Introduction to Mi Cloud Cloud-ML Services

Introduction to Xiaomi Cloud-ML

Xiaomi Cloud-ML Services, or Xiaomi Cloud-ML, is Xiaomi's distributed cloud service, with Mi Eco Cloud optimized for high performance machine learning.

Developers can use the GPU training model in the cloud and start distributed training tasks in just seconds. It is compatible with deep-learning frameworks, such as TensorFlow. Deploying trained models only takes one click. Developers can create GPU-based development environments to provide model development, training, debugging, optimizing, testing, deploying, and predicting, with the one-stop solutions we provide. Cloud-ML Services also provides multiple ways to provide access to the service, such as APIs, SDKs, command lines, and Web consoles. It also supports flexible billing by seconds, providing convenience of usage to AI experts.

Xiaomi Cloud-ML Features

Ease of Use

Supports easy-to-use command line tools on Linux/Mac/Windows operating systems or Docker, as well as using Cloud-ML Services via APIs, SDKs, or Web consoles.

Compatibility

Supports standard APIs like that used by TensorFlow and other such deep-learning frameworks; compatible with Google Cloud-ML samples codes. The same model codes can be trained on different cloud platforms, to avoid being bound to a single vendor.

High Performance

Supports ultra-high GPU computing performance. Maximum support to 56 core CPU with 128G memory. Support data parallel and model parallel transfer, and single-machine, multiple-card and multiple-machine, multiple-card distributed training.

Flexibility

Supports on-demand application and allocation of CPU, memory, and GPU resources. Second-level metering and billing functions can be implemented based on task runtime.

Security

Supports multi-tenant authentication and authorization mechanism using Access key/Secret key. User Quotas can be dynamically adjusted online.

Integrity

Supports cloud training. Users can do coding and submit it with a press of a key to cloud training. Supports CPU-based or GPU-based training. Supports 17 mainstream deep-learning frameworks with super-parameter automatic fine-tuning and other functions.

Supports model services. A user-trained model can be deployed to the cloud platform with one click, providing general-purpose, high-performance gRPC services, and supporting functions such as online upgrade of models and load balancing of multiple instances.

Supports development environment. Users can create deep-learning environments, such as TensorFlow, on the platform, automatically allocate CPU, memory, and GPU resources, and support functions such as Notebook and password encryption.