hyperion research update: research highlights in hpc, … · 2020-01-09 · hyperion research...
TRANSCRIPT
HYPERION RESEARCH UPDATE:
Research Highlights In HPC, HPDA-AI, Cloud
Computing, Quantum Computing, and
Innovation Award Winners
Earl Joseph, Steve Conway,
Bob Sorensen, and Alex Norton
ISC19
Important Notes
▪ If you want a copy of the slides:• Please register at our homepage
(www.HyperionResearch.com ):
→ or leave a business card
▪ Check out our web sites:
• www.HyperionResearch.com
• www.hpcuserforum.com
© Hyperion Research 2
Agenda
1. Example Hyperion Research Projects
2. Update on the HPC Market
3. Cloud Computing Update and Our New
Scorecard Tool
4. Quantum Computing Update
5. The Exascale Race
6. The ISC19 Innovation Award Winners
7. Conclusions and Predictions
© Hyperion Research 3
The Hyperion Research Team
Lloyd Coen
Earl Joseph
Steve Conway
Bob Sorensen
Alex Norton
Mike Thorp
Kurt Gantrish
Jean Sorensen
Tom Christian
Nishi Katsuya
Research studies & strategic consulting
Strategic consulting, HPC UF, Big Data, AI
Special studies, new data analysis, clouds
Business manager
Survey design & executive interviews
Japan research and studies
Global sales management
Global sales management
Strategic research, government studies, QC
4© Hyperion Research
Hyperion Research HPC Activities
• Track all HPC servers sold each quarter • By 28 countries
• 4 HPC User Forum meetings each year
• Publish 85 plus research reports each year
• Visit all major supercomputer sites & write reports
• Assist in collaborations between buyers/users and vendors
• Assist governments in HPC plans, strategies and direction
• Assist buyers/users in planning and procurements
• Maintain 5 year forecasts in many areas/topics
• Develop a worldwide ROI measurement system
• HPDA program (includes ML/DL/AI)
• HPC Cloud usage tracking
• Quarterly tracking of GPUs/accelerators
• Cyber Security
• Quantum Computing
• Map applications to algorithms to architectures
5© Hyperion Research
The HPC User Forum: www.hpcuserforum.com
6© Hyperion Research
HPC User Forum Steering Committee
7© Hyperion Research
Plan On Joining Us This Year:2019 HPC User Forum Meetings
(www.hpcuserforum.com)
Plus China in August
September 9-11
Argonne National Laboratory
(Chicago area)
© Hyperion Research 2019 8
October 7-8
CSCS: Swiss National
Supercomputing Center (Lugano)
October 10-11
EPCC
(Edinburgh)
9
Sample Projects
© Hyperion Research
Precision Medicine
Automated Driving Systems
Fraud and anomaly detection
Affinity Marketing
Business Intelligence
Cyber Security
IoT
© Hyperion Research 10
HPDA & AI Key Takeaway:
Most Economically Important Use Cases
High Growth Areas:
HPDA-AI (May 2019)
11
• HPDA is growing faster than the overall HPC market
• The AI subset is growing faster than all HPDA
© Hyperion Research 2019
THE ROI From HPC & Success Storieswww.HyperionResearch.com/roi-with-hpc/
12© Hyperion Research
HPDA Algorithm Report: Mapping Algorithms
to Verticals & System Requirements https://hyperionresearch.com/proceed-to-download/?doctodown=hpda-algorithm-report
13© Hyperion Research
HPDA Algorithm Report: Mapping Algorithms
to Verticals & System Requirements https://hyperionresearch.com/proceed-to-download/?doctodown=hpda-algorithm-report
14© Hyperion Research
HPDA Algorithm Report: Mapping Algorithms
to Verticals & System Requirements https://hyperionresearch.com/proceed-to-download/?doctodown=hpda-algorithm-report
15© Hyperion Research
Finding HPC Centers (Map)www.HyperionResearch.com/us-hpc-centers-activity/
16© Hyperion Research B
Our Forecast On When & Where Exascale
Systems Will Be Installed
17© Hyperion Research
QC Documents are Starting to Roll Out
18© Hyperion Research
A New Cloud Application Assessment Toolhttps://hyperionresearch.com/cloud-application-assessment-tool/
19© Hyperion Research A
Price/Performance Trends of the
Largest Supercomputers
20© Hyperion Research
▪ This trendline projects
that the cost of a Linpack
PF will go from $100
million in 2009, to $110
thousand by 2028, a
roughly 1000X
improvement in a decade.
Tipping Points: How Quickly HPC
Buyers Can Change
© Hyperion Research 21
Tracking New AI Hardware: Emergence
of AI-Specific Hardware Ecosystem
22© Hyperion Research 2019
23
HPC Market Update
© Hyperion Research E
Top Trends in HPC
2018 was a very strong year with over 15% growth --
$13.7 billion (US$) in revenues!
• Supercomputers grew 23% = $5.4 billion in 2018
The top systems have started growing again after
over 4 years of softness
• The profusion of Exascale announcements are
generating a lot of buzz
Big data combined with HPC is creating new solutions
• Adding many new users/buyers to the HPC space
• AI/ML/DL & HPDA are the hot new areas
© Hyperion Research 24
The Worldwide HPC Server Market:$13.7 Billion in 2018
Departmental ($250K - $100K)
$3.9B
Divisional ($250K - $500K)
$2.5B
Supercomputers(Over $500K)
$5.4B
Workgroup(under $100K)
$2.0B
HPC
Servers
$13.7B
▪ Record revenues!
25© Hyperion Research
2018 HPC Market By Verticals
26© Hyperion Research
HPC Market By Vendor Shares
27© Hyperion Research
HPC Market By Regions (US $ Millions)
28© Hyperion Research
The Broader HPC Market Forecast
29© Hyperion Research
The Total HPC Market Including Public Cloud Spending ▪ TOTAL HPC spending grew from $22B in 2013 to $26B in 2017, and
is projected to reach $44B in 2022
30© Hyperion Research
31
HPC Cloud Forecasts
© Hyperion Research A
Major Trend
HPC in the Cloud
▪ Over 70% of HPC sites run some jobs in public clouds
• Up from 13% in 2011
▪ Over 10% of all HPC jobs are now running in clouds
• Public clouds are cost-effective for some jobs, but up to
10x more expensive for others
• Key concerns: security, data loss
▪ Private and hybrid cloud use is growing faster
▪ Large public clouds are going heterogeneous – and are
getting much better at running a broader set of HPC
workloads
• AWS with Ryft FPGAs, Google with NVIDIA GPGPUs,
Microsoft with Cray & Bright
32© Hyperion Research
Top Reasons for Using External Clouds
All Sites
1. Extra capacity (“surges”)
2. Cost-effectiveness
3. Isolate R&D projects
4. Special HW/SW features
5. Management decision
6. No on-premise data center
Government Only
1. Special HW/SW features
2. Extra capacity (“surges”)
3. No on-premise data center
4. Cost-effectiveness
5. Management decision
6. Isolate R&D projects
33© Hyperion Research
How we Developed the HPC Cloud
Forecast
▪ A public cloud, as we define it, is any large, third-
party cloud that sits outside of an organization's
firewall and is accessible in return for payment
by anyone over the public Internet
• And is owned by a different organization
• It excludes private clouds
▪ Our market forecast is from the point of view of
the end user
• The amount end users spend in public clouds
• It’s not a measure of the hardware within the clouds
© Hyperion Research 34
HPC Public Cloud Forecast
© Hyperion Research 35
What We Often Hear
36
“My boss says, why can’t you do everything in the cloud?”
“Clouds are only for embarrassingly parallel jobs.”
“It’s much more expensive to run jobs in the cloud.”
“It’s much cheaper to run jobs in the cloud.”
“I’m worried about data security and data loss in the cloud.”
© Hyperion Research
So we created a tool/scorecard to help understand
the fit of public clouds for HPC applications…
37
The Hyperion Research
Cloud Friendliness
Scorecard Tool
© Hyperion Research
About This Tool
▪ This tool is designed to help users decide between two popular environments for running HPC workloads: • On premise facilities or in public clouds.
▪ The metrics and observations are drawn from our global research and interactions with end users, CSPs and other members of the worldwide HPC community.
▪ The scorecard at the end of this slide deck allows users to record their responses concerning each application or workload they evaluate using this tool.
Hyperion Research offers this tool without charge to the HPC community, in the hope that it will be useful.
We welcome feedback on the tool: Contact Alex Norton, [email protected]
© Hyperion Research 38
How to Use This Tool
▪ To use the tool, you select a number on each
scale (between 0 to 10) for each attribute, and
then average them for an overall score.
• The tool consists of a series of “sliders" that you can
select between the two end points for each workload
you use the tool to evaluate: “Public Cloud Favored”
or “On Premise Favored.”
• The scorecard at the end shows where your overall
evaluation falls:
▪ A lower score favors the on-prem option
▪ A higher score favors the public clouds option
© Hyperion Research 39
© Hyperion Research 40
Tool Scorecard
Next Steps
✓ Provided as an online tool ➢Go To: www.HyperionResearch.com (Special Projects)
▪ Collect data on a large, diverse set of HPC, HPDA & AI applications (just started) • And score them using this tool
▪ Then create a directory of HPC applications, and identify what are the major road blocks for each area:• Those that are public cloud friendly
• Those that fit better on-prem
• Those that are in the middle
▪ Then conduct custom studies for clients
© Hyperion Research 41
© Hyperion Research 2018 PRIVATE INFORMATION FOR DESIGNATED U.S. GOVERNMENT RECIPIENTS 42
Quantum Computing
B
This Effort is an Amplification of Long-Term
Coverage of QC by Hyperion Research
43© Hyperion Research
The Promise of QC is Substantial…
QC systems have the potential to exceed the performance of conventional computers for problems of importance to humankind and businesses alike in areas such as:
• Cybersecurity
• Materials science
• Chemistry
• Pharmaceuticals
• Machine learning
• Optimization
And the list grows longer each day
© Hyperion Research 2019 44
… But So Are the Challenges Ahead
• Formidable technical issues in QC hardware and
software
• Uncertain performance gains
• Unclear time frames
• Wild West in algorithm/application progress
• Looming workforce issues
This complicates treating QC as a stable market along
side more traditional IT sectors
• Making a business case is tough
© Hyperion Research 2019 45
Errors … Errors Everywhere
© Hyperion Research 2019 46
Many Developers, Many Approaches
© Hyperion Research 2019 47
• D Wave
• IBM
• Quantum Circuits
• ionQ
• Intel
• Rigetti
• Microsoft
• Quantum Diamond
Technologies
• ATOS/Bull
• More….
▪ All are dealing with errors: NISQ is near-term
solution
Key Questions To Be Addressed
48© Hyperion Research 2019
• What are the leading development trends in QC hardware and software?
• Who are the most significant QC developers and how do they compare with counterpart efforts?
• What is the likelihood of success and time frame for leading QC development efforts?
• Which verticals can benefit most from near-term QC advances?
• What is the expected time frame/schedule for various QC hardware and software availability?
• What are some of the key technology drivers and inhibitors in QC development?
• What applications will benefit most from QC in both the near and far term?
What We Are Hearing, INPO
▪ Current business case -- FOMO
• Luckily, barrier to entry is low
• Divining ROI is not yet possible
▪ Provable algorithms may not be necessary
• ML/DL developers don’t lose sleep over this
• Best/worst case vs average case (GE)
▪ Quantum Supremacy/Advantage not necessary
• Demonstrated speed up over classical is all that matters
• The enterprise space doesn't want to have to look too far down the stack
▪ Living in the age of NISQ
• Pair with classical to get near-term results
• Concentrate on algorithms that are fault-tolerate
© Hyperion Research 2019 49
How This Plays Out Near-Term
▪ Near-term QC Ramp Up
• Many exploring applications/use cases and not just for traditional HPC but also in enterprise computing environments
• The QC/SME application developer may be the next unicorn
▪ Cloud Access Model looks to be short-term preference
▪ Crowd sourcing of algorithms/applications is current order of the day
▪ Will this succeed?
▪ We are not at the Moore’s Law stage yet
▪ National agendas are a double-edged sword
• As are national security considerations
▪ A rising tide lifts all boats
• For now
© Hyperion Research 2019 50
© Hyperion Research 2018 PRIVATE INFORMATION FOR DESIGNATED U.S. GOVERNMENT RECIPIENTS 51
The Exascale Race
Projected Exascale System Dates
U.S.
▪ Sustained ES*: 2022-2023
▪ Peak ES: 2021
▪ ES Vendors: U.S.
▪ Processors: U.S. (some ARM?)
▪ Cost: $500M-$600M per system (for early systems), plus heavy R&D investments
52
China▪ Sustained ES*: 2021-2022
▪ Peak ES: 2020
▪ Vendors: Chinese (multiple sites)
▪ Processors: Chinese (plus U.S.?)
▪ 13th 5-Year Plan
▪ Cost: $350-$500M per system, plus heavy R&D
EU▪ PEAK ES: 2023-2024
▪ Pre-ES: 2020-2022 (~$125M)
▪ Vendors: US and then European
▪ Processors: x86, ARM & RISC-V
▪ Initiatives: EuroHPC, EPI, ETP4HPC, JU
▪ Cost: Over $300M per system, plus heavy R&D investments
Japan
▪ Sustained ES*: ~2021/2022
▪ Peak ES: Likely as a AI/ML/DL system
▪ Vendors: Japanese
▪ Processors: Japanese ARM
▪ Cost: ~$1B, this includes both 1 system
and the R&D costs
▪ They will also do many smaller size
systems
* 1 exaflops on a 64-bit real application 52© Hyperion Research
Projected Exascale Investment Levels
(In Addition to System Purchases)
U.S. ▪ $1 to $2 billion a year in R&D
(around $10 billion over 7 years)
▪ Investments by both the government & vendors
▪ Plans are to purchases multiple exascale systems each year
53
China▪ Over $1billion a year in R&D (at
least $10 billion over 7 years)
▪ Investments by both governments & vendors
▪ Plans are to purchases multiple exascale systems each year
▪ Investing in 3 pre-exascale systems starting in late 2018
EU▪ About 5-6 billion euros in total (around
$1 billion a year)
▪ EU: 486M euros, Member States: 486M euros, Private sector: 422M euros
▪ Investments in multiple exascale and pre-exascale systems
▪ Large EU CPU funding
Japan▪ Planned investment of over
$1billion* (over 5 years) for both the R&D and purchase of 1 exascale system
▪ To be followed by a number of smaller systems ~$100M to $150M each
▪ Creating a new processor and a new software environment
* Note that this includes both the system and R&D53© Hyperion Research
US Exascale Plans
A21 A22 Frontier (OLCF5)
El Capitan (ATS-4) NERSC-10
NSF Frontera Follow-on
Location ANL ANL ORNL LLNL LBNL/NERSC TACC
Planned Delivery Date/ Estimated 2022 Q1 2022 2022 Q1 2022 2024 2024
Early Operation 2022, Q2 2023 2022, Q3 2023 2025 2025
Planned/Realized Performance (Pflops) ~1,000 1,300 or
higher1500-3000 4000-5000 8000-
12000 500
Linpack Performance (PFlops) 800-900 780-1040 900-2100 2000-3000 2000-3000
Linpack/Peak Performance Ratio
(%)80-90 (est.)
60-70 (est.)
60-70 (est.) 50-60 50-60 50-55
High Performance Conjugate Gradient
(PFlops/s)20.0-22.5 19.5-26.0 18-36 48-72 52-78 74
GF/Watt 40 60-100 134-200 266-480
54© Hyperion Research
Chinese Exascale Plans
Sunway 2020 Sugon Exascale NUDT 2020
Key User/Developer Sunway/NRCPC Sugon/AMD NUDT
Planned Delivery Date/ Estimated
2020, 4Q (could slip 1-1.5 years)
2020, 4Q (could slip 1-1.5 years)
2020, 4Q (could slip 1-1.5 years)
Planned/Realized Performance (Pflops) 1000 1024 1000
Linpack Performance (PFlops) 600-700 627-732 700-800
Linpack/Peak Performance Ratio (%) 60-70 60-70 (est.) 70-80
High Performance Conjugate Gradient
(Pflops/s)6-7 9.4-10.1 14-16
GF/Watt 30 34.13 20-30
Linpack GF/Watt 20-23 20.9 23.3-32.0
© Hyperion Research 55
56
EU Exascale Plans
© Hyperion Research
In Comparison: European Growth Compared To
USA Growth In Acquiring Supercomputers
▪ Europe has close to doubled their supercomputer purchases since 2005 (94.9% growth)
▪ Compared to the US growing by only 18.6% since 2005
5757© Hyperion Research
The European JU Plan
New procurements for 8 systems: http://europa.eu/rapid/press-release_IP-19-2868_en.htm
The JU along with the hosting sites, plan to acquire 8 supercomputers: • 3 precursor to exascale machines (more than 150 Petaflops) • And 5 petascale machines (at least 4 Petaflops)
The precursor to exascale systems are expected to provide 4-5 times more computing power than the top PRACE supercomputers
• Together with the petascale systems, they will double the supercomputing resources available for European-level use
• In the next few months the Joint Undertaking will sign agreements with the selected hosting sites
• The supercomputers are expected to become operational during the second half of 2020
© Hyperion Research 58
New EU Processors
© Hyperion Research 59
EPI Roadmap Targets Inclusion in Pre-
and Full-Exascale Supercomputers
© Hyperion Research 2019 60
EPI General Purpose Processor
(GPP) and Variants
© Hyperion Research 2019 61
The
HPC Innovation Award
and Winners
62© Hyperion Research A
Our Award Program
63© Hyperion Research
Examples Of Previous Winners
64© Hyperion Research
The Trophy For Winners
65
HPC Award Program Goals
▪ #1 Help to expand the use of HPC by showing real ROI
examples:
▪ Expand the “Missing Middle” – SMBs, SMSs, etc. by providing
examples of what can be done with HPC
▪ Show mainstream and leading edge HPC success stories
▪ #2 Create a large database of success stories across
many industries/verticals/disciplines
▪ To help justify investments and show non-users ideas on how
to adopt HPC in their environment
▪ Creating many examples for funding bodies and politicians to
use and better understand the value of HPC → to help grow
public interest in expanding HPC investments
▪ For OEMs to demonstrate success stories using their products
66© Hyperion Research
Users Must Show the Value of the
Accomplishment
▪ Users are required to submit the value achieved with their HPC system, using three broad categories: a) Dollar value of their HPC project
– e.g., made $$$ in new revenues
– Or saved $$$ in costs
– Or made $$$ in profits
b) Scientific or engineering accomplishment from their project
– e.g. discovered how xyz really works, develop a new drug that does xyz, etc.
c) The value to society as a whole from their project
– e.g. ended nuclear testing, made something safer, provided protection against xyz, etc.
… and the investment in HPC that was required
67© Hyperion Research
ISC19 Winners:
HPC User Innovation Awards
68© Hyperion Research
Massively Parallel Evolutionary Computation
for Empowering Electoral Reform: Quantifying
Gerrymandering via Multi-objective
Optimization and Statistical Analysis
▪ Wendy K. Tam Cho, NCSA
▪ Yan Liu, NCSA
© Hyperion Research 2019 69
Gerrymander: manipulate the
boundaries of electoral districts
so as to favor one party or group
Project DisCo (Discovery of Coherent
Structures)
▪ Adam Rupe, UC-Davis, NERSC
▪ Karthik Kashinath, NERSC
▪ Nalini Kumar, Intel
© Hyperion Research 2019 70
▪ James Crutchfield, UC-Davis
▪ Ryan James, UC-Davis
▪ Prabhat, NERSC
Can machine learning help find modes of
laser-driven fusion that have been
undiscovered by traditional methods?
▪ Brian Spears, et al., LLNL
© LLNL, 2019 71
© Hyperion Research
Hidden Earthquake Research
Caltech team: Zachary Ross, Daniel Trugman, Egill Hauksson, Peter Shearer
© Hyperion Research 72
We Invite Everyone
To Apply For The Next Round Of
Innovation Awards!
73© Hyperion Research
74
In Summary:
Some Predictions
For the Next Year
© Hyperion Research
The Exascale Race
Will Drive New Technologies
▪ The global ES race is boosting funding for the Supercomputers market segment and creating widespread interest in HPC
▪ Exascale systems are being designed for HPC, AI, HPDA, etc.
• This will drive new processor types, new memories, new system designs, new software, etc.
▪ And (in some cases) that HPC is too strategic to depend on foreign sources
• This has led to indigenous technology initiatives
© Hyperion Research 2019 75
▪ Hyperion Research studies confirm that this
convergence is occurring and speeding up
▪ Competitive forces are driving companies to aim more-
complex questions at their data structures and push
business operations closer to real time
▪ Important HPC capabilities: scalable parallel processing,
ultrafast data movement and ultra-large, capable
memory systems
▪ The HPC and commercial sectors are also converging
around a shared need to extremely data-intensive AI-ML-
DL workloads, both on the simulation and analytics side
© Hyperion Research 2019 76
The HPC-Enterprise Market Convergence Will Drive HPC Products into the Broader Enterprise Market
▪ Choices of processing elements (CPUs, accelerators)
will increase
• x86 will remain the dominant HPC CPU, but indigenous
CPUs will gain ground
▪ In AI, startups and large companies are developing
processors designed for analytics workloads
▪ NVIDIA is the dominant accelerator, but three-quarters
of respondents in our recent survey expect NVIDIA
GPUs to face "serious competition" in the next 4-5 years
▪ Processors exploiting ARM IP are planned for Europe
(EPI), Japan (Post-K computer) and China
© Hyperion Research 2019 77
Many New Processors/Accelerators
Are on The Way
▪ Data-intensive HPC is driving new storage requirements
▪ Iterative methods will expand the size of data volumes needing to be stored
▪ Future architectures will allow computing and storage to happen more pervasively on the HPC infrastructure
▪ Metadata management will deal with data stored in multiple geographic locations and environments
▪ Physically distributed, globally shared memory will become more important
▪ More intelligence will need to be built into storage software
© Hyperion Research 2019 78
Storage Systems Will Increasingly
Become More Critical
▪ Running HPC workloads in CSP environments will grow
in step-function like leaps
• As clouds get better at running HPC workloads
▪ HPC on premise and cloud environments will more
closely resemble each other: the containerization of HPC
▪ 5G will be important for reducing latency in AI use cases
that rely on coupled local-cloud environments, such as
automated driving systems, precision medicine, fraud
detection, cyber security and Smart Cities/IoT
▪ Hybrid cloud environments will grow to be a highly viable
option for HPC users
© Hyperion Research 2019 79
Cloud Computing For HPC Workloads
Will Grow Fast, Via Tipping Points
▪ The AI market is at an early stage but already highly useful (e.g., visual and voice recognition)• Once better understood, there are many high value use cases
that will drive adoption
▪ Advances in inferencing will reduce the amount of training needed for today's AI tasks, but the need for training will grow to support more challenging tasks
▪ The trust (transparency) issue that strongly affects AI today will be overcome in time
▪ Learning models (ML, DL) have garnered most of the AI attention, but graph analytics will also play a crucial role with its unique ability to handle temporal and spatial relationships
© Hyperion Research 2019 80
Artificial Intelligence Will Grow
Faster Than Everything Else
▪ HPC is a high growth market• Growing recognition of HPC’s strategic value
• HPDA, including ML/DL, cognitive and AI
• HPC in the Cloud will lift the sector writ large
▪ HPDA, AI, ML & DL is growing very quickly • The HPDA, AI, ML & DL markets will expand
opportunities for vendors
▪ Vendor share positions shifted greatly in
2015, 2016, 2017 & again in 2019 and may
continue to shift • e.g., HPE acquisition of Cray
▪ Software continues to lag hardware and
systems designs/sizes
Conclusions
81©Hyperion Research
Important Dates For Your Calendar
▪ 2019 HPC USER FORUM MEETINGS:
• Late August, Hohhot and Beijing, China
• September 9 to 11, Chicago Illinois, Argonne National Laboratory
• October 7 to 8, Lugano, Switzerland at CSCS
• October 10 to 11, Edinburgh, Scotland at EPCC
▪ And Join us for the Hyperion SC19 Briefing
© Hyperion Research 82
Visit Our Website: www.HyperionResearch.com
Twitter: HPC_Hyperion@HPC_Hyperion
Hyperion Research Holdings, LLC:▪ Owns 100% of the IDC HPC assets▪ Tracking the HPC market since 1986 ▪ Headquarters:
365 Summit Ave., St. Paul, MN 55102
84© Hyperion Research