Share
H2O & MapD Meetup at NVIDIA: GPU Accelerated AI
The GPU Open Analytics Initiative (GOAI) was formed to create common data frameworks enabling developers and statistical researchers to accelerate data science on GPUs.
GOAI will foster the development of a data science ecosystem on GPUs by allowing resident applications to interchange data seamlessly and efficiently.
Special thanks to the NVIDIA for hosting this meetup.
Speaker Bios:
Vinod Iyengar is the Director of Business Development and Partnerships at H2O.ai. Vinod comes with over 7 years of marketing and data science experience in multiple startups. He was the founding employee for his previous startup, Activehours, where he helped build the product and bootstrap the user acquisition with growth hacking. He has seen the user base for his companies grow from scratch to millions of customers. He’s built models to score leads, reduce churn, increase conversion, prevent fraud and many more use cases.
When he is not busy hacking, Vinod loves painting and reading. He is a huge foodie and will eat anything that doesn’t crawl, swim or move.
Bill Maimone is the VP of Engineering at MapD. Bill is responsible for leading the engineering team in architecting and delivering the the company’s suite of database, visualization and GIS solutions. Prior to MapD, Bill was the VP of Engineering at Anaplan and held similar roles with Salesforce and Actian. Bill began his career at Oracle, where he spent two decades, finishing as a Vice President with responsibility for over five hundred members of the R&D team across four continents. He holds a M.S. and a B.S. in Computer Science and a Bachelor of Science in Journalism from the Massachusetts Institute of Technology.
OmniSci
Transcript
43 slides
-
H2O & MapD: GPU Accelerated AI Meetup at NVIDIA Bill
Maimone, VP of Engineering at MapD Vinod Iyengar, Director of Business Development and Partnerships at H2O.ai August 29, 2017 -
Confidential & Proprietary 3 Who is MapD? MapD was incubated
in the MIT CSAIL database group under the advisory of Michael Stonebraker and Sam Madden (Vertica). MapD has captured the imagination of some of the most sophisticated investors in Silicon Valley and beyond. “It's completely amazing to access databases so large completely in-memory and to interact with it, create graphs out of it, query it with AI, visualize it, all in real time. Completely revolutionary stuff.” Jensen Huang, CEO -
Confidential & Proprietary 4 GPUs offer a way forward GPU
Processing Power 50% per year Data Growth 40% per year CPU Processing Power 20% per year -
Confidential & Proprietary 5 GPU Processing CPU Processing (Traditional) 5,120
Cores Cores and memory define the difference 24 Cores -
Confidential & Proprietary 6 MapD: software optimized for the fastest
hardware + 100x Faster Queries Speed of Thought Visualization MapD Core MapD Immerse A fast, relational, column store database powered by GPUs A visual analytics engine that leverages the speed + rendering capabilities of MapD Core -
Confidential & Proprietary 7 Analytics ANALYTICS 3.0 - ACCELERATED -
PREDICTIVE - DYNAMIC ANALYTICS 2.0 ANALYTICS 1.0 1990s 2017 2000s The evolution of analytics GPUs will be as transformative to Analytics as Broadband was to the Internet -
Confidential & Proprietary 8 Where does MapD fit in? Complementing
your entire data ecosystem -
MapD Core 9 The world's fastest in-memory GPU database powers
the world's most immersive data exploration experience -
Confidential & Proprietary 10 Keeping Data Close to Compute MapD
Core: Performance starts with memory management SSD or NVRAM STORAGE (L3) 250GB to 20TB 1-2 GB/sec CPU RAM (L2) 32GB to 3TB 70-120 GB/sec GPU RAM (L1) 24GB to 256GB 1000-6000 GB/sec Hot Data Speedup = 1500x to 5000x Over Cold Data Warm Data Speedup = 35x to 120x Over Cold Data Cold Data COMPUTE LAYER STORAGE LAYER Data Lake/Data Warehouse/System Of Record -
Confidential & Proprietary 11 10111010101001010110101101010101 00110101101101010101010101011101 Query Compilation with LLVM
Traditional DBs can be highly inefficient • each operator in SQL treated as a separate function • incurs tremendous overhead and prevents vectorization MapD compiles queries w/LLVM to create one custom function • Queries run at speeds approaching hand-written functions • LLVM enables generic targeting of different architectures (GPUs, X86, ARM, etc). • Code can be generated to run query on CPU and GPU simultaneously -
Confidential & Proprietary 12 Mark Litwintschik benchmarked MapD vs. major
CPU systems on a billion-row taxi data set and found it to be 6x to 12,500x faster than the fastest CPU databases. MapD GPU-Powered Queries are Fastest MapD Comparative Query Acceleration* System Query 1 Query 2 Query 3 Query 4 BrytlytDB & 2-node p2.16xlarge cluster 36x 47x 25x 12x ClickHouse, Intel Core i5 4670K 49x 58x 32x 25x Redshift, 6-node ds2.8xlarge cluster 74x 24x 14x 6x BigQuery 95x 38x 6x 6x Presto, 50-node n1-standard-4 cluster 190x 75x 61x 41x Amazon Athena 305x 117x 37x 13x Elasticsearch (heavily tuned) 386x 343x n/a n/a Spark 2.1, 11 x m3.xlarge cluster w/ HDFS 485x 153x 119x 169x Presto, 10-node n1-standard-4 cluster 524x 189x 127x 61x Vertica, Intel Core i5 4670K 685x 607x 203x 132x Elasticsearch (lightly tuned) 1,642x 1,194x n/a n/a Presto, 5-node m3.xlarge cluster w/ HDFS 1,667x 735x 388x 159x Presto, 50-node m3.xlarge cluster w/ S3 2,048x 849x 164x 86x PostgreSQL 9.5 & cstore_fdw 7,238x 3,302x 1,424x 722x Spark 1.6, 5-node m3.xlarge cluster w/ S3 12,571x 5,906x 3,758x 1,884x *All speed comparisons are to the “MapD & 1 Nvidia Pascal DGX-1” benchmark Source: http://tech.marksblogg.com/benchmarks.html -
Confidential & Proprietary 14 Basic charts are frontend rendered using
D3 and other related toolkits Scatterplots, pointmaps + polygons are backend rendered using the Iris Rendering Engine on GPUs Geo-Viz is composited over a frontend rendered basemap MapD Immerse: our hybrid approach -
Confidential & Proprietary 16 GOAI: End-to-end analytics on the GPU
GPU Open Analytics Initiative – Fusing Machine Learning and GPU Analytics -
Confidential & Proprietary 17 GOAI: End-to-end analytics on the GPU
GPU Open Analytics Initiative – Fusing Machine Learning and GPU Analytics MapD + pymapd + H2O (GPU) CPU ML Pipeline GPU ML Pipeline GPU ML Pipeline on GOAI Pointer in VRAM Python (CPU) Spark (CPU) CPU ML (CPU) CPU MapD (GPU) Python (GPU) ML (GPU) -
MapD Proprietary. Get involved with GOAI ➢ GOAI website: gpuopenanalytics.com/
➢ GitHub: github.com/gpuopenanalytics ➢ Discussion Group: groups.google.com/forum/#!forum/gpuopenanalytics -
“Confidential and property of H2O.ai. All rights reserved” Company Overview
Founded 2011 Venture-backed, debuted in 2012 Products • H2O-3 • Sparkling Water • H2O4GPU • Driverless AI Mission Operationalize Data Science, and provide a platform for users to build beautiful data products Team 70 employees • Distributed Systems Engineers doing Machine Learning • Kaggle Winning Data Scientists • World-class visualization designers Headquarters Mountain View, CA -
CONFIDENTIAL 3 A.C. Nielsen A1 Telekom Austria AAPT Abovenet Communications
Academic Administrative and Research Network Academic Computer Centre Cyfronet H Accelerated Data Works Accenture Accenture Services Ace Ina Holdings Ace International Ace Telecom Acton Acxiom Oration Adamo Telecom Iberia Administracion Nacional De Telecomunicaciones Admiral Objekt Waesche & Arbeitskleidung Adobe Systems Adobe Systems India Adsl Maroc Telecom Advanced Cable Communications Advanced Computer Solutions Affecto Afrihost-Dynamic Ainet Telekommunikations-Netzwerk Betriebs Air Bank A.S. Air Liquide Sa Airess Cesko Akamai Technologies Aktia Saastopankki Oy Aktiv-I Szolgaltato Al-Shahad Information Technology Albert Einstein College of Medicine of Yeshiva University Albert-Ludwigs-Universitaet Freiburg Alexander & Alexander Information Technology Algar Telecom Aliyun Computing Allbusiness.Com Allianz Maned Operations & Services Se Bell Canada Beltelecom Beyond The Network America Bezeq International- Bh - Tec Bharti Airtel Bibliotheque Nationale De France Big Fish Games Bigleaf Networks Biglobe Bilink Bimeh Dormitory Sharif University of Technology Bio-Rad Laboratories Biocontrol Bisiness Network Jv Bite Communications Biznet Biznet Metronet Blekinge Institute of Technology Blue Line Infotech Blueconnect Boingo Wireless Bol.Com Bv Boots UK Retail Boranet Borlange Energi Boston Scientific Oration Bouygues Telecom Division Mobile Bouygues Telecom Sa Brain Telecommunication Bright House Networks Brighthouse Networks Cfl Division Brighthouse Networks Indianapolis Bristish Petroleum British Sky Broadcasting Broadriver Communication Broadstripe Brutele Sc Bryant University Case Western Reserve University Catalina Marketing Catalina Marketing Oration Cect-Chinacomm Communications Cedars-Sinai Health Systems Celgene Oration Center For Governmental Research Centerbeam Central Telegraph Public Joint-Stock Centre De Calcul El-Khawarizmi - Cck Centre For Advanced Computing Centro De Tecnologia Da Informa O Renato Archer Ceom Israel Cerfnet Cerner Oration Certara USA Ceu Cgi Group Champaign Telephone Charles University Charlesbrauer Charter Communications Chegg Chengdu West Dimension Digital Technology Cheonanjeonhwakukjang Chico Board of Trade China Digital Kingdom Technology China Education and Research Network Chinatelecom Group Beijing Co Chongqing Times Newper Office Chs - Bna Lan Chunghwa Telecom Data Communication Business Group Cik Telecom Cisco Cisco Systems Cisco Systems Ironport Division Citadel Investment Group L.L.C. Citrix Systems City University Delft University of Technology Network Delhi Technical University(Dce) Deloitte Deloitte Services Deloitte Touche Tohmatsu Services Deloitte and Touch Regional Consulting Services Delphon Industries Delta Dental Plan of Michigan Delta Leasedline Network Deluxe Oration Den Networks Dena Deutsche Telekom Deutsches Reisebuero Develon Dhirubhai Ambani Institute of Information Dialog Axiata Digi Tavkozlesi Es Szolgaltato Digia Digital Entertainment Digital Hosting Technology Digital Network Associates - Franchisee Digital Ocean Digital Realm Digital River Digital-Entertainment-Industry-Development-Co--Zhongshan Zho Digitalocean Cloud Direct Supply Discoveries In Sight Dishnet Wireless Distributel Communications Disy Informationssysteme Diverge Consulting Dna Oy Doclernet Dongbeicaijingdaxue-Dl-Ln Doorway As Dotomi Drivetime Enbridge Pipelines Ency For Science Technology and Research End-User Numericable Energy Sciences Network Enom Orporated Ensync Business Solutions Pty Entanet International Enterprise Teaming Enzu Eotvos Lorand University of Sciences Epam Systems Epm Telecomunicaciones E.S.P. Epsilon Data Manement Dba Equant Equinox Consulting Erasmus Mc Erasmus University Rotterdam Ericsson Business Communications Ericsson Network Systems Escout Consulting Espn Estate Valuations and Pricing Systems Etapa Ep Etex Communications Etheric Networks Ethio Telecom Ethz Swiss Federal Institute of Technology Zurich Etisalat Lanka (Private) European Bioinformatics Institute Evergy Excell Media Exe2 Newton Abbot Exetel Act Dsl Exponential-E FPL Fibernet Facebook Fachhochschule Dortmund Fachhochschule Nordwestschweiz Faculty of Sciences University of Lisbon Companies Using H2O.ai 2015 2016 N ow 2017 G oal 14,000 10,281 6,427 3,810 H2O.ai Users 2015 2016 N ow 2017 G oal 140,000 97,620 54,163 38,257 10,000+ Companies use H2O — World Wide Community Adoption -
CONFIDENTIAL H2O.ai is a Visionary in the Gartner Magic
Quadrant for Data Science Platforms 4 “Overall customer satisfaction is very high.” “H2O is especially suited to IoT edge and device scenarios.” “H2O had the highest reference customer analytics support score of all the vendors.” “H2O.ai has significant adoption by large enterprises such as Macy’s, Comcast, and Capital One.” “H2O.ai is best known for developing open source, cluster-distributed ML algorithms at a time (2011) when big data demanded them, but no one else had them.” H2O.ai is a Strong Performer in the Forrester Predictive Analytics & Machine Learning H2O ranked #1 for Advanced Prototyping in Gartner Critical Capabilities for Data Science Publish: January 2017 “H2O received an outstanding score in machine learning, largely driven by its popular best-in-class machine-learning implementations” H2O.ai Strongly Positioned in Key Analyst Reports “It also scored very highly for flexibility, extensibility and openness, as well as for delivery.” -
CONFIDENTIAL AI in Financial Services 5 Wholesale / Commercial Banking
• Know Your Customers (KYC) • Anti-Money Laundering (AML) Card/Payments Business • Transaction Frauds • Real-time Targeting • Credit Risk Scoring • In-Context Promotion Retail Banking • Deposit Fraud • Customer Churn Prediction • Auto-Loan IT Infrastructure • Security Cyberlake • DoS Detection and Protection • Master Data Management -
CONFIDENTIAL AI in Healthcare 6 Flu Season Prediction Personalized Drug
Matching Medical Claim Fraud Detection Emergency Room and Hospital Management Drug Discovery Remote Patient Monitoring Early Cancer Detection / Oncology Medical Imaging and Diagnostics Product Recommendation -
“Confidential and property of H2O.ai. All rights reserved” Cloud Neutral
Big Data Ecosystem H 2 O.ai Makes A Difference as an AI Platform Open Source Flexible Interface Interpretability GPU Enablement Rapid Model Deployment Smart and Fast Algorithms H 2 O Flow • Highly portable models deployed in Java (POJO) and Model Object Optimized (MOJO) • Automated and streamlined scoring service deployment with Rest API • Reason code for predictions • Variable importance for models • Visual Intelligence -
9 H2O GPU Solutions H2O4GPU • 100% open source •
Enterprise Support Algorithms • GLM • Gradient-boosted Tree • K-means • Random Forest Coming Soon • Linear SVM • PCA • Nearest Neighbors • Multi-GPU / multi-node Driverless AI • Enterprise License Solution functionality • Automatic Feature Engineering • Automated ML • Machine Learning Interpretability (MLI) • Automatic Visualization Coming Soon • Time Series / Text input • Model Stacking -
“ConfidenZal and property of H2O.ai. All rights reserved” Model Model
Building Features Target Modeling Table Data Quality & Transformation Data Integration + Typical Enterprise Machine Learning Workflow -
“ConfidenZal and property of H2O.ai. All rights reserved” Model Model
Building Features Target Modeling Table Data Quality & Transformation Data Integration + Driverless AI Typical Enterprise Machine Learning Workflow -
“ConfidenZal and property of H2O.ai. All rights reserved” Data VISUAL
MODEL INTERPRETATION AUTO DL Distributed Multi-GPU Model Repository Deploy MODEL FITNESS data.table Kaggle Grandmaster in a Box Automatic Feature Engineering Pipeline Export English Explanations Reason Codes K-Lime, LOCO, Partial Dependence Driverless AI Recipes H2O4GPU -
“ConfidenZal and property of H2O.ai. All rights reserved” 3 Pillars
Speed Accuracy Interpretability -
“ConfidenZal and property of H2O.ai. All rights reserved” Speed •
GPU acceleration to achieve up to 40x speedups vs CPU • Multi GPU - XGBoost, GLM, K-Means and more • Achieve best performance in shortest time H2O4GPU -
“ConfidenZal and property of H2O.ai. All rights reserved” 0 0.1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Allstate BNP Paribas Amazon Homesite Otto Group Relative error: Lower is Better Kaggle Grandmaster Best AutoDL GBM Baseline • Automatic feature engineering to increase accuracy - AlphaGo for AI • Automatic Kaggle Grandmaster recipes in a box for solving wide variety of use-cases • Automatic machine learning to find and tune the right ensemble of models Accuracy Preliminary results - untuned, single model Driverless AI -
“ConfidenZal and property of H2O.ai. All rights reserved” Interpretability •
Interpretability for debugging, not just for regulators • Get reason codes and model interpretability in plain english • K-Lime, LOCO, partial dependence and more -
“Confidential and property of H2O.ai. All rights reserved” Automatic Visualization
• Statistically sound and relevant visualizations generated automatically to highlight important characteristics in the data. • Automatic Scagnostics and other visualizations to generate the most relevant visualizations for each dataset • Can handle billions of rows using the aggregator function -
CONFIDENTIAL GPU OPEN ANALYTICS INITIATIVE github.com/gpuopenanalytics GPU Data Frame (GDF)
Ingest/ Parse Exploratory Analysis Feature Engineering ML/DL Algorithms Grid Search Scoring Model Export -
FEATURE GTC Q2 2017 Q3 2017 Q4 2017 GLM (Single
GPU) Python API for training & scoring Python .whl installation file GBM (Single GPU) Inference on GPU (GLM) GLM (Multi GPU - Multiple Models in Parallel) GBM (Multi GPU - Multiple Models in Parallel) Inference on GPU (GBM) k-Means Clustering, Nearest Neighbors - Single GPU PCA (Single GPU) Quantiles (Single GPU) GLM (Multi GPU - Single Model) Sort (Single GPU) GBM (Multi GPU - Single Model) Random Forest (Multi GPU - Single Model) Kalman Filters Sort, Quantiles (Multi GPU) SVM (Single GPU) Connectors to GPU Open AI Data Frames H2O4GPU Roadmap -
“Confidential and property of H2O.ai. All rights reserved” Links &
Resources Driverless AI Webinar Interpretability Webinar Josh Patterson Blog Andy Steinbach Blog