NPU to GPGPU, emphasizing the evolution of NVIDIA's GPU architecture and the strategic decisions made by domestic companies in response to industry challenges [1][2]. Group 1: GPU and NPU Architecture - NVIDIA's GPU development has shown a clear cycle, evolving from fixed pipeline DSA architecture to unified Shader architecture, and now to Tensor Core for AI, maintaining its industry position through continuous optimization of the CUDA ecosystem [1]. - NPU is designed specifically for AI computations, offering advantages in energy efficiency and speed compared to traditional CPUs and GPUs, making it suitable for mobile, edge computing, and embedded AI scenarios [3]. - The complexity of general-purpose CPUs is significantly higher than that of GPUs and NPUs, with NPU design being simpler, focusing mainly on matrix multiplication and convolution operations [4]. Group 2: Software and Ecosystem Challenges - The software complexity of NPU exceeds that of its hardware, making it crucial to evaluate software usability rather than just comparing computational power [5]. - NPU's multi-level memory architecture presents challenges, such as limited L1 cache size and storage conflicts, requiring precise data segmentation to maximize performance [5]. - The fragmented ecosystem of NPU poses a barrier to optimization, as software developed for one NPU may not be easily transferable to another, increasing application deployment costs [5]. Group 3: Evolution and Future Directions - The evolution of GPUs from simple dedicated calculators to complex systems with independent control units highlights the need for NPUs to develop similar capabilities to handle the increasing complexity of AI tasks [6][7]. - The shift in AI tasks from inference to a combination of training and inference necessitates a move towards architectures that support efficient computation and flexible control [7]. - The rise of NPU is seen as a natural progression in AI computing, with a trend towards integrating SIMT front-end capabilities to enhance control units, aligning more closely with GPU architectures [7].