What is Alibaba Cloud Platform for AI? - Platform For AI

Alibaba Cloud Platform for AI (PAI) is a one-stop machine learning platform that provides data labeling, model development, model training, and model deployment services. This topic describes what PAI is.

What is PAI?

PAI is a one-stop machine learning platform for developers. With its core modules, such as Machine Learning Designer, Data Science Workshop (DSW), Deep Learning Containers (DLC), and Elastic Algorithm Service (EAS), PAI provides an all-in-one solution for machine learning, covering data labeling, model development, model training, and model deployment. PAI supports multiple open-source frameworks and AI optimization capabilities. PAI is flexible and easy to use.

Features
Benefits
Billing
Scenarios

Features

AI development phase	Related module	Description
Data preparation	PAI-iTAG	In the data preparation phase, iTAG provides intelligent data labeling services. You can create data labeling tasks for the following data types: image, text, video, and audio. You can also create multimodal data labeling tasks. iTAG provides various content and question components for data labeling. You can use the preset templates provided by iTAG or custom data labeling templates based on your business requirements. iTAG also provides fully managed data labeling services that are outsourced.
Model development	PAI-Designer	Machine Learning Designer provides more than 140 mature algorithms and allows you to develop AI models by performing visualized drag-and-drop operations in a low-code environment.
Model development	PAI-DSW	DSW allows you to develop models through interactive programming. DSW is a cloud integrated development environment (IDE) embedded with Notebook, VS Code, and Terminal. DSW also grants you sudo permissions for flexible management.
Model training	PAI-DLC	You can use general computing resources and Lingjun resources for model training based on scenarios and computing power types. General computing resources: Alibaba Cloud general computing resources, such as Elastic Compute Service (ECS), Elastic Container Instance, and Elastic GPU Service (EGS). DLC supports multiple training frameworks such as Tensorflow, PyTorch, and MPI and provides flexible, stable, and easy-to-use model training services. Lingjun intelligent computing resources: Based on the integrated optimization technology of software and hardware, DLC allows you to run ultra-large distributed deep learning jobs and provides benefits such as high performance, high efficiency, and high utilization. Both Lingjun AI Computing Service Serverless Edition on the Alibaba Cloud public cloud and Lingjun AI Computing Service Dedicated Edition for a single tenant are supported. An AI engineering end-to-end platform with software-hardware heterogeneously integrated computing power is provided.
Model deployment	PAI-EAS	EAS allows you to deploy models as online inference services or AI-powered web applications. EAS is suitable for multiple scenarios, such as real-time inference, asynchronous inference, and offline inference.
Model deployment	PAI-EAS

View more features

Benefits

End-to-end AI-powered R&D

Supports data labeling, model development, model training, model optimization, model deployment, and AI O&M as a one-stop AI platform.
Provides over 140 types of optimized built-in algorithm components.
Provides core capabilities such as multiple modes, deep integration with big data engines, multi-framework compatibility, and custom images.
Provides cloud-native AI development, training, and deployment services.

Multiple open-source frameworks

Supports Flink, a stream computing framework.
Supports TensorFlow, PyTorch, Megatron and DeepSpeed, which are optimized deep learning frameworks based on the related open-source frameworks.
Supports Parameter Server, a computing framework that can process hundreds of billions of samples in parallel.
Supports Spark, PySpark, MapReduce, and other mainstream open-source computing frameworks.

Industry-leading AI optimization

Supports high-performance training framework, sparse training scenarios, billions to tens of billions of sparse features, tens to hundreds of billions of samples, and distributed incremental training of thousands of workers.
Supports acceleration of mainstream framework models such as RestNet50 and Transformer language model (LM) by using PAI Blade.

Diverse service modes

Supports fully managed and semi-managed services for public cloud.
Provides high-performance AI computing clusters and lightweight service modes.
Supports periodical scheduling by using DataWorks. You can run scheduled tasks in the production or development environment. This enables data isolation.

Billing

Billing method	Description	Involved module
Pay-as-you-go	If you use the pay-as-you-go billing method, you are charged based on the actual usage of each module. The pay-as-you-go billing method is suitable for short-term or uncertain workloads. It allows you to pay for resources based on the actual amount of resources that you use. The pay-as-you-go billing method is suitable for test environments, development environments, unexpected requirements, or projects in the early phases.	Machine Learning Designer, DSW, DLC, and EAS
Subscription	The subscription billing method is suitable for long-term and stable workloads. You must pay in advance to use resources for a specific period of time, such as a month or a year. The subscription billing method is more cost-effective than the pay-as-you-go billing method for long-term use.	DSW, DLC, and EAS
Resource plan	Resource plans refer to quota plans of specific resources that you can purchase in advance. Resource plans are suitable for scenarios in which you want to use a large number of specific resources. You can purchase quota plans for specific resources at more favorable prices.	DSW
Savings plan	You can purchase savings plans in advance, which offer specific discounts or benefits. Savings plans provide discounted pay-as-you-go rates in exchange for committing to a specific spending amount within a specific period of time.	DSW and EAS
Pay-by-inference-duration	You are charged based on the actual inference duration. The resource specifications support automatic scaling based on the number of service requests. This billing method is suitable for inference tasks that require indefinite quantities and is appropriate for high-concurrent requests and dynamic loads.	EAS

View more information about billing

Scenarios

Scenario	Use case
Large language model (LLM)	Quickly deploy LLMs in EAS
Retrieval-Augmented Generation (RAG)-based LLM chatbot	RAG-based LLM chatbot
AI painting	Use Stable Diffusion web UI to deploy an AI painting service
AI video generation	Generate a video
Distributed training	Submit a standalone training job that uses PyTorch

View more scenarios

References

If you use PAI for the first time, you must activate PAI and create a default workspace. For more information, see Activate PAI and create a default workspace.