Huawei Releases Open-Source AI Container Technology Flex:ai: Make Idle Computing Power “Move” and Allocate One GPU Card for Multiple Tasks | Latest News

Text | Fu Chong

Editor | Su Jianxun

Currently, the issues of "insufficient computing power" and "wasted computing power" are occurring simultaneously.

36Kr learned that at the recently held 2025 AI Container Application Implementation and Development Forum, Huawei officially launched the AI container technology - Flex:ai, which improves the utilization rate of computing power resources through three technological innovations. At the same time, Huawei, in collaboration with Shanghai Jiao Tong University, Xi'an Jiaotong University, and Xiamen University, jointly announced the open - sourcing of the results of this industry - academia cooperation.

The XPU pooling and scheduling software Flex:ai, which was released and open - sourced this time, is built on Kubernetes (a widely used container management platform). Simply put, through the refined management and intelligent scheduling of intelligent computing power resources such as GPUs and NPUs, it unifies scattered computing power into a "resource pool" on one hand, and intelligently allocates AI tasks of different scales into it on the other hand.

Specifically, Flex:ai mainly has three core capabilities:

In addressing the issue of resource waste in the small - model training and inference scenarios, the XPU pooling framework jointly developed by Huawei and Shanghai Jiao Tong University divides a single GPU or NPU computing power card into multiple virtual computing power units with a precision of 10%, achieving "cut as much as needed", and increasing the overall average utilization rate of computing power in such scenarios by 30%;

In order to aggregate the idle computing power on different machines in the cluster, the cross - node remote virtualization technology developed by Huawei and Xiamen University aggregates the idle XPU computing power of each node in the cluster to form a "shared computing power pool", enabling general - purpose servers without intelligent computing capabilities to participate in AI computing by remotely accessing GPU/NPU resources through the network;

Facing the challenge of unified scheduling of heterogeneous computing power resources of multiple brands and specifications in the computing power cluster, the Hi Scheduler intelligent scheduler jointly launched by Huawei and Xi'an Jiaotong University can sense the status of computing power resources of multiple brands and specifications in the cluster. Based on parameters such as task priority and computing power requirements, it automatically selects appropriate local or remote resources, achieving time - sharing multiplexing and globally optimal scheduling, and enabling the system to decide "which card should handle which task".

Huawei stated that the full open - sourcing of Flex:ai this time will open all core technological capabilities to developers from the industry, academia, and research fields, jointly promoting the construction of standards for the docking of heterogeneous computing power virtualization and AI application platforms, and forming a standardized solution for the efficient utilization of computing power.