共創(chuàng)AI開發(fā)世界

1、技術(shù)創(chuàng)新，變革未來共創(chuàng)AI開發(fā)世界Huaweis AI portfolioCANN(Compute Architecture for Neural Networks)AscendAllScenarios全場景MindSporeTensorFlowPyTorchPaddlePaddleAscend-MaxAscend-MiniAscend-TinyAscend-LiteAscend-NanoFull Stack 全棧華為AI解決方案AI Applications AI 應(yīng) 用基于統(tǒng)一、可擴(kuò)展架構(gòu)的系列化AI IP 和 芯片Application enablement:Full-pipeline

2、services (ModelArts), hierarchical APIs, andpre-integrated solutions應(yīng)用使能：提供全流程服務(wù)(ModelArts), 分層API和預(yù)集成方案MindSpore：Unified training and inference framework for device, edge, and cloud (both standalone and cooperative)支持端、邊、云獨(dú)立的和協(xié)同的統(tǒng)一訓(xùn) 練和推理框架CANN：Chip operators library and highly automated operators d

3、evelopment toolkit芯片算子庫和高度自動(dòng)化算子開發(fā)工具Ascend：AI IP and chip series based on unified scalable architectureApplication Enablement 應(yīng)用使能Framework框架Chip Enablement芯片使能ModelArtsGeneral APIsAdvanced APIsPre-integrated SolutionsHiAI EngineHiAI ServiceIP & ChipIP和芯片Consumer Device消費(fèi)終端Public Cloud公有云Private Clou

4、d私有云Industrial IoTDeviceIoT 行業(yè)終端Edge Computing 邊緣計(jì)算Ascend 910:Greatest computing density in a single chip華為昇騰910：單芯片計(jì)算密度最大Ascend-Max Architecture: Da VinciHalf-Precision (FP16): 256 TeraFLOPS Integer-Precision (INT8) : 512 TeraOPS128 Channel FHD Video Decoder H.264/265Max Power: 350W7nm2019 Q2Ascend

5、-Max架構(gòu): 達(dá)芬奇半精度 (FP16): 256 TeraFLOPS整數(shù)精度 (INT8) : 512 TeraOPS128 通道 全高清 視頻解碼器 - H.264/265最大功耗: 350W 7nm2019 Q2華為昇騰910Ascend-Mini 架構(gòu)：達(dá)芬奇半精度 (FP16): 8 TeraFLOPS整數(shù)精度 (INT8): 16 TeraOPS16 通道 全高清 視頻解碼器 - H.264/2651 通道 全高清 視頻解碼器 - H.264/265最大功耗：8W 12nm FFC現(xiàn)在Ascend-Mini Architecture: Da VinciHalf-Precision

6、 (FP16): 8 TeraFLOPS Integer-Precision (INT8): 16 TeraOPS16 Channel FHD Video Decoder H.264/2651 Channel FHD Video Encoder H.264/265Max Power: 8W12nm FFCAvailable NowAscend 310:Extremely efficient AI SoC for low-power computing華為昇騰310：極致高效計(jì)算低功耗AI SoC華為昇騰310As of Sep 30, 2018 / 截至2018年9月30日 Available

7、 on services服務(wù)functions 功能4518platform services 平臺類15vision services 視覺 8language services 語言4decision-making services決策 14236platform functions 平臺98APIs8pre-integrated solutions預(yù)集成方案HUAWEI CLOUD EI華為云 EIMindSporeGPUModelArtsGeneral APIsAdvanced APIsASRTTSModerationImageOCRFaceCBSImageSearchVCMVCTVA

8、RIDSPre-integrated SolutionsNLPAISVGSVCCBasic PlatformsCityInternetHomeVehicleLogisticsHealthcareCampusManufacturingAI開發(fā)當(dāng)前最大的挑戰(zhàn)是什么？What are the biggest challenges for AI development?Larger data sets數(shù)據(jù)量越來越大More complicatedmodels模型越來越復(fù)雜The long process from ideation to production hinders AI developmen

9、t從設(shè)想到實(shí)現(xiàn)，“慢”阻礙了AI產(chǎn)業(yè)發(fā)展Long training time訓(xùn)練耗時(shí)長Limited resources資源緊張More and more computing計(jì)算過程越來越耗時(shí)Acceleration resources are expensive and hard to acquire加速資源昂貴難獲取Long learning cycle工具繁多學(xué)習(xí)周期長ModelArts更快的普惠AI開發(fā)平臺HiLens更懂開發(fā)者的視覺AI應(yīng)用開發(fā)平臺Atlas 200 Developer Kit與線上體驗(yàn)一致的離線開發(fā)套件HiQ探索未來的量子計(jì)算模擬器與編程框架數(shù)據(jù)準(zhǔn)備效率百倍提升10

10、0 x higher data preparation efficiency40TB data processing：xxxx man-days - xx man-days40TB數(shù)據(jù)處理：xxxx人天-xx人天Video視頻Frameextraction視頻處理Images圖片集Intelligentfiltering智能篩選Difficult cases難例集Labeling標(biāo)注ModelArts更快的AI開發(fā)平臺Training 訓(xùn)練模型訓(xùn)練耗時(shí)降低一半50% less time for model trainingOptimized algorithms算法優(yōu)化Fast快Distrib

11、uted speedup ratio 0.8 with 1000 nodes1000規(guī)模集群，訓(xùn)練加速比0.8Simplifiedhyperparameter tuning簡化調(diào)參Simplified簡Simplified code簡化代碼DeepLearningModelArts更快的AI開發(fā)平臺模型一鍵部署到云、邊、端One-click model deployment to devices, edge, and cloudEdge reasoning邊緣推理Batch reasoning批量推理Online reasoning在線推理ModelArts更快的AI開發(fā)平臺AI models

12、AI模型Deploy部署IEF用AI方式加速AI開發(fā)過程 自動(dòng)學(xué)習(xí)Accelerating AI development with AI techniques(automatic learning)Automatic machine learning自動(dòng)機(jī)器學(xué)習(xí)Automatic deep learning自動(dòng)深度學(xué)習(xí)自適應(yīng)訓(xùn)練白盒化UI向?qū)IIntuitive UI guideExeMLWhite boxModelArts更快的AI開發(fā)平臺快亦有道 匠心打造全流程管理End-to-end process managementAPIAutomatic visualizationof the

13、development flow開發(fā)流程的自動(dòng)可視化Easy resuming from a training checkpoint訓(xùn)練斷點(diǎn)重啟Easy comparison ofthe training results訓(xùn)練結(jié)果輕松比對ModelArts更快的AI開發(fā)平臺快亦有道 幫開發(fā)者實(shí)現(xiàn)AI資源復(fù)用Helping developers reuse AI resourcesAI MarketplaceAI共享平臺APIData數(shù)據(jù)Model模型APIs應(yīng)用Enterprise-wide sharing企業(yè)內(nèi)共享Higher efficiency效率提升External market外部市場

14、AIOpen ecosystem開放生態(tài)ModelArts更快的AI開發(fā)平臺f(x)Fast to use上手快Fast training訓(xùn)練快Fast TTM上線快Data preparation數(shù)據(jù)準(zhǔn)備ExeML自動(dòng)學(xué)習(xí)Model deployment 模型部署Model optimization 模型優(yōu)化Hyperparameter tuning訓(xùn)練調(diào)參AI Marketplace 共享平臺End-to-end processmanagement全流程管理A Faster and Inclusive AI Development PlatformModelArts更快的普惠AI開發(fā)平臺E

15、xeML視覺感知是智能世界基本能力Visual perception:Basic ability of the intelligent worldCameras enable visibility in everything各行各業(yè)部署攝像頭，萬物都安上了眼睛Observe 觀測Understand 理解Change 改變Text文字TPictures圖片Voice & images語音&圖像Videos視頻Online, real-time video streaming在線實(shí)時(shí)視頻流The digital world is lookingat all aspects of the phys

16、ical world數(shù)字世界正在全景式地“看”物理世界Model training模型訓(xùn)練ModelArtsSkill developmentSkill開發(fā)Function developmentFunction開發(fā)Split, quantization, pruning optimization 拆分、量化、剪枝優(yōu)化Skill LibrarySkill MarketAI-enabled devices 智能終端Automatic deployment 自動(dòng)部署Diverse pre-built skills豐富的預(yù)置技能Upload上傳Download下載Model optimization

17、 模型優(yōu)化給AI以慧眼HiLensHiLens: Vision AI application development platform慧眼-視覺AI應(yīng)用開發(fā)平臺用于開發(fā)者的HiLens慧眼支持深度學(xué)習(xí)推理執(zhí)行的視覺裝備HiLens: Visual devices for developers, supporting deep learning reasoning executionEmbedded Ascend 310 chipset: High-performance reasoning Face recognition withinmilliseconds Hundreds of fram

18、es within 1s搭載昇騰310芯片: 高性能推理能力 人臉檢測ms級1s百幀Built-in lightweight containers: Fewer resources, lower demand for network bandwidth, quicker downloading推理驗(yàn)證內(nèi)置輕量級容器： 資源占用少，降低網(wǎng)絡(luò)帶寬 下載啟動(dòng)時(shí)間快給AI以慧眼HiLensSkills accelerate the development of visualperception applicationsSkill加速視覺感知類開發(fā)Campus園區(qū)Home家庭Enterprise企業(yè)給A

19、I以慧眼HiLensMovement imitation demo機(jī)器人動(dòng)作模擬Demo演示obots that understand you讓機(jī)器人讀懂視頻看懂你Model: Key human body parts tagging, movement classification模型：人體關(guān)鍵點(diǎn)標(biāo)注，動(dòng)作分類Skill: Movement imitationSkill：動(dòng)作模仿Skill marketSkill市場Model training模型訓(xùn)練Skill developmentSkill開發(fā)AI app development without any basic knowledge

20、of AI零基礎(chǔ)開發(fā)AI應(yīng)用Robots that understand you讓機(jī)器人讀懂視頻看懂你Movement imitation demo機(jī)器人動(dòng)作模擬Demo演示SkillmarketSkill市場Modeltraining模型訓(xùn)練SkilldevelopmentSkill開發(fā)ModelArts一站式AI開發(fā)平臺Robots that understand you讓機(jī)器人讀懂視頻看懂你Movement imitation demo機(jī)器人動(dòng)作模擬Demo演示SkillmarketSkill市場Modeltraining模型訓(xùn)練SkilldevelopmentSkill開發(fā)HiLens

21、一鍵下發(fā)SkillRobots that understand you讓機(jī)器人讀懂視頻看懂你Movement imitation demo機(jī)器人動(dòng)作模擬Demo演示讓機(jī)器人讀懂視頻看懂你Movement imitation demo機(jī)器人動(dòng)作模擬Demo演示Robots that understand youModelArts 現(xiàn)場公開賽“保時(shí)捷”花落誰家？ModelArts model development competitionWho will be the winner?Number of Yun Bao: 2云寶數(shù)量: 2Use ModelArts to develop a mode

22、l that identifies the number of Yun Bao基于ModelArts平臺開發(fā)一個(gè)識別云寶數(shù)量的模型Small, Powerful, Multiple Interfaces小巧，強(qiáng)大，接口豐富Embedded Ascend 310chipset搭載昇騰310芯片Reasoning & verification 推理驗(yàn)證Operator development算子開發(fā)Atlas 200 Developer KitModelArts 線下開發(fā)套件豐富的算子庫和易用的開發(fā)IDEDiverse operators library and easy-to-use IDEsS

23、ame as Huawei Cloud與華為云一致的Automatic performance optimization自動(dòng)性能優(yōu)化DSL abstract interfacesPythonDSL級抽象接口Python通用語言Diverse pre-builtoperators library豐富的預(yù)置算子庫Tensor EngineDSLAPIModelSyntaxEngineTVMs-compatible frameworks & network interfaces各類框架及網(wǎng)絡(luò)接口兼容TVMHuawei Cloud and AI explorations華為云和AI未來探索AI + Q

24、uantum computingAI+量子計(jì)算Computing capability 算力Todays AI 今天的AIFuture AI 未來的智能Classical computing cloud經(jīng)典計(jì)算云Quantum computing cloud 量子計(jì)算云N qubits = 2N classical bitsExponential growth in computing capabilityN量子比特=2N經(jīng)典比特指數(shù)增長算力Quantum AINP-hard optimization量子AINP復(fù)雜優(yōu)化問題Giant molecularsimulation for material and drug design 材料、藥物大分子模擬Natural intelligence自然智能探索AI結(jié)合量子計(jì)算致未來：To explorers：Explorations into AI & quantum computingHiQ: Quantum computing simulator & programming frameworkHiQ 量子計(jì)算模擬器與編程框