Introduction

Application enablement

MindX DL

MindX DL is a deep learning component reference design that integrates Atlas 800 AI training servers, Atlas 800 AI inference servers, and GPU-based servers. It provides basic functions such as Ascend AI Processor resource management and monitoring, Ascend AI Processor scheduling optimization, and distributed training collection communication configuration generation, enabling partners to quickly develop deep learning systems. You can log in to the Ascend developer zone to obtain source code and reference documents.

ModelZoo

ModelZoo provides common network models based on the mainstream deep learning frameworks such as MindSpore, TensorFlow, and Caffe, including ResNet50 and VGG16. After obtaining the model, you can perform customized training or develop an application based on the model and Ascend AI Processors. Click ModelZoo for login.

Data center platform

Container engine plug-in (Ascend Docker)

Ascend Docker is a basic component of the Atlas data center solution, and can mount devices and drivers of Ascend AI Processors to containers so that AI jobs can run smoothly on Ascend devices as Docker containers.

Kubernetes device plugin

Kubernetes provides a device plugin framework mechanism for third-party devices. The Ascend AI Processor uses the Kubernetes device plugin to discover and report resources to kubelet. In addition, the Ascend device can be used in a K8S POD and container to periodically check the health status of the device.

System tool

CANN-based tools that are easy to deploy and maintain

CANN

AscendCL, GE, operator/acceleration library, HCCL, Runtime, and Driver

• The Ascend Computing Language (AscendCL) provides C++ or Python API libraries for users to train neural networks, including device management, memory management, model loading and execution, and operator loading and execution.
• During model training, AscendCL calls the API provided by the GE executor to compile the model, and loads the compiled and optimized model to the Ascend AI Processor through Runtime for calculation.
• Operator/Acceleration library: The operator library in the network model includes Caffe and TF operators. An independent acceleration library is provided, which is represented by the AscendCL APIs, for example, the matrix multiplication API.
• Huawei Collective Communication Library (HCCL) is based on the high-performance set communication library of the Ascend AI Processor. The HCCL provides single-server multi-device and multi-server multi-device set communication primitives and implements set communication over PCIe, HCCS, and RoCE high-speed links, efficiently completing gradient update tasks in AI training services.
• Runtime: provides functions such as chip memory management and device management.
• Driver: provides device driver interfaces for Runtime.
• The training hardware, as the hardware computing capability basis of Ascend AI processors, performs matrix-related computation of neural networks, general computation of control operators, scalars, and vectors, execution control, as well as image and video data preprocessing. The CANN guarantees execution of deep neural network (NN) computing.

Auxiliary development tool (CLI)

Auxiliary tools for model or service development, including the precision comparison tool, operator development tool, and profiling performance debugging tool

Full-pipeline development toolchain

MindStudio

MindStudio provides one-stop efficient development and simplified deployment capabilities for operators, models, and applications. For details, see the MindStudio Documentation.

O&M management tool

SmartKit

Users can use SmartKit to deploy, maintain, and upgrade training hardware devices in batches, simplifying operations and improving work efficiency. For details about SmartKit, see the FusionServer Tools 2.0 SmartKit User Guide.

FusionDirector

FusionDirector is the management software for intelligent O&M throughout the server lifecycle. It provides intelligent version management, intelligent deployment, intelligent asset management, intelligent energy efficiency management, and intelligent fault management. The visualized management GUI provides ultimate O&M experience for customers. For details about FusionDirector, see related documents.

Hardware

Atlas 300T AI training card

For details, see Table 2-1.

Atlas 800 AI training server

Atlas 900 AI cluster