Оптимизированные решения huawei и … for mfti v1.pdfВклад в r&d...

Оптимизированные решения HUAWEI

и эффективные дата-центры ИИ

Boundless Computing Inspires an Intelligent World

Михаил Плескунин

[email protected]

+7(985)765-54-40

mailto:[email protected]

Agenda

Аппаратные

решения

Решения для

ИИО компании ARM

Взгляды на ИИ

Bring digital to every person, home and organization

for a fully connected, intelligent world.

170+Стран

70Рейтинг в списке

Fortune Global 500

180,000сотрудников

36Совместных

инновационных центров

16НИОКР

центров

80,000НИОКР

сотрудников

Вклад в R&D

Входит в первые 10 компаний,

инвестирующих в НИОКР

Более

50 млрд долл. СШАЗа последние 10 лет

180 000 сотрудников:

45% являются инженерами в НИОКР

Объем инвестицийс 2008 по 2017 (млрд. долл. США)

1.62.1

2.73.8

4.85.2

20172008 2009 2010 2011 2012 2013 2014 2015 2016

6.6

9.2

11.0

13.8

Юг

Центр

Северо-запад

Поволжье

Сибирь

Дальний восток

Урал

Huawei в России – 20 лет

Головной офис

Центр обучения

Региональный офисЦентр поддержки

Центр разработок

10 региональных офисов

Инновационный центр

2000+ Сотрудников2 Центров разработок1 Центр инноваций

1 Центр обучения5 Сетевых академий (20 до 2020)2 Центра поддержки58 Складов зап. частей

СОТРУДНИКИ ОБУЧЕНИЕ И ПОДДЕРЖКАОФИСЫ РАЗРАБОТКА

Consumer Products & Services

Iconic global technology brand

Carrier Products & Services

Best strategic partner for carriers

Enterprise Products & Services

Enabler and preferred partner for digital transformation

A Global Leader in ICT Products and Solutions

Information Distribution & Presentation Information Transmission Information Storage & Processing

Data center infrastructure

Big data analytics

AI platforms

Global carriersHundreds of millions of consumersGlobal enterprises, government, and

industry

ICT Solutions and Services for Three Customer GroupsInformation Distribution & Presentation, Transmission, Storage & Processing

Managed services and system integration

Cloud Products & Services

Cloud partner with reliable, trusted, evolvable services

Cloud services

gPaaS

Cloud OS

Smartphones

MBB & home appliances

Wearables

Vehicle devices

Wireless networks

Fixed networks

Carrier software

Core networks

Enterprise networks

IoT connection management

Agenda


решения




Infrastructure

Hardware devices

System environment

Cluster management

Industry applications

Service

Capabilities

Focus on HPC Hardware Platforms and Infrastructure

Data center Liquid cooling infrastructure

Compute Storage

BladeHigh-

densityFat

nodeNVMe SSD

FPGAFile

storage

Convergedstorage

Network

IB switch

OPA switch

Operating systems

WindowsSUSE, Cent OS,

Red HatFusionSphere

File systems

NFS Lustre GPFS

Monitoring/Scheduling/Web portal

BCMIBM

Platform

Runtime library environment

MPI OpenMPIMVAPIC

H2PBS

Works

Ext3, ext4, XFS

Grid Engine

Manufacturing

FLUENTPAM-

CRASH

Life science

BLAST Gaussian NAMDSTAR-CCM+ Oil & gas

exploration

Halliburton

Schlumberger

Meteorology

WRF MM5

Power supply devices

Modular equipment room

Liquid coolingservers

Liquid cooling cabinet

Solution design

Application tuning

System deployment

Training and O&M

Joint innovation

Huawei HPC Platform

Huawei HPC Platform - FusionServer V5 Series Servers

Multi-Node Server Blade Server Rack Server

Supports SKU with integrated OPA All-flash acceleration

Converged architecture with network

switching EDR IB & OPA 100G

4 CPUs on a single node, with simplified

network solution Balance between compute density and

power consumption loads

X6000 2U 4-node

E9000 12U multi-node

2488 2U 4-socket

Fat Node Server Heterogeneous Server

Multi-core compute on a single node Massive in-memory computing

Supports multiple CPU:GPU ratios Fully modular design

KunLun

8- to 32-socket

GPU-accelerated server, 4U rack+

Bas

ic P

latf

orm

Exte

nd

ed P

latf

orm

Supports full series of

Intel® Xeon® Scalable processors

Board-level liquid cooling, PUE ≤ 1.1

Green and energy-saving

Supports rear-access aggregated

management network port

Simplifies cablingSupports 2 kW/3 kW PSUs and power capping

TDP = 205 W CPU

Supports EDR IB & OPA standard

cards

High-speed interconnect

Supports 2.5'' NVMe SSDs and 3.5'' HDDs

Suitable for different acceleration scenarios

4 x 1U rack servers

=

X6000 High-Density Server - Fully Upgraded with New Features

XH321 V5 Half-Width Dual-Socket Compute Node

Up to 72 Nodes in a Single Cabinet

New

features!

Supports SKU with integrated OPA

Boost cluster performance

1 x 2U X6000

• 2 full-series Scalable Processors, 16 DDR4 DIMMs

• 6 x 2.5'' NVMe SSDs or 3 x 3.5'' HDDs + 2 x M.2

SSDs

• 2 x GE + 2 x 10GE LOM ports, 2 PCIe x16 slots

• Supports air cooling or liquid coolingLiquid

cooling!

Rich Compute Node Types

E9000 Blade Server - Converged Architecture Computing Platform

All-flash compute nodeCH225 V5: full-width, 12 NVMe/SAS/SATA HDDs/SSDs

Balanced compute nodeCH121 V5: half-width, 2 Scalable Processors

Powerful Integrated Switch

Capabilities

• Available in two form factors: half-

width and full-width

• Supports CPU, GPU, and all-flash

accelerated compute nodes

• Integrated IB EDR/OPA and Ethernet

high-speed switch modules

• 32T passive midplane switching

capacityStorage

Compute Network

Mgmt

GPU-accelerated compute nodeCH221 V5: full-width, 2 dual-slot GPUs

CX620 100G IB EDR

Switch module

CX320 10GE/40GE

Ethernet switch module

CX820 100G OPA

Switch module

Advantages of E9000 Blade Nodes and Chassis

• 12U chassis, supporting up to 16 half-width

or 8 full-width compute nodes

• Supports compute, storage, and GPU nodes

Half-width slot

Full-width slot

Mgmt

module

Switch

module

PSU

Fan

module

CH121 V5 Half-Width 2S Blade

• Fully modular design

• Redundancy design for key

modules:

• Management modules

support 1+1 redundancy

• PSUs support N+N

redundancy

• Fan modules support

N+1 redundancy

Support Long-Term Evolution Chassis

Capabilities

• 2 full-series Scalable Processors, 24

DDR4 DIMMs

• 2*2.5”NVMe SSDs• 1 PCIe x16 slot, supporting air cooling or

liquid cooling

CH242 V5 Full-Width 4S Blade

• 4 full-series Scalable Processors, 48 DDR4 DIMMs

• 4*2.5”NVMe SSDs• 1 PCIe x16 slot

Newproduct!

Liquid cooling!

Brand New Blade Nodes

Integrated

high-speed switch!

• FusionServer G2500

Super Intelligence

• Supports up to 16 NVIDIA® Tesla® P4 in a 4U chassis

• Provides 256-channel intelligent video analytics of

people/vehicles/things

Ultra-Large Storage

• Supports 24 x 3.5'' hard drives, up to 240 TB

• Enables massive data storage and real-time

information retrieval

Edge Deployment Capability

• Supports 55°C working temperature

• 675 mm deep chassis, supporting installation in short cabinetsSafe citySmart

transportation

Facial

recognition

• FusionServer G5500

Outstanding Heterogeneous Computing Performance

• Supports 8 V100 training or 32 P4 inference accelerator cards

• GPUDirect RDMA, P2P, and NVLink interconnects

One-Click Switching of Heterogeneous Topologies

• One-click topology switching for AI and HPC

• Supports diversified applications and reduces hardware

investment

Fully Modular Architecture

• Decoupled design for CPU and heterogeneous resources,

supporting long-term technology evolution

• Efficient heat dissipation designAI HPC Cloud GPU

acceleration

• 8 V100/P100/P40/P4

GPU cards based on

PCIe

G5500 half-width model

1 chassis, 3 types of compute nodes, and 4 types of heterogeneous nodes

G5500 full-width model

G530 V5 half-width dual-socket

compute node

• 8 V100 GPU cards

based on NVLink

• 4 V100 or 8 V100 GPU cards

(150 W)/P4 based on PCIe

G560 full-width dual-socket

compute node

• 16 P4 GPU cards

based on PCIe

New

New

G560 V5 full-width dual-socket

compute node

GP608

heterogeneous node

GS608

heterogeneous node

GP308

heterogeneous node

GP316

heterogeneous node

G5500 Product Portfolio

KunLun 32S Fat Node Computing Platform— Ushering in a New Mission Critical Server Era

KunLun Fat Node Computing Platform

• Scale-up

• Fault-tolerant

• Failure Analysis

Engine

• Underlying feature

security

32S/576C, 32 TBElastically scale-up

architecture

RAS 2.0Open platform with the

highest-level reliability

NC Interconnect

ChipManagement Chip

Stability and ReliabilityUltimate Performance

KunLun Platform - Large-Capacity In-Memory Computing with Low Latency

Cores

Shared memory

Cores

Shared memory

Cores

Shared memory

Cores

Shared memory

…

Inner-server interconnect Shared memory

Inner-server interconnect

…NC interconnect chipCores

Up to 8 CPUs per system

Milliseconds of latency for inter-server

data transmission, including latency from

CPU processing, NIC processing, and

link transmission

KunLun 9032 supports 32 CPUs in a single

system, with 576 compute cores and memory

of up to 24 TB

CPU high-speed network reduces data

transmission latency to nanoseconds,

enabling faster service response

Prefabricated Data Center

• Cabinet-level deployment, taking only

2 hours to install onsite

• 1–6 cabinets, supporting HPC

systems of 10 to 100 TFLOPS

All-In-Cabinet Small-Sized HPC All-In-Room Medium- to Large-Sized HPC

FusionModule500

FusionModule800

All-In-Container HPC

• Supports single- or double-row

confined cold/hot channel deployment

in an equipment room of 500 m2 or

less

• 2–48 IT cabinets, supporting HPC

systems of 100 TFLOPS to 1 PFLOPS

• Factory prefabricated, pre-verified, and

onsite delivery, reducing deployment

cycle by 80%

• 8 IT cabinets, supporting HPC systems

of 10 to 100 TFLOPS

FusionModule2000

Single row

FusionModule1000A

Double rows

University of Nebraska

University of Tennessee

Digital Domain Company

Macao Meteorology Bureau

Singapore Global Foundries

Singapore Institute of Science

and Technology Research

Philippines Meteorology Bureau

(phase 1)

Singapore A-Star

University of Victoria

The University of Queensland

Deakin University

CASSAC Observatory in Chile

Brazil Mckinsey University

Cuba oil CUPET

Venezuela PDVSA

Ministry of water resource, Mexico

Ministry of Agriculture, Mexico

ULAKBIM

Yildiz Technical University (YTU), Turkey

Istanbul Technical University (ITU), Turkey

Harran University, Turkey

Yeditepe University, Turkey

Turkish Petroleum (TPAO)

China

Europe

Asia Pacific

North America

Latin America

Central Asia

Saudi Arabia MOI

Africa

Middle East

Zimbabwe Ministry of

Higher Education and

Technology

Development

South Africa CHPC

Institute of Disaster

Prevention, China

Meteorological

Administration

Environmental Protection

Bureau, Hebei Province

Beijing Institute of Data

Communication

Beijing Jiaotong University

Beijing University of

Aeronautics and

Astronautics

Southwest University

Capital Medical University

China Electric Power

Research Institute

China Meteorology Bureau

Shanghai Observatory

ZhongXin Biotechnology

Shanghai

Bureau of Geophysical

Prospecting INC.

Tsinghua University

Beijing Genomics Institute

BGP

UK Newcastle University

Imperial College London

University of Hamburg,

Germany

University of Lubeck,

Germany

University of Burgos, Spain

Illumination Mac Guff

Daimler Mercedes-Benz,

Germany

German IIya Ehrenburg

Netherlands Deltares

Italy CNR

University of Warsaw, Poland

Poland PCSS

University of Gdańsk, Poland

University of Silesia, Poland

Poland Cyfronet

Poland Qumak University

Saint-Petersburg State

University, Russia

HPC Installation Experience in Industries Worldwide

University Waterloo Cluster Launch in 2017

Compute Canada GP3Graham SuperComputing Cluster

(Sharcnet / University of Waterloo)

# 95

Rank

• High-density Servers

• Storage, Switches, Management systems

• 30+ Cabinets

• Liquid cooling

• 33,000 compute cores

• 1,228 TFLOPS (1.2 PFLOPS)

Graham HPC Cluster @ UWaterloo: Overview

Agenda


решения




ARM Ecosystem Goes Towards 100B Level

22 Years

4 Years

4 Years

1st 50 Billion ShipmentEmbedded

2nd 50BEmbedded, mobile

Next 100B+AI, IOT, Data Center

Source: ARM

20171991 2013 2021

2008 20192014 20162005

1st ARM

wireless base

station

1st 64-bit

server class

ARM CPU

1st ARM CPU

supporting

multiple sockets

1st ARM CPU

powered by AI

(Under planning)

1st ARM

mobile CPU

1st ASIC chip for

optical networks

TaiShan Server Highlights

Energy-efficient computing

High-performance Huawei ARM processors deliver the same

performance as that of x86 E5-2650 v4 using 20% less power.

Huawei's own chips

Huawei's own suite of compute, storage, transmission, and

management chips, underlying system-level design

CPU

Optimized for distributed computing

application workloads

Many-core computing architecture, suitable for parallel

computing scenarios such as big data analysis, distributed

storage, ARM cloud service, and HPC

Compute-storage balanced

Typical scenario: big data analytics

TaiShan Server — Various Specifications to Meet Diversified Requirements

Product TaiShan 2280 TaiShan 5280 TaiShan X6000

Form Factor 2U 2-socket rack server 4U 2-socket rack server 2U with 4 x 2-socket nodes

Processor 2 HiSilicon Hi1616 processors (16 nm/32 cores/2.4 GHz)

Memory 16 DDR4 DIMM slots

Local StorageConfiguration 1: 16 x 3.5-inch drives

Configuration 2: 27 x 2.5-inch drives40 x 3.5-inch drives 24 x 2.5-inch drives

LOM Network Ports 2 x GE electrical ports + 2 x 10GE optical ports

PCIe Expansion Up to 5 PCIe 3.0 x8 slots Up to 2 PCIe 3.0 x8 slots

Storage-intensiveTypical scenario:

distributed storage

Compute-intensiveTypical scenario: HPC

Ever-Improving Proprietary Processor Competitiveness for Mainstream

Applications in Data Centers

Hi1612 32 cores/2.1 GHz/16 nm

Hi1616 32 cores/2.4 GHz/16 nm

• Huawei-developed cores, 2- and 4-socket

interconnects

• 8 DDR4 2933 MHz memory channel

controllers

• Integrates I/O functions such as 25GE,

100GE, PCIe 4.0, and CCIX

• ARMv8 architecture, 2-socket interconnects

• 4 DDR4 2400 MHz memory channel

controllers

• Integrates I/O functions such as 10GE,

SAS/SATA, and PCIe

• ARMv8 architecture

• 4 DDR4 memory channel controllers

Hi162032 & 48 Cores/2.6 GHz/7 nm

Entry-level performance Mid-range performance High-end performance

• Big data analytics

• Distributed storage

• HPC

• In-memory database

• Big data analytics

• Distributed storage

• ARM cloud service

• Web application


• Web application


• Web application

Brings Valuable Solutions by Innovation on Silicon

Hi 1822

Smart Network

Controller

BMCSSD Controller

Hi1812 Hi1822 Hi1710

Computing

ARM 64bit CPU

Hi16XX

Storage Networking Management

Node

Controller

CPUBMC

NIC

controller

SSD

controller

NC

interconne

ct chip

CPU CPU

CPU

CPU

Storage/Network etc.

I/O controller chips

Server

management

chip

ARM CPU

1

2 3

4

1 2 3 4

Equivalent Performance with Better Efficiency

• Data reported by Huawei lab benchmark tests

Computing efficiency comparison

• Hi1616 provides equivalent computing performance

and better computing efficiency

• Hi1620 provides higher computing performance and

60% higher computing efficiency85w

105w

150 w 150w

Hi 1616 E5-2650 v4 Hi 1620 SkylakeGold 6148

CPU TDP Power

1 X1,04 X

1,9 X

3,02 X

E5-2650 v4 Hi 1616 SkylakeGold 6148

Hi 1620

CPU SPECINT Performance

60%

20%

Supreme Reliability

• Multiple measures to ensure 99.999% availability

• Key chips of the computing platform being

developed by Huawei, ensuring secure and reliable

data storage

Ultimate Performance

• In-depth optimization of software and hardware

based on application scenarios

• Huawei-developed high-performance SSDs,

delivering industry-leading IOPS and latency

Easy to Deploy and Scale

• Wizard-based deployment, simple and quick

• Capacity expansion by drive, node, cabinet, or

resource pool

New AppsTraditional

Apps

Object Storage ServiceFile Storage ServiceBlock Storage Service

TaiShan Platform

FusionStorage

iSCSI SCSI NFS CIFS FTP HDFS NDMP S3/Swift

OpenStack

TaiShan supports FusionStorage block storage.

TaiShan+FusionStorage Distributed Storage Solution

Huawei IT Data Center: Support for the Huawei R&D, Production,

and Office Clouds

FusionStorage

FusionSphere/FusionStageCompute

Storage

TaiShan

500 servers

730 servers

DCDC

DC

DC

DC

R&D

Terminal

mobile phone

simulation

HiSilicon

chip EDA

simulation

SVRP

simulation/storage

simulation

Production/office services

Tomcat/MySQl/Zabbix/PHP/JVM

doubled

1.45 million kWh

Huawei IT

Data Center

• Deployed 1,230 TaiShan

compute/storage servers

• Provides 18,000+ hard drives with a total

of 72 PB data capacity

• Allows for access of 75,000 R&D usersTaiShan

Microsoft Azure Data Center: Use TaiShan Servers to Support

Three Types of PaaS Service

• Deployed 200 TaiShan storage servers

• Based on CentOS and Windows Server

• Provides three types of PaaS services

storage, search, and Big Data

Storage service Bing search Big data

Shanghai

DC

Deploy TaiShan-based Azure cloud

infrastructure in Shanghai DC

Hatch Gaming Cloud: The Annual Revenue Is Expected to Double Thanks to the Cloud Gaming Community Service

Business model: The gaming social platform is

moved to cloud, and user data generates value.

Precise advertisement pushing and subscription

bring sustainable revenue sources.

• Average number of active users every day:

100M• Revenue from commercial advertisement

push: USD8/1,000 users• Daily revenue: USD800K• Annual revenue: 292M$ (142M€ in 2015)

Data source: Hatch

• Deployed 100 TaiShan compute servers

• Performance 4x that of x86 servers

• The system has been running stably for more than

one year after the system went live.

Agenda


решения




Railways

Iron steamship

Internal combustion

engine

Electricity

9000 BC~1000 AD 15th ~18th Century 19th Century 20th Century 21st Century

Multiple uses across the economy Many technological complementarities and spillovers

Domestication of plants

Domestication of animals

Smelting of ore

Wheel

Writing

Bronze

Iron

Water wheel

Three-masted sailing ship

Printing

Factory system

Steam engine

Automobile

Airplane

Mass production

Computer

Lean production

Internet

Biotechnology

Business virtualization

Nanotechnology

Artificial intelligence

(A set of technologies)

Public safety

• Safe City

• Intelligent transport

• Disaster prediction

Education

• Personalization

• Attention improvement

• Robo teacher

Healthcare

• Early prevention

• Diagnosis assistance

• Precision cure

Media

• Real-time translation

• Abstraction

• Inspection

Logistics

• Routing planning

• Monitoring

• Auto sorting

Finance

• Doc process

• Real-time fraud prevention

• Up-sell

Pharmacy

• Fast R&D

• Precise trial

• Targeted medicine

Insurance

• Auto detection

• Fraud prevention

• Innovative service

Retail

• Staff-less shops

• Real-time inventory

• Precise recommendations

Manufacturing

• Defect detection

• Industrial internet

• Predictive maintenance

Telecom

• Customer service

• Auto O&M

• Auto optimization

Agriculture

• Fertilization improvement

Remote operation

• Seeds development

Oil & Gas

• Localization

• Remote maintenance

• Operation optimization

Leaders

Managers / Experts

Junior Managers / Senior Professionals

Junior Employees

Managers / Experts

/ Data Scientists

Junior Managers / Senior Professionals

/ Data Science Engineers

Leaders

Junior Employees

business and academic events in 2018

Speech recognition: On par with

machine learning papers in 201720k

# of AI papers keeping up with Moore’s law in past 8 years

Object detection: Outperforming humans

humans

Translation: Approaching humans

countries with national AI plans22+

253+new AI startups in 20171,100+

AI-related M&As in 2017US$24 bn

AI-related VC investments in 2017US$14 bn

of B2B companies employ AI to augment sales processes

of enterprises have invested in or deployed AI4%

of retailers have invested in and deployed AI

of higher education institutions use AI to augment experience

of smart city implementations are using AI

of customer service operations integrated virtual assistants in 2017

of consulting and SI service projects were AI-related in 2017

of smartphones with AI capabilities in 2017

of B2C/B2B2C apps in China include AI in 2018

– Available AI talent vs. Global demand1%

~2%

5%10%

~5%

4%

2%

~10%

~10%

Inspiring opportunities we see

To BeAs Is

AI: Mostly in cloud, some at the edge

10 changes that will shape the future

Today’s basic algorithms invented before the 1980s

No labor, no intelligence

Models perform better in tests

Updates not in real time

Inadequate integration with other technologies

Only highly-skilled experts can work with AI

Scarcity of data scientists

Training in days or even months

Scarce & costly computing power

Pervasive AI for all scenarios. Respects and protects user privacy

Data and energy-efficient, secure, and explainable algorithms

Automated / semi-automated data labeling

Industrial-grade AI, perform excellently in execution

Real-time, closed-loop system

Synergy between AI and cloud, IoT, edge computing, blockchain, big data, databases, etc.…

AI as a basic skill, supported by one-stop platforms

Data scientists + Subject matter experts + Data science engineers

Training in minutes or even seconds

Abundant & affordable computing power

4 Billion Internet Users8 Billion Mobile Users

100 Billion IOT connections

1.7GB Data /people-day

3000 Billion US$ of mobile

payment

5G

AR/VR

SDN

IOT

Autopilot

Cloud

• Precise Network Control• Expert-Level Automatic O&M

•Low Latency Network•Cloud Real-Time Decision Making

• Edge Computing• Intelligent Network Slicing

•High Throughput(50Gbps/Site)•Low Latency(1ms)

• Whole Network Real Time insight• Intelligent Defense Against Risk

2025

• Smart Cache on Edge• Low Jitter Network

5G

SDN

IoT

AR/VR

Autopilot

Cloud Security

AI will Be Applied in Huawei's Future Business

TERMINAL AI

Achievements

Personal Devices: A New Round of Innovation is Coming…

After 2019

Intelligent phonesIntegrating real + digital

1983

First mobile phones

1995

Feature phones:Voice + text

Digit

2007

SmartphonesApp-based

services

Analog Mobile Internet Smart Internet

Two major innovations: Active sensing of the real world; interactive engagement and proactive services.

AI and natural interaction transform user experience.

AIPhysical world

Natural interaction

Brain wave

Digital world

Touch

Voice

Visual CognitionSensing

VR/AR

Mouse

New technologies drive transformative changes.

Communication

User experience

Camera

BatteryChips

Design

ID

AI

First AI Processing Core in Mobile (Kirin with NPU)

AI App

Framework

Library

ML acceleration API library for NPU on

mobile phone

Each API will be implemented by calling

ML OS library

Mobile OS assign different tasks to drivers

of physical Chips (NPU, GPU, CPU, etc)

NPU (Neural Network Processing Unit)

http://allforcellphones.com/wiki/

http://allforcellphones.com/wiki/

AI Camera: Simple & Fast for Best Picture

Consumer Experiences

Smart camera assistant

Preview real-time identification of ten scenesAutomatic adjustment of professional parameters

ENTERPRISE AI Achievements

Successful Practices of HUAWEI CLOUD EI

iPower

30X of Data Analysis Efficiency Within One Target Area

iLogistics iTransporation

50X of Delivery Capability per Vehicle through HUAWEI EI Path Optimization Service

1.3X of Vehicle Traffic Flow

Intelligent logistics Solution Panorama - Internal

SupplierHong Kong.

Raw material warehouseCountry-level warehouseSupply Center Site/CustomerCustomsPort

International transportation Last 1 km

The factory /EMS Intermediate transit point

Local distribution

Manage customs clearance servicesWarehouse management

(China)

Transportation&Shipping management (in China)

P/L

Material management

Delivery Management

P/L Monitoring

Warehouse management

(For server rooms outside China, country warehouse)

Transportation management

(Overseas)P/L Transport

solution

P/LProcessing forecast

Material preparation

forecast

Warehousing plan

Optimization of positions

Picking instruction

Document Identification

The customs compliance check

Classification of imported or exported

materials

Document Identification

Distribution plan

Goods quantity based on

estimation

Estimated Packing

It is recommended that the transport

route

Path optimization

Transport resource capability analysis

Transportation cost estimate.

Bid Risk Warehousing plan

Intelligent packing

Document Identificatio

n

Path planning

Exceptional expenses

30 %.

The packingrate

Increasedby 15%

Import efficiency

10X

Risk early-warning rate

99 %.

Operation efficiency

10%

Intelligent logistics service application – Results

Library XX disease

Diagnostic Screening

Disease factor analysis

Chronic disease management

Application and treatment.

Survival, and its prognosis.

HIS CIS EMRHand hemp.

..

The service bus

HL7 engine

Patients with uniform flag.

Structured Unstructured documents

Access controlRepository

engine

Dictionary service

Semi-structured documents.

Privacy protection

PACS

Drug research

Assists Clinic

Correlation analysis

Fuzzy clustering

algorithm and category

PredictionAnomaly detection

Medical service quality

managementMedical Research

Based on the medical data, analysis of disease, symptoms, relationships between departments and build a big data application

Application of correlation analysis

Disease star charts

Intelligent guiding

High blood

pressure.

Department of Cardiovascular Medicine (43%)Cardio-cerebral surgery (37%)

Internal Medicine (6%)

Others (4%)

Auxiliary diagnosis

High blood

pressure.705 Keywords

coronary heart

disease876

Indigestion.1492

Precordial pain

Shortness

Abdominal

distentionLeft

shoulder pain

I cann't sleep

Tinnitus

Fatigue

I have a headac

he.

Constipation

— — health data has widely application prospect, disease screening, diagnosis, prognosis and epidemic analysis, management optimization

Huawei and Phillips are enabling the Digital Health Solutions in China

Huawei Radiographic AI ranks third in global configuration in lung nodus

BSI-сенсоры и собственные чипы обработки Huawei со специально разработанными программными алгоритмами

Серия камер Huawei H.265 с высокой светочувствительностью

BSI CMOS удваивает светочувствительность

ИИ в системах безопасности

Видеосинопсис

Система быстро анализирует архивную запись за определённый период времени по одному видеоканалу и выдаёт “видеосинопсис” – все

детектированные по определённым критериям объекты в одном кадре или коротком видеоролике. Найденная выборка может быть

подвергнута повторному отсеиванию по новым аналитическим критериям.

1 hours

5 minutes

Интеллектуальная фильтрация и поиск по критериям

Увеличение эффективности многофакторного поиска и фильтрации результатов с помощью

моделирования фона и анализа метаданных

Анализ

Мо

дел

ир

ован

ие

фо

на

Исходное видео

Фоновое изображение

Передний план (движение)Проверка типа объекта – ТС

или пешеход

Ручная фильтрация

Ан

али

з м

етадан

ны

х Цветовые параметры Результат в видео списка

Охранные видеоаналитические алгоритмы

Анализ и отслеживание объекта наблюдения в реальном времени с автоматическим превентивным оповещением оператора при фиксации подозрительного поведения согласно

заданному критерию проверки

Вторжение в зону

Обнаружение праздношатания

Подсчёт людей

Пропавший объект

Детекция слишком быстрого движения

Оставленный объект

Движение в запрещённом направлении

Детекция образования толпы

Прорисовка маршрута движения

Пересечение виртуальной линии

20

2035 40 45

5555 60

6065 65 70 75 80

5535 25

4535 30 25

15 20 20 20 15

10101010101010

10252015

India

5

South Korea

5

Japan

5 5

Sweden

5

Germany

5

5

55

China

Singapore

5

5

5

Netherlands

Belgium

UK Italy

France

USA

5

5

5

New mobility formats have high acceptance specifically in Singapore, China and India

Mobility behavior: Mobility usage as a percentage of distance driven

"Think about the last couple of weeks, what % of distance (miles or kilometers) driven did you useCar sharing, ride hailing, taxi vs. Public transport vs. Own car/a car of a friend vs. other?"

Demand-driven modes (like car sharing, ride hailing, taxi)

Public transportation

Private car

Other

Participants by country: Belgium 1,019; China 1,033; France 1,025; Germany 1,039; India 1,019; Italy 1,017; Japan 1,016; Netherlands 1,006; Singapore 1,009 ; South Korea 1,007; Sweden 1,048; UK 1,015;

Ratio of shared vehicles by far highest in Singapore and China –All other regions with very low demand for share new vehicle

Amount of shared vehicles: Share of shared vehicles on car parc [in %]

4,0

9,9

1,1

2,3

7,1

2,8

0,90,10,10,30,20,20,2 0,50,50,70,80,7

China

0.1

4.0

0.1

South Korea

0.0

0.4

2.4

India

0.10.4

1.4

Netherlands

0.00.3

1.1

USA

0.0

1.1

Sweden

0.0

0.9

UK

0.0

0.7

Germany

0.0

0.70.4

France

0.0

<0.7

0.4

Japan

0.0

0.50.4

Italy

0.0

0.5

0.0

11.1

0.3

Singapore

0.0

10.2

Belgium

Taxi and on demand

Car sharing

Car Rental

Core capabilities: Uses sensors and computing

devices of better economics, combined with 5G core

technologies such as sliced network and edge

computing, to replace the high-cost sensor locating

technologies such as 64-laser lidar and expensive

computing devices. In this way, build a connected

automatic driving system that integrates big data,

artificial intelligence, and the Internet.

Sense Fusion Suite + TITAN Domain Controller +

ATHENA Software

Integrated, Multi-Dimensional

Autonomous Driving Application Solution

Chip-level solutionAutonomous driving AI algorithmAI hardware

acceleration

Connected

communication

Board-level solution

Collaboration

Plan

Automotive-grade

hardware computing

board

Autonomous driving software

app suite

ATHENAMDC

Sensor Fused Locating Based on Lidar + Camera

基于激光雷达和摄像头的融合定位

Locating Algorithm Based on Lane Detecting

Trajectory-Level Planning

Successfu

l Practices

Agenda


решения




Atlas Intelligent Computing Platform Portfolio

Atlas 200 Atlas 800

AI Accelerator Module AI Appliance

Atlas 300

AI Accelerator Card

Atlas 500

AI Edge Station

G5500

Heterogeneous Server

G2500

Intelligent Video Analytics Server

GPU FPGANPU

• Atlas 200 AI Accelerator Module

Cameras Drones Robots

• Powered by the Huawei Ascend

AI Processor, it supports real-time

analysis of 16-channel HD videos

in the size of half a credit card

• Positioned for terminal devices such

as cameras and drones, and edge

devices such as AI edge stations, it

has a thermal design power (TDP) of

only 13 W

16 TOPS INT8 @ 13 W

Supports real-time analysis of 16-channel HD videos,

and JPEG encoding/decoding

8 GB memory, PCIe 3.0 x4 interface

Operating temperature: -25°C to +80°C

Dimensions: 52 mm x 38 mm x 10.2 mm

Large Capacity in

a Small Size

Built for Intelligent

Devices and Edge

• Provides 64 TOPS of INT8 and

64-channel HD video real-time

analysis on a single card, fueling

deep learning and inference with

superior compute power

• Supports JPEG and video hardware

encoding/decoding, delivering a leap

in image and video application

performance

• Large memory capacity and high

bandwidth, meeting memory

requirements in feature matching

scenarios and reducing application

latency

64 TOPS INT8 @ 75 W

Supports real-time analysis of 64-channel HD videos,

and JPEG encoding/decoding

32 GB memory, 204.8 GB/s memory bandwidth

PCIe 3.0 x16, half-height half-length card

Superior

Compute Power

Hardware

Encoding/Decoding

• Atlas 300 AI Accelerator Card

64 channels per card

Best choice for high-density video inference

• Atlas 200 Developer Kit

• Inspired by the Huawei Ascend AI

Processor, it integrates rich

peripheral interfaces and the

MindStudio software to help

developers quickly familiarize with

the development environment

• The MindStudio provides a user-

friendly programming interface and

graphics-based debugging capability,

allowing automatic management of

offline models with a simulation

environment

16 TOPS INT8 @ 24 W

• 1 USB type-C | 2 camera interfaces | 1 GE network

port | 1 SD card slot

8 GB memory

Operating temperature: 0°C to 45°C

Dimensions: 125 mm x 80 mm x 24 mm

High Integration Easy-to-Use Software

EnvironmentQuickly build development environment

in 30 minutes

• Atlas 500 AI Edge Station

AI Edge Station

Safe City | Smart Transportation |

Unattended Retail

Intelligent Edge

• An industry-leading edge

product that integrates AI

processing capabilities

• Works stably at -30°C to +60°C

outdoors without fans

Large Capacity in a

Compact Size

• Capable of processing 16-

channel HD videos in the size of

an STB

• Delivers a 4x performance lead

over competing products in the

industry

Edge-Cloud Collaboration

• Works with Huawei private cloud

and HUAWEI CLOUD, receiving

applications and updated

algorithms pushed from the cloud

• Unified device management and

firmware upgrade on the cloud

• AI Edge Station for Various Edge Applications

Smart Transportation(Traffic light tuning and intelligent traffic diversion)

Smart Manufacturing(Intelligent quality inspection and flexible

manufacturing)

Smart Security Surveillance(Boundary violation and compliance detection,

detention center surveillance)

Intelligent Care(Kindergarten and elderly care)

Unattended Retail(Unmanned retail and smart store)

Safe City(Facial recognition and license plate recognition)

• Atlas 800 AI Appliance

AI Appliance

Deep Learning | Model Training |

Recommendation System

Out-of-the-Box

• Ready to work in 2h with pre-

installed AI development

environment, underlying software

library, and development

framework

• Automatic generation of AutoDL

models, hyperparameter tuning,

and one-click model deployment

Ultimate Performance

• Provides an optimized AI environment

based on the standard framework and

programming environment

• Unlocks ultimate performance with

GPU scheduling algorithms, with over

15% resource utilization

Integrated Management

• Comprehensive management

of GPU utilization, health

monitoring, and job

scheduling

• Easy-to-use WebUI, more

intuitive and efficient than the

command line interface (CLI)

THANK YOU

Copyright©2018 Huawei Technologies Co., Ltd. All Rights Reserved.

The information in this document may contain predictive statements including, without limitation, statements regarding the future financial and operating results, future product portfolio, new

technology, etc. There are a number of factors that could cause actual results and developments to differ materially from those expressed or implied in the predictive statements. Therefore, such

information is provided for reference purpose only and constitutes neither an offer nor an acceptance. Huawei may change the information at any time without notice.

Оптимизированные решения huawei и … for mfti v1.pdfВклад в r&d...

Documents