Today we share a research report titled "Research Report on Computing Power Measurement for AI Model Inference in Computing Networks" jointly compiled by China Unicom Research Institute, China Information Technology Designing & Consulting Institute, and China Unicom Digital Technology Co., Ltd. The report systematically proposes theoretical and methodological frameworks for measuring computing power in AI model inference, providing technical support for performance evaluation, intelligent scheduling, and billing in computing networks.

The report focuses on AI model inference services in computing networks and innovatively proposes a dual-dimensional measurement model of "computing power consumption" and "computing power usage," including resource-perspective computing power consumption measurement covering three-tier indicators: model inference services, computing network nodes, and computing network resources; and user-perspective computing power usage measurement covering two categories of indicators: model inference usage and computing power usage units.

The report also systematically organizes key technologies such as model profiling, parallel inference, and basic operation counting, and conducts empirical analysis using typical models like ResNet50 and DeepSeek R1 for inference, validating the feasibility and accuracy of the measurement methods.

Partial content: Due to space limitations, only partial content is displayed. For the complete file, please visit Knowledge Island to scan the QR code and join [XG Cloud Intelligence Knowledge Island] to download the latest 5G, 6G, and digital transformation materials anytime! Materials have been uploaded with daily continuous updates for anytime, anywhere access.