Core Insights - Huawei is set to hold the 2025 AI Container Application Implementation and Development Forum next week, where it will unveil breakthrough technologies in the AI field aimed at improving computing resource utilization efficiency [1] - The new technology is expected to enhance the utilization rate of computing resources such as GPU and NPU from the industry average of 30%-40% to 70%, significantly unlocking the potential of computing hardware [1] - Huawei's technology aims to compete with Nvidia's acquisition of Israeli company Run:ai, focusing on software innovation to unify resource management across different computing platforms [1] Group 1 - Huawei's new technology follows the innovative approach of "software compensating for hardware," which is designed to optimize the utilization of computing resources [1] - The technology will provide more efficient resource support for AI training and inference by "masking" hardware differences among various computing resources [1] - Nvidia announced the acquisition of Run:ai for $700 million, which focuses on optimizing GPU resource utilization through a software platform built on Kubernetes [1] Group 2 - In the context of limitations in advanced processes and performance gaps in single-chip computing compared to foreign counterparts, Huawei is focusing on software innovation to compensate for deficiencies in chip technology [2] - The Scale-up ultra-large-scale super-node computing platform, ranked first in Huawei's sixth "Top Ten Inventions" selection, exemplifies the use of system architecture and interconnection technology to address single-chip performance shortcomings [2] - This invention allows heterogeneous parallel processors, CPUs, memory, and storage within super-nodes to form a fully peer-to-peer interconnection architecture, enabling flexible resource allocation based on task requirements [2]

Core Insights - Huawei is set to release a groundbreaking technology in the AI field on November 21, which aims to address the efficiency challenges in computing resource utilization [2] - The upcoming technology is expected to enhance the utilization rate of GPU and NPU resources from the industry average of 30%-40% to 70%, significantly unlocking the potential of computing hardware [2] - The technology will enable unified resource management and utilization of computing power from Nvidia, Ascend, and other third-party sources through software innovation, thereby providing more efficient support for AI training and inference [2] - Huawei's technology shares commonalities with the core technology of Israeli AI startup Run:ai, which was acquired by Nvidia for $700 million at the end of 2024 [2] - Run:ai has focused on GPU scheduling technology since its establishment in 2018, aiming to create a platform that allows AI models to run in parallel, regardless of whether the hardware is on-premises, in the cloud, or at the edge [2] Technical Insights - Managing workloads for generative AI, recommendation systems, and search engines requires complex scheduling to optimize system and underlying hardware performance [3] - Run:ai's core product is a software platform built on Kubernetes, designed for scheduling GPU computing resources. It employs dynamic scheduling, pooling, and sharding techniques to optimize GPU resource utilization, enabling efficient execution of deep learning training and inference tasks in enterprise environments [3]

具体来看，华为即将发布AI领域的突破性技术，可将GPU（图形处理器）、NPU（神经网络处理器） 等算力资源的利用率，从行业平均的30%至40%提升至70%，显著释放算力硬件潜能。 据透露，华为即将发布AI领域的突破性技术，是通过软件创新实现英伟达、昇腾及其他三方算力的统 一资源管理与利用，屏蔽算力硬件差异，为AI训练推理提供更高效的资源支撑。 记者11月16日获悉，华为将在11月21日发布AI领域的突破性技术，有望解决算力资源利用效率的难题。 同时，华为即将发布AI领域的突破性技术，与以色列AI初创公司Run:ai的核心技术路线有共同性，后者 在2024年底被英伟达以7亿美元资金收购。 Run:ai的核心产品是基于kubernetes（开源容器编排平台）构建的软件平台，用于调度GPU的计算资源， 通过动态调度、池化、分片等技术，实现GPU资源利用率的优化，让深度学习训练与推理任务在企业级 环境中高效运行。 （文章来源：中国基金报） 公开资料显示，Run:ai自2018年成立以来，一直专注于GPU调度技术，并致力于打造一个能将AI模型拆 分并行运行的平台，无论硬件位于本地、云端还是边缘。 据悉，管理生成式AI、 ...