见证历史！华为 重大发布

Core Insights - Huawei has officially launched a groundbreaking AI technology called Flex:ai, which can enhance the average utilization rate of computing resources by 30% in AI workloads where full card computing power is not utilized [1][3] - Flex:ai is a software for pooling and scheduling various types of processors (XPU) based on Kubernetes, and it is set to be open-sourced to the Moqing community, contributing to a complete ModelEngine open-source ecosystem [1][6] - The technology aims to facilitate the industrial application of AI, with Huawei's VP stating that it can unlock the potential of infrastructure and accelerate the democratization of AI [1][6] Group 1: Flex:ai Capabilities - Flex:ai presents three key capabilities: resource slicing, multi-level intelligent scheduling, and cross-node resource aggregation [3][5] - The technology allows for the slicing of a single GPU/NPU card into multiple virtual computing units, achieving a precision of 10%, enabling multiple AI workloads to run simultaneously on a single card [4][5] - Flex:ai's intelligent scheduling can automatically sense cluster loads and resource states, optimizing the allocation of virtualized GPU and NPU resources based on workload priority and requirements [5][6] Group 2: Competitive Advantages - Compared to Nvidia's Run:ai technology, Flex:ai has two unique advantages: virtualization and intelligent scheduling [5][7] - The "remote virtualization" feature allows for the aggregation of idle XPU resources across the cluster into a "shared computing pool," providing ample support for high-demand AI workloads [5][9] - Flex:ai can also ensure that high-priority AI workloads receive the necessary computing resources, even reallocating resources from lower-priority tasks when needed [5][6] Group 3: Ecosystem Development - The release and open-sourcing of Flex:ai will help Huawei build a complete ModelEngine open-source ecosystem, which includes previously released tools like Nexent, AppEngine, DataMate, and UCM [6][9] - The ModelEngine platform offers a one-stop service for data preprocessing, model training, optimization, and deployment, catering to the needs of large model training and inference [6][9] - The open-source nature of Flex:ai encourages collaboration among developers from academia and industry to establish standards for heterogeneous computing virtualization and AI application integration [9]