APPLICATION OVERVIEW Foundation: protects and stabilizes ...
Digital Annealer Introduction - fujitsu.com · Annealing Method Classical Approach Metal is heated...
Transcript of Digital Annealer Introduction - fujitsu.com · Annealing Method Classical Approach Metal is heated...
Copyright 2019 FUJITSU LIMITED
Digital Annealer
Introduction
December 21, 2018
FUJITSU Quantum-Inspired Computing Digital Annealer
Patent pending including related technology
Fujitsu Limited
Quantum Computing is one promising prospect as a next generation computer
Copyright 2019 FUJITSU LIMITED
1990 2010 20202000
Data Volume
103
102
1
101
104
10nm
20nm40nm
180nm
250nm
90nm
Today
GAP
UnstructuredData
StructuredData
350nm
2025
2002
Moore’s Law foretells of improved CPU density and performance
Performance improvement hasslowed due to power constraints
The limit ofMoore’s Law
Power efficiency performance of general CPU “Relative value”
Reaching the Limit of Moore’s LawComputers must process increasingly massive and complex data at higher and higherspeeds in order to support digital transformation in society and business. Moore’s Law* is approaching its limit, threatening the drastic compute performance required in the coming future.
*Moore's Law: An empirical rule in the semiconductor industry stating that the number of transistors in a dense integrated circuit doubles every 18~24 months.
Power Efficiency
Performance
Year
1
Copyright 2019 FUJITSU LIMITED
Digital AnnealerA new architecture that solves "combinatorial optimization problems" at high speed with digital circuits inspired by quantum phenomena
Stable operation with digital circuit, and easy miniaturization
Easy mapping of more complex problems with a fully-connected architecture
Digital Annealer
Difficult to maintain a quantum state
Limits in connection and expansion
Quantum ComputersEasy to apply to actual problems
Still in the research stage ...
2
Copyright 2019 FUJITSU LIMITED
Digital Annealer Positioning
Quantum Computing Technology
DevelopmentOrganization
Method
Stable operation atroom temperature
Available in traditionalrack / server room
Requires cryogenic environment, but still technically unstableLarge cooling device required
Quantum Architecture
FujitsuDigital Annealer
D-Wave systemsD-Wave 2000Q
NECHitachi, Ltd.
CMOS AnnealingMachine
Non-Quantum Architecture
Google Microsoft IntelIBM
IBM Q
Quantum Gate SystemComputation using quantum gates and
quantum bits that can be in a superposition of 0 and 1 states
量子イジングマシン方式
Superconducting Circuit
Ising Machine SystemMap a problem to a mathematical model of interacting magnets, then search for solutions
Specialized for combinatorial optimization problems
Annealing
Digital Circuit
LaserNetwork System
National Institute of Informatics, NTT and others
Japan Cabinet Office ImPACT"Quantum Neural Network"
Optical Parametric Oscillation
Quantum Annealing
Analyzed by Fujitsu
Unlike quantum computers, Digital Annealer does not require an extremely low temperature environment. Digital Annealer operates stably at room temperature.
Digital Annealer makes use of the annealing method, specialized for combinatorial optimization.
3
Annealing Process in Metallurgy
Copyright 2019 FUJITSU LIMITED
What is Quantum Annealing? The algorithm’s name comes from the annealing process used in metallurgy
Blocks are placed randomly, then the entire system is“shaken”. The shaking is gradually reduced, and the shapes quickly fit together.
Blocks are placed in sequence. Process restarts if a solution is not found. Repeated until a solution is found.
Annealing Method Classical Approach
Metal is heated to a high temperature, and the structure stabilizes as it is slowly cooled(=low energy)
The most stable state has minimum energy The minimum value
4
Choose the number of cities Visit each city only once
Evaluate based on travelling distanceShortest route
Copyright 2019 FUJITSU LIMITED
Combinatorial Optimization ProblemsSeek combinations or sequences that satisfy given constraints, with the goal of finding
the best out of all available combinations
City CityCity
City
City CityCity
City
City CityCity
City
With 5 cities 120 possible routes. With 32 cities 2.63x1035 possible routes
Number of cities: N Constraint Optimal Solution
“Which is the shortest route that visits each city exactly once and returns to the origin city?”
Example: The Traveling Salesman Problem
The number of combinations increases exponentially!5
The resulting bits show the shortest path solution
City B ⇒ City A ⇒ City D ⇒ City C ⇒ City E
Create an expression to define the interaction between bits (distance
between cities)
Use Digital Annealer to Solve Combinatorial Optimization Problems
“Which is the shortest route that visits each city exactly once and returns to the origin city?”
INPUT OUTPUT
Formulation
City A City B City C City D City ECity A City B City C City D City E
0th
1st
2nd
3rd
4th
Cities to Visit
0th
1st
2nd
3rd
4th
0
1
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
Cities to Visit
Each represents one bit
Order to Visit
The Traveling Salesman Problem
Copyright 2019 FUJITSU LIMITED
3. Execute on Digital Annealer2. Build Formula1. Capture Problem in Ising Model
6
Applicable to New Areas
Improve Precision and Reduce Time to solve existing combinatorial
optimization problems
Copyright 2019 FUJITSU LIMITED
Combinatorial Optimization Problems Across All Industries & Business
Advanced Healthcare
New Material Development
Distribution RoutePlanning
UtilitiesManagement
Demand Forecasting
Materials & Compounds
Network ConfigManagement
LogisticsHospitality
Management
DrugDiscovery
Portfolio Management
DigitalMarketing
Applicable Area Examples
AutonomousVehicles
7
Copyright 2019 FUJITSU LIMITED
Problem and Application Examples Well-Suited for Digital Annealer
Optimization Problems
Maximum IndependentSet Problem
Minimum Set Cover Problem
Candidate Overlay Problem
Max Cut Problem
1. Molecular Similarity Search
2. Big Data Visualization Toolkit
3. Traffic Route Optimization
4. Financial Portfolio Optimization (QHRP Method*1)
5. Shelf Location Optimization
Molecular similarity identificationin compounds
Visualizing and clustering big data
Optimize route selection while minimizing route overlap to ease congestion
Select investment portfolio assetsbased on low correlativity
Optimize shelf placement to improve factory parts pickup
*1 QHRP: Quantum-inspired Hierarchical Risk Parity
8
Copyright 2019 FUJITSU LIMITED
1. High Precision Molecular Similarity Search for Drug Discovery
The conventional Finger Print method determines the presence or absence of an atomic group, but does not consider the molecular shape. Thus, a precise search cannot be performed.
By converting the molecular structure to a graph and handling atomic groups as nodes and bonds as edges:• Precision is improving by considering molecular shape• Calculations are performed at high speed by Digital Annealer
• Highly precise molecular similarity search becomes possible• Expected to improve the efficiency of drug development leading to new
highly effective medicines
Issues
Technique
Results
Use Case
Combinations of atomic groups results in a huge volume of calculations
Improve efficiency with Digital Annealer
Conventional Method
Digital Annealer Benefit
Finger Print
Matching without consideration of molecular shape
Graph Theory
Highly accurate by matching molecular
shapes
Chemical & Pharmaceutical
Contributing to the development of highly effective medicines
Finger Print: A method of representing the presence or absence of an atomic group as 0 or 1 and expressing the molecule as a Boolean vector
9
Copyright 2019 FUJITSU LIMITED
2. High-Speed Clustering for Big Data Utilization
As the importance and prevalence of big data increases, high-speed data processing is necessary to effectively derive business insights
High precision clustering with hierarchical structures is implemented by compressing high dimensional data and segmenting it into portions that can be clustered
• Clustering is accelerated from several hours with conventional methods to just a few minutes with Digital Annealer
• Large scale data sets can be visualized and analyzed• The level of clustering can be changed to enhance analysis
Issues
Technique
Results
Use Case
Marketing
Visualizing large-scale datasets for more accurate analysis
Conventional Digital Annealer
Customer data for 40,000 people Grouped by similar data
Highly accurate clustering with Digital Annealer
Changeable according to the level of clustering
Coarse grained
Fine grained
10
Copyright 2019 FUJITSU LIMITED
3. Route Optimization to Reduce Traffic Congestion
Total Travel time
Shortest Route
With Digital Annealer
With conventional routing systems, there is a tendency to assign the shortest distance route, leading to traffic congestion in the city center
• Optimize route selection to avoid overlap• Prioritize route options by adding conditions, such as: speed
limits, number of lanes, etc.
• Reduce traffic congestion by up to 40% by dispersing traffic• Apply to cases of iterative simulation used for road development planning• Applicable to other routing problems, such as warehouse collection and
distribution, AGV (Automated Guided Vehicles), and network traffic 0
50
100
150
DigitalAnnealer
58hrs
Total Travel Time
101hours
58hours
Issues
Technique
Results
Use Case
LogisticsReduce overall travel time by distributing routes throughout a city
or factory to avoid congestion
Conventional Digital Annealer
Select Shortest Path
Optimized Route Selection
Degree of Congestion
High
Low
Example Max
Speed (km/h)
10
20
40
50
60
Degree of Congestion
High
Low
Example Max
Speed (km/h)
10
20
40
50
60
Shortest Route
Method101hrs
40% Decrease
Immediately recalculate routesaccording to traffic condition changes
11
Copyright 2019 FUJITSU LIMITED
4. Investment Portfolio Optimization Through Risk Diversification
The commonly used Minimum Variance (MV) method for portfolio optimization is susceptible to the influence of market fluctuation
Quantum-inspired Hierarchical Risk Parity (QHRP) portfolio optimization method provides for:• The clustering of assets into a tree diagram based on risk
correlations • Composition of risk-diversified portfolios with low correlativity
• Create portfolios with resistance to market fluctuations that continue to provide stable returns
• 60% higher Sharpe Ratio compared to MV method
Issues
Technique
Results
Use Case
Finance
Instant clustering for the correlation of 500 stocks to compose a risk-resistant portfolio
Risk/Return Ratio
Results (Sharpe ratio comparison)
Digital Annealer Technique
Clustering Assets from CorrelationsAsset Price Change Variance Matrix
Company
ACompany
BCompany
C …
Company A 1 0.23 0.85
Company B 0.23 1 0.64
Company C 0.85 0.64 1
…
…
Express all correlations among assets(price changes between two assets) using
variance-covariance matrices Companies: A C D B
Lehman Shock
QHRP
MV
S&P500 Investment Results (2005 – 2011)
12
Copyright 2019 FUJITSU LIMITED
5. Factory Parts Pick Up Optimization
• High-mix, low-volume factory production requires a large variety of parts for each product. Time and labor required dependent on the experience level of each worker
• Inconsistent and inefficient parts pick up process
• Routes and shelf population are minimized as combinatorial optimization problems
• Correlation of frequently used shelves identified
• Even inexperienced workers can realize efficient parts picking• Travel distances reduced by up to 45% per month through route and
shelf location optimization• Optimization methods to be deployed to other factories, as well as
other processes such as warehouse management
Issues
Technique
Results
Use Case
Manufacturing
Reduce travel distance for warehouse parts pick up by up to 45%Now in use at Fujitsu IT Products
Very complicated parts picking routes required experienced workers
Digital Annealer provides optimum picking routes displayed on a tablet
Warehouse area: 1000m²Number of parts: 3000
13
Copyright 2019 FUJITSU LIMITED
5. Case Study:
Factory Parts Pick Up Optimization - At Fujitsu IT Products
Case Study: http://www.fujitsu.com/global/digitalannealer/case-studies/201804-fjit/
Company Profile
Digital Annealer Project Schedule
Manufacturing
Company Name
Location
Capital
Establishment
Industry
Employees
Fujitsu IT Products Limited
1-1, Kasajima-to, Kahoku-shi, Ishikawa, 929-1196, Japan
100 million yen (wholly owned subsidiary of Fujitsu Limited)
April 1, 2002 455 people
Manufacture of servers, supercomputers, storage systems , software, etc.
Servicein February 2018
PoC fromOctober 2017
Discussion fromSeptember 2017
Source: http://www.fujitsu.com/jp/group/fjit/ (Japanese)
14
Copyright 2019 FUJITSU LIMITED
5. Case Study:
Factory Parts Pick Up Optimization - ConfigurationManufacturing
Before
With Digital Annealer
Tablet
Request a formula and distance between shelves from Digital Annealer via REST-API
ManagementServer
Optimum route provided
a b c d e f
a 0 1 2 3 6 7
b 1 0 1 4 5 6
c 2 1 0 5 4 5
d 3 4 5 0 4 4
e 6 5 4 4 0 1
f 7 6 5 4 1 0
①
②
③
④
⑤ ⑥
Shelf positionDistance
between shelves
Part1: Shelf aPart3: Shelf cPart6: Shelf f
Shelf ac: 2Shelf cf: 5Shelf af: 7
Shelf Position DB
Output Optimal route
Pick up done based on the best route
Shelf positions and distances between shelves are put in the Shelf Position DB
1
2
3
4
5
Tablet
Shelf a → c → fPart 1 → 3 → 6
Shelf c→f→aPart 3→6→1
Shelf ac: 2Shelf cf: 5Shelf af: 7
+ Mathematical Expression
Newly Implemented
Digital AnnealerCloud Service
0
1
1
1
0
0
1
1
Digital Annealer
Shelf Position DBComponent Parts DB
Part1: Shelf aPart3: Shelf cPart6: Shelf f
Part1: Shelf a Part2: Shelf bPart3: Shelf c Part4: Shelf dPart5: Shelf e Part6: Shelf f
Product A-Part 1-Part 3-Part 6
Order Input
Product A
Sales Factory
Goal: Leverage existing operation, and create a new shelf position databaseto use Digital Annealer to calculate optimal pick up routes
Based on information displayed on the tablet, parts picked up
15
Copyright 2019 FUJITSU LIMITED
Easy to use with total bit couplingConnectivity
High precision with 64bit graduationsPrecision
Stable operation at roomtemperature with digital circuits
Stability
Applicable to real world problems with the stability and
balance of scale, connectivity, and precision
Digital Annealer Advantages
Ready to scale up for 8,192bit problemsScale
16
Scale Connectivity Precision
Copyright 2019 FUJITSU LIMITED
What Are Scale, Connectivity and Precision?
The Traveling Salesman Problem
City CityCity
CityUnder the constraint that each city must be visited only once, find the shortest route (minimum distance)
Scale:Number of cities to visit
that can be handled
Connectivity:All distances betweencities can be defined
Gradations:Accuracy of distance
between cities
8192 bits Total Coupling 64bit Gradations
0th
1st
2nd
3rd
4th
City A B C D E
0th
1st
2nd
3rd
4th
VisitOrder
Cities to Visit
0th
1st
2nd
3rd
4th
Near Medium Far
VisitOrder
VisitOrder
For example, 3 gradations can be expressed as near, medium and far.
Cities to VisitCities to Visit
City A B C D E City A B C D E
17
Copyright 2019 FUJITSU LIMITED
Digital Annealer Features
Differe
ntia
tors
Higher escape probability when local solution is detected
StateOptimal Solution
Local Solution
Offset
Escape!B
C
A
0
1
1
1
0
0
1
0
Bit 1
Bit 2
…
Bit i
…
Total Coupling
• Increased escape probability from local minimum energy states using hardware offsetting
• Represent large-scale problems of up to 8192 bits
• High precision with full coupling and 264 bit gradations
• Easy to use with fully connected bits
• Next bit inversion found using parallel search to increase processing speed
• Stochastic parallelism provides significant acceleration
Current set of bits
Parallel Search
Set bits aftersingle bit updated
・・・
Trial1 Trial 2 Trial N
Evaluation
Bit 1Inversion
Bit 2Inversion
Bit NInversion
Evaluation Evaluation
Eval
uati
on V
alue
Ready for Many Business Tasks
High Speed Processing withLocal Solution Escape and Parallel Search
18
Copyright 2019 FUJITSU LIMITED
Specification ComparisonApplicable to real world problems with the stability and balance of
scale, connectivity, and precision
Digital Annealer(2nd generation)
Company A Company B
ImplementationTechnology
Digital Circuit(Using existing technology)
Superconductive Circuit(Cryogenic cooling required)
Non-Linear Optical(1km ring device)
Number of Bits 8192 2048 2048
Amount of Coupling Total Coupling6 - Partly Coupled(64bit total coupling
equivalent)
Total Coupling
EvaluationAccuracy 64bit Gradations 32 Gradations 3 Gradations
19
Combinatorial Optimization Problems Faced in Actual Business
Copyright 2019 FUJITSU LIMITED
Digital Annealer Service OverviewCloud Service and Technical Service launched in May 2018
TechnicalService
CloudService
On-Premises(planned)
FUJITSU Quantum-Inspired Computing Digital Annealer Cloud Service Digital Annealer Technical Service
Realize high speed processing forcombinatorial optimization problems
Customer Support forDigital Annealer utilization
21
Digital Annealer Cloud ServiceFUJITSU Quantum-Inspired Computing Digital Annealer
Ising model
Mathematical expressions
Variables
Constraint conditions
Digital AnnealerCloud Service
Customer
WebAPI
WebAPI
Formulation
0
1
1
1
0
0
1
1
Find Optimal Solutions
Request
Response
Netw
ork
Via Web API, submit Ising model mathematical expressions and constraints,
then receive solutions in seconds
Copyright 2019 FUJITSU LIMITED22
Copyright 2019 FUJITSU LIMITED
*It may take time to start calculating the mathematical model.
Help Desk Service is mandatory for all Service Levels
Cloud Service MenuFUJITSU Quantum-Inspired Computing Digital Annealer
1st generation: 1024bit 2nd generation: 8192bit
Service Level Gen Fee Category Web API Type Remarks
Premium 1st Fixed monthly Dedicated SynchronousPlease contact Fujitsu in advance
to schedule environment deployment
Standard
1st Basic fee (fixed monthly)+metered rate (by usage)
Shared*
Synchronous
Asynchronousoptional
2nd
Selectable
Basic fee (fixed monthly)+metered rate (by usage)
Fixed monthly
Trial
1st Basic fee (fixed monthly)+metered rate (by usage)
Shared*
Synchronous
Asynchronousoptional
Limited to 6 months2nd
Selectable
Basic fee (fixed monthly)+metered rate (by usage)
Fixed monthly
Academic1st
Fixed monthly Shared*
Synchronous
Asynchronousoptional
Limited to universities andgraduate schools for educational
and research purposes2nd
23
The API requests calculation results synchronously.The API returns when the calculation process has completed.
A calculation request is issued and the API returns. A separate request is issued to obtain the calculation result.
Cloud Service Web APIs FUJITSU Quantum-Inspired Computing Digital Annealer
Synchronous Web API (basic) and Asynchronous Web API (optional) offerings:
Multiple requests can be made for large problemsThe problem is processed in real time
Synchronous Web API Asynchronous Web API
Ising Model Mathematical
Expressions Variables Constraint
Conditions
Digital AnnealerCloud Service
CustomerRequest
OptimalSolution
…
Ising Model Mathematical
Expressions Variables Constraint
Conditions
CustomerRequest
0
1
1
1
0
0
1
1
Digital AnnealerCloud Service
0
1
1
1
0
0
1
1…
Netw
ork
Netw
ork
OptimalSolution
Response(with Acceptance ID)
Request for result corresponding to ID
Copyright 2019 FUJITSU LIMITED24
Copyright 2019 FUJITSU LIMITED
Help Desk ServiceFUJITSU Quantum-Inspired Computing Digital Annealer
Service Details for Digital Annealer Cloud Service Subscribers:
Weekdays 9:00-17:00 (JST) *Closed on Japanese holidays
Inquiries received outside service hours will be processed after 9:00am on the following business day
The following inquiries are excluded from the Help Desk Service:・Processing speed tuning (performance evaluation resulting from customer’s APP design, implementation and operation, etc.)・Consulting (advice on creation, design, implementation and operation of mathematical models)・Disclosure of information related to our cloud service environment and logs・Calculation result accuracy
Service Hours
• Questions about specifications, settings and usage of the Digital Annealer Cloud Service• Questions, investigation, workarounds when the Digital Annealer Cloud Service does not function correctlyAcceptable Questions
EmailInquiry Contact Method JapaneseLanguage
Important Notes
Calculated as 5% of the Web-API usage service fee Service Fee
Trouble Notification Maintenance Notification Service Detail / Update Guide
Detail
Method
Notification from the time the issue occurs, functions impacted, etc.
Portal/Email
Detail
Method
Detail
Method
Notification of Digital Annealer Cloud Service planned or emergency maintenance date/time, functions impacted, etc.
Portal/Email
Notification of new functionality and improvements made to the Digital Annealer Cloud Service
Portal/Email
25
Digital Annealer Technical Service
Complete support from introduction to operation, leveraging Digital Annealer for customer business
Digital Annealer Technical Service - OutlineDigital Annealer Technical Service
Technical experts with advanced mathematics knowledge and data analysis capabilities support you in resolving your optimization problems
Training for EngineersTraining
Tuning Support ServiceTuningSupport
Convert customer issues into formula
Apply formulation to customer issues
Applicationimplementation
Maintenance,Support
Assessment Deployment Support Implementation Operation
Copyright 2019 FUJITSU LIMITED
Operation PhaseImplementation Phase
26
Copyright 2019 FUJITSU LIMITED
Digital Annealer Technical Service DetailsDigital Annealer Technical Service
Verify the customer’s problem can be converted to a combinatorial optimization problem, and generate a mathematical model to solve with Digital Annealer.
PoC planning for customer Digital Annealer use. Evaluate mathematical model implementation and results. Support requirements definition for customer Digital Annealer deployment.
Develop and deploy application for use with Digital Annealer: Deploy on customer systems, establish connection to Digital Annealer, process input / output data, and perform post processing of output.
Formulation Verification Development Support Implementation
Answer questions and solve problems for the deployed Digital Annealer application.
F2F instructor-led Digital Annealer training:• Basic: Digital Annealer features and
mechanisms•Hands-on: Solve optimization
problems using Digital Annealer Cloud Service
Problem solving support including parameter tuning for Digital Annealer Cloud Service customers.
Operation Training Tuning Support
27
Combinatorial Optimization Problems Faced in the Real World
Copyright 2019 FUJITSU LIMITED
Complete Solutions for Real-World Business Problems
Issues Solutions
Ising ModelMathematical Formula
Find OptimalSolution
OptimizationCalculation
BusinessApplication
Technical ServiceCloud ServiceTechnical ServiceFormulation
Analysis/Verification
28
Copyright 2019 FUJITSU LIMITED
Digital Annealer Project Example
Typical Project Duration: 4-6 months
• Procedure Agreement
• Conclude NDA (As needed)
• Pre-meeting(2-3 times)
• Technical Introduction• Technical Verification
• Select target business• Agreement on Expectations
• PoC Plan Agreement
• Formulation• Ising Model Construction
• PoC on Digital Annealer-Confirm Functions-Performance, UX, etc
• PoC Result Summary
Available Digital
Annealer Services
2~3 months
NextStepPoCModel ConstructionSelect business
target for PoCTechnical Check
Pre-PhasePhase 1 Phase 2 Phase 3 Phase 4
2~3 months
Trial: Maximum 6 months (Fee-based)
Technical Service (Fee based)
Premium/Standard (Fee based)
29
A new service in thequantum computing area
Solving combinatorial optimization problems to control Automated
Guided Vehicles (AGVs)
Touhoku University
New marketing technology research and development for real-world business problems
RecruitCommunications
Co., Ltd.
Technical verification of portfolio optimization in asset
management
Mitsubishi UFJ TrustInvestment Technology
Institute Co., Ltd.
Technical verification of production lines for high-mix
low-volume models in factories
Fujifilm Corporation
Fixstars Corporation
Copyright 2019 FUJITSU LIMITED
Digital Annealer Case Studies
Participation inthe MITOU Program
of METI / IPA
Using Digital Annealer to foster talent and start-up companies
in technical fields
30
Copyright 2019 FUJITSU LIMITED
Digital Annealer Roadmap
Scale: 1024 bit
Digital Annealing Unit
1st Gen Cloud Cloud2nd Gen
2018
Technical Service
May On-Premises Service 4QDec
2019
Expand applications from technical verification to real-world business value
Next Generation
Precision: 16 bit 65536 Gradations
Max Scale: 8192 bit
Max Precision: 64 bit1.845x1019 Gradations
Large-Scale Parallel Processing
32
Size (bit)
Precision (bit)
Company A
Company B
Company C
2nd Generation
Priority for Size
1st Generation
Copyright 2019 FUJITSU LIMITED
Priority for precision and priority for size
can be switched
16bit
64bit
Application Areas Surpassing the Competition - 2nd Generation
8192bit1024
bit
33
Accuracy
2nd generation
Priority for size
1st generation
1024bit
16bit
8192bit
64bit
Size
Precision
2nd Generation
Priority for Size
1st Generation
1024bit
16bit
1,000,000 bit
Supports up to 1,000,000 bitswith software partition technology & multichip support
Copyright 2019 FUJITSU LIMITED
Application Areas Surpassing the Competition – Next Generation
Develop further large-scale application technology2019
8192bit
64bit
34
Copyright 2019 FUJITSU LIMITED
Tackling Extremely Large Scale Problems
Hardware technology links multiple Digital Annealer Units (DAUs) to execute large scale parallel processing
DAU
Depending on the characteristics of the problem, multiple splitting
methods available
Find the optimal solutionusing a solution search flow
Large scale problems split into manageable portions by software
DAU DAU DAU
Fujitsu is developing a problem-splitting method to solve extremely large-scale problemsThis method extracts a portion of a problem according to problem characteristics; processing each portion on Digital Annealer, then assembling the output from multiple calculations to derive the total optimal solution using search flow.
This method allows a single 2nd generation (8192 bit) Digital Annealer to solve 100,000 bit scale problems
Further software and hardware enhancements will expand the scale to 1 million bits (planned for CY2019)
35
Drug candidate search can be drastically shortened from 6 months to just a few days,accelerating the development of middle molecule medicine
Copyright 2019 FUJITSU LIMITED
Accelerate the Development of Middle Molecule Medicine
Search for the most stable structure binding relationship within amino acids, when each amino acid is modeled and placed on a lattice point
Linear molecular structureof amino acids
Represents one amino acid
Search for optimal structure
Return to the molecularstructure
Amino acidtype
Apply Digital Annealer + the problem-splitting method for simulation to find stable structure middle molecule drug candidates jointly with ProteinQure Inc.
Effectiveness demonstrated with 30K bit scale problem
36
Digital Annealer Global Rollout
Japan
• Cloud Service • Technical Service
launched May2018
EMEIA
• Cloud Service (Planned)
• Technical ServiceCustomized Support
APAC &Oceania
• Cloud Service (Planned)
• Technical ServiceCustomized Support
Copyright 2019 FUJITSU LIMITED
Digital Annealer Cloud Service rolling out in each region starting from Japan
Technical Service is under development globally Americas
• Cloud Service (Planned)
• Technical ServiceCustomized Support
37
New AI Headquarters (AI HQ) - Established October 2018
Creating a core base in Vancouver to roll out Fujitsu's AI business and accelerate the application of AI to customers worldwide using Digital Annealer as a key technology
Americas
AIHQ
APAC
EMEIA Japan
Deployment of advanced global use cases
• Distribution of technology and solutions
• Expert technical support
Leading an AI ecosystem
• Expansion of platform users
Creation of AI core business
• Planning, execution, and roll out • Company-wide AI strategies
Oceania
Copyright 2019 FUJITSU LIMITED38
Copyright 2019 FUJITSU LIMITED
Expanding Digital Annealer Applications Through Partnerships
1QBit middleware implementedon Digital Annealer
Collaborative researchin cutting-edge areas
Confronting new issues in society
Digital Annealer
Promotion of combinatorial optimization as a way to
solve societal issuesExpansion of application areas
40
Digital Annealer
Copyright 2019 FUJITSU LIMITED
Partnership with 1QBit
HardwareThe processing power
to solve problemswith high speed and
high precision
MiddlewareMathematical formulas
and algorithms for computation
The world’s #1 vendor of software for quantum computers
Conducting joint business around the world
Planning to incorporate Digital Annealer in 1QBit Cloud Service in CY2019
41
Copyright 2019 FUJITSU LIMITED
Partnership with University of Toronto
Smart Traffic
Finance
Network
Medical
World-class research university in the fields of AI and quantum computing
New joint research center established with the University of Toronto in November, 2017:
Fujitsu Co-Creation Research Laboratory at the University of Toronto
Joint research in smart transportation, networks, finance, and healthcare fields
42
Copyright 2019 FUJITSU LIMITED
Joint Research with the University of Toronto- Cancer Radiation Therapy
Huge number of irradiation patterns (number of combinations) with variations such as range,
direction, and intensity of irradiation
Even when the beams are from only one direction, the number ofcombinations would be: 10150
In case of irradiation against a 1cm2 tumor with a precision of 1mm2 and 32 intensity levels from one direction
Huge computational load required for simulation before a treatment plan can be made.
Current technology needs multiple hours to a few days to calculate the combinatorial optimum.
Digital Annealer takes only a few minutes
Cancer
Vital Organs
Intensity Modulated Radiation Therapy (IMRT)
43
Copyright 2019 FUJITSU LIMITED
Partnership with Waseda University
Candidate Topics
Traffic/Logistics Delivery Optimization
DigitalMarketing Data Analytics
FinanceStock Price
Fluctuation Prediction
Comprehensive Collaborative Activity Agreement Established
Solve Real-World Societal Issues
Healthcare Disaster Remediation
Chemical, Pharma Traffic/
Logistics
ManufacturingEnergy
SecurityDigital
MarketingFinance
Private Company
Joint research into solving combinatorial optimization problems with Digital Annealer
ResearchTopics
Starting April 2019: Collaborative research using Digital Annealer to solve real-world combinatorial optimization problems
Drawing research topics from Waseda University’s entire body of research
Joint research site opened in March 2018: Fujitsu Co-Creation Research Laboratory at Waseda University
44
Copyright 2019 FUJITSU LIMITED
Digital Annealer Reference Sites & Press Releases
Digital Annealer Websites:
Japanese English
http://www.fujitsu.com/global/digitalannealer/
http://www.fujitsu.com/jp/digitalannealer/
Press Releases
Fujitsu Laboratories and Waseda University Agree to Comprehensively Collaborate on Digital Annealer Research
September 19, 2018Fujitsu Technology to Solve Combinatorial Optimization Problems for Medium-Sized Drug Discovery
September 18, 2018Fujitsu Quantum-Inspired Digital Annealer Cloud Service to Rapidly Resolve Combinatorial Optimization Problems
May 15, 2018
Fujitsu Initiates Joint Research with Recruit Communications on Marketing Technologies Using "Digital Annealer"
January 29, 2018
Exhibited at "CEATEC JAPAN 2017“
September 27, 2017Fujitsu Technology Facilitates Application of Combinatorial Optimization Methods to Real-World Problems
September 20, 2017
Fujitsu Laboratories and University of Toronto Enter Strategic Partnership
September 20, 2017Fujitsu and 1QBit Collaborate on Quantum Inspired AI Cloud Service
May 16, 2017Fujitsu Laboratories Develops New Architecture that Rivals Quantum Computers in Utility
October 20, 2016
http://www.fujitsu.com/global/about/resources/news/press-releases/
YouTube
Digital Annealer public channel
http://www.youtube.com/channel/UCo0c9YwYOHXLwJnNA_moEJQ
Fujitsu and TC3 Promote Quantum Inspired Digital Annealer Next-Generation Architecture in Topcoder Contest
January 15, 2019Fujitsu Launches Next Generation Quantum-Inspired Digital Annealer Service
December 21, 2018Fujitsu Drives Quantum-Inspired Project to Help Solve NatWest’s Complex Optimization Challenges
October 2, 2018
45