在行业大模型的构建过程中，为进一步提高算力资源利用率，第四范式先知AIOS 5.1版本新增GPU资源池化(vGPU)能力，拥有五大技术亮点：Today, Fourth Paradigm (06682) officially released the AIOS 5.1 version of Prophet.
According to the Zhitong Finance APP, today, Fourth Paradigm (06682) officially released the AIOS 5.1 version of Prophet. This version adds the GPU resource pooling (vGPU) capability, realizing on-demand allocation and rapid scheduling of computing resources in hardware clusters, saving up to 80% of hardware costs, and improving GPU utilization by up to 5-10 times.
Fourth Paradigm's Prophet AIOS 5 is an industry-leading platform for large-scale model development and management. With the goal of enhancing the core competitiveness of enterprises, the platform provides end-to-end services for constructing, deploying, and managing large industry models, based on supporting access to various modal data from the enterprise, providing low-threshold modeling tools for large-scale model training and fine-tuning, an innovation service system for scientists, the Polaris strategy management platform, large model governance platform, main computing power optimization, and other capabilities.
In the process of building large industry models, to further improve computing resource utilization, Prophet AIOS 5.1 version of Fourth Paradigm adds the GPU resource pooling (vGPU) capability, which has five technical highlights:
Fully adapt to domestic and non-domestic computing power, support mixed deployment, and unified scheduling.
Compute and memory superimposition, with computing split to 1% precision and memory split in M bytes.
Having distributed scheduling and management capabilities on kilocalorie level.
Support custom isolation strategies to achieve shared or exclusive computing pools.
Improve inference performance by more than 10 times using multi-task sharing and processing optimization technology.
