Huaweis FullStack AllScenario AI Solution AI applications Full

  • Slides: 8
Download presentation
Huawei’s Full-Stack, All-Scenario AI Solution AI applications Full stack Ascend. Nano Ascend. Tiny Ascend.

Huawei’s Full-Stack, All-Scenario AI Solution AI applications Full stack Ascend. Nano Ascend. Tiny Ascend. Lite Model. Arts Application enablement Mind. Spore Framework CANN Chip enablement Ascend. Mini Ascend. Max IPs and chips Atlas All scenarios Consumer devices Public cloud Private cloud Edge computing Industry Io. T devices Application enablement: whole-process services (Model. Arts), layered APIs, and pre-integration solution Mind. Spore: unified training and inference framework for device/edge/cloud (independent or collaborative) CANN: chip operator library and highly automated operator development tool Ascend: a series of AI IPs and chips with unified and scalable architecture Atlas: various products built on Huawei Ascend AI processors for device-edgecloud AI infrastructure for all scenarios

Atlas AI Computing Portfolio Superior computing power Atlas 900 AI cluster Cloud-edge-device collaboration All-scenario

Atlas AI Computing Portfolio Superior computing power Atlas 900 AI cluster Cloud-edge-device collaboration All-scenario deployment Atlas 800 AI server Model 9000/9010 Model 3000/3010 Cloud-device synergy Ascend 310 AI processor Atlas 300 AI accelerator card Atlas 500 AI edge station Model 3000/9000 Model 3000 Edge Ascend 910 AI processor Atlas 200 AI accelerator module Model 3000 Device

Atlas Accelerates AI Training Ascend 910 AI processor World’s most powerful training card World's

Atlas Accelerates AI Training Ascend 910 AI processor World’s most powerful training card World's most powerful training server Atlas 300 AI accelerator card Atlas 800 AI server Model 9000/9010 World's fastest AI training cluster Atlas 900 AI cluster

Atlas Accelerates AI Inference Ascend 310 AI processor Highest density, 64 video inference channels

Atlas Accelerates AI Inference Ascend 310 AI processor Highest density, 64 video inference channels Edge intelligence and cloud-edge synergy AI inference platform with ultimate computing power Atlas 200 AI accelerator module Atlas 300 AI accelerator card Atlas 500 AI edge station Atlas 800 AI server Model 3000/3010 Intelligent devices with 7 x higher performance

CANN: High-Performance Chip Operator Library and Automated Operator Development Tool CANN: CANN Compute Architecture

CANN: High-Performance Chip Operator Library and Automated Operator Development Tool CANN: CANN Compute Architecture for Neural Networks Ø Includes the chip operator library and highly automated operator development tool for optimal development efficiency and Ascend performance matching. Ø Fusion Engine: Ascend internal storage reduces operator calling overheads and memory migrations while improving performance Fusion Engine Task information Operator fusion TBE operator development tool TIK TVM Operator management CCE operator library Convolution Control flows Matrix multiplication Ø Vectors CCE operator library: high-performance operator library based on indepth collaborative optimization the Ascend chip TBE operator development tool: various APIs for custom operator development and automatic optimization, improving operator development efficiency CCE compiler Ø Compiler front end AI core AI CPU CCE compiler: compiler and binary tool set using heterogeneous hybrid programming language (C/C++ extension) to optimize performance and programming, enabling Ascend to support all scenarios

Mind. Spore: All-Scenario AI Computing Framework All-scenario AI application ecosystem Mind. Spore All-scenario unified

Mind. Spore: All-Scenario AI Computing Framework All-scenario AI application ecosystem Mind. Spore All-scenario unified APIs Automatic differentiation Automatic parallelism Automatic optimization Mind. Spore IR computational graph expression On-device execution Pipeline parallelism Deep graph optimization Device-edge-cloud, synergistic, distributed architecture (deployment, scheduling, communication, etc. ) Processors: Ascend, GPU, and CPU User-friendly development: AI algorithm as code Efficient execution: optimized for Ascend and GPU Flexible deployment: all-scenario on-demand collaboration

1 Platform + 3 Plans Support Ascend Industry Partners and Developers CNY 3 billion

1 Platform + 3 Plans Support Ascend Industry Partners and Developers CNY 3 billion investment, 3000 partners, and 1 million developers in 5 years Business partners Developers Universities Solution Partner Program Developer Enablement Plan AI Talent Development Plan Platform for Ascend industry development Industry cooperation Open-source Technical support Marketing support

Atlas Products: Built on Ascend 310 and Serving Many Industries 50+ Atlas Industry Solutions

Atlas Products: Built on Ascend 310 and Serving Many Industries 50+ Atlas Industry Solutions …… Security Finance Electric power Transportation Safe city Smart banks Unmanned inspection of high-voltage lines Free flow at provincial toll stations Internet Intelligent recommendati ons Carriers Smart customer service center