At the Huawei Developer Conference (HDC.Cloud) 2021, the company unveiled MindSpore 1.2, the first AI computing framework that supports the training of large models with 200 billion parameters in China.
MindSpore 1.2 enables fully automated parallel, full-scene AI and interpretable recommendation models in AI.
Huawei says MindSpore is the industry's first fully automated parallel framework based on automatic awareness of network topology and cluster resources and has developed Chinese pre-trained models with 200 billion parameters.
MindSpore enables fast application, efficient operation, and effective collaboration of hardware devices in different scenarios, allowing Huawei Watch GT's wrist lift recognition rate to be improved by 80%, with a latency of less than 5ms and a model of less than 1KB.
Huawei says MindSpore's built-in semantic-level interpretable recommendation model TB-Net is based on original knowledge graph bi-directional conduction technology, which can identify core features and key paths that influence user behavior, and its evaluation index is 63% higher compared to other models.
After being open-sourced in March last year, the MindSpore community has more than 170,000 developers, with more than 240,000 software downloads, and is used at scale in more than 10 industries, Huawei said.