重磅！华为刚刚发布AI新技术！

Core Viewpoint - Huawei has launched Flex:ai AI container software, which utilizes computing power slicing technology to divide a single GPU/NPU power card into multiple virtual computing units, enhancing resource utilization significantly [1][3]. Group 1: Product Features - Flex:ai allows a single power card to support multiple AI workloads simultaneously, improving hardware resource utilization [3]. - The software aggregates idle XPU computing power from various nodes within a cluster to form a unified "shared computing pool," enabling global scheduling and flexible allocation of computing resources [4]. - The core technology integrates both hardware and software, enhancing the typical utilization rate of GPU/NPU from 30%-40% to 70%, embodying the concept of "software compensating for hardware" [4]. Group 2: Technical Innovations - Flex:ai deeply integrates Huawei's self-developed Ascend AI processors, optimizing performance and power consumption through collaborative design [4]. - In large model training scenarios, Flex:ai innovatively manages and schedules various heterogeneous computing resources, including NVIDIA GPUs and Ascend NPUs, effectively addressing the efficiency bottleneck in current large model training [4]. Group 3: Community Engagement - Flex:ai will be open-sourced in the Magic Engine community to promote technology sharing and ecosystem development [3].