AI in Health Care using IBM Systems/OpenPOWER systems

© 2020 IBM Corporation
IBM Cognitive Systems
Ander Ochoa – ander.ochoa.gilo@ibm.com
Cognitive Systems Architect for SPGI
OpenPOWER Foundation member
https://es.linkedin.com/in/anderotxoa
AI in Healthcare

Agenda
§Healthcare and IBM
§IBM built supercomputers building blocks
§IBM Power solutions unique advantages
§Benefits summary of IBM based solutions
§IBM software solutions leverage AI
§DEMO

Healthcare & IBM

IBM Supercomputer Summit Attacks Coronavirus…
4
https://newsroom.ibm.com/IBM-helps-bring-supercomputers-into-the-global-fight-against-COVID-19
We invite you, as well, to check
how IBM technology helps
researchers generate potential
new drug candidates for COVID-19
https://covid19-
mol.mybluemix.net/
https://www.ibm.com/blogs/nordic-msp/ibm-supercomputer-summit-attacks-coronavirus/
https://newsroom.ibm.com/US-Dept-of-
Energy-Brings-the-Worlds-Most-Powerful-
Supercomputer-the-IBM-POWER9-based-
Summit-Into-the-Fight-Against-COVID-19

IBM SPGI & Universitat Pompeu Fabra - Barcelona
5
https://www.upf.edu/es/web/focus/noticies/-/asset_publisher/qOocsyZZDGHL/content/id/236038347/maximized#.YK-PUeuxVaJ
https://www.bsc.es/bsc%E2%80%99s-tool-visualizes-the-
relationship-between-mobility-and-covid-19-spread
Other uses in healthcare
• Public sector Healthcare in northern Spain for
rare deseases.
• Public sector customer genomics.
• Spanish Hospital Severo Ochoa for protein
folding molecular analysis
(https://www.ibm.com/downloads/cas/6PR8Z7L8 )

IBM and R&D
6
https://www.research.ibm.com/patents/
Artificial Intelligence Hybrid Cloud Quantum Computing

Three of the 10 most POWERFUL HPC systems: Summit, Sierra & Marconi
The United States Department of Energy together with Oak Ridge National Laboratory and Lawrence
Livermore National Laboratory have contracted IBM and Nvidia to build two supercomputers, the Summit and
the Sierra, that are based on POWER9 processors coupled with Nvidia's Volta GPUs. These systems went
online in 2018.
http://www.teratec.eu/actu/calcul/Nvidia_Coral_White_Paper_Final_3_1.pdf
IBM Summit #2 !!
IBM Sierra #3 !!
IBM Marconi100 #9 !! Online in
2020
#1 and #2
from 2018-
2020

CINECA
MARCONI100
8
© Copyright IBM Corporation 2020
https://www.ibm.com/case-studies/cineca-systems-power-hpc-exascale
IBM POWER9 + NVIDIA
A new accelerated HPC system will be installed at
Cineca in February 2020. This system, acquired by
Cineca within the PPI4HPC European initiative,
opens the way to the pre-exascale Leonardo
supercomputer expected to be installed in 2021.
MARCONI100, is based on the IBM Power9
architecture with NVIDIA Volta GPUs. Specifically,
each node will host 2x16 cores IBM POWER9
AC922 at 3.1 GHz with 256 GB/node of RAM
memory and 4 x NVIDIA Volta V100 GPUs per
node, Nvlink 2.0, 16GB. The number of nodes will
be 980, totallying 31360 cores. Internal Network:
Mellanox Infiniband EDR DragonFly+
Model: IBM Power AC922
(Whiterspoon)
Racks: 55 total (49 compute)
Nodes: 980
Cores: 31360
Processors: 2 x 16 cores IBM
POWER9 AC922 at 3.1 GHz
Accelerators: 4 x NVIDIA Volta
V100 GPUs, Nvlink 2.0, 16GB
Cores: 32 cores/node
RAM: 256 GB/node
Peak Performance: about 32
Pflop/s
Internal Network: Mellanox
Infiniband EDR DragonFly+
Disk Space: 8PB Gpfs storage

Part of Mare Nostrum 4
• 3 Racks
• 54 Power9 Systems
• 54 AC922 servers
• 4x Nvidia V100
• 512 GB RAM
• 6.4 TB NVMe storage
• 1.48 Pflops !
BSC Mare Nostrum 4

IBM's Pangea III is the world's most powerful commercial supercomputer
10
Total’s Supercomputer Ranked First in
Industry Worldwide
The new IBM POWER9-based supercomputer (25 PFLOPS & 50
Pbytes) will help Total more accurately locate new resources and
better assess the potential of new opportunities.
According to Total Pangea III requires 1.5 Megawatts, compared
to 4.5 MW for its predecessor system. Combined with the
increased performance of Pangea III, Total has reported that they
have observed that the new system uses less than 10% the
energy consumption per petaflop as its predecessor.
• Higher Resolution Seismic Imaging in exploration and
development phase
• Reliable Development and Production Models
• Asset Valuation and Selectivity

Satori PowerAI Cluster at MIT
satori.mit.edu is the name of a new scalable AI oriented
hardware resource for research computing at MIT. It is made
possible by a donation through IBM Global Universities
Program. Provided as a gift from IBM it will help further the
aims of the new MIT Stephen A. Schwarzman College of
Computing and other campus initiatives that are combining
supercomputing power and AI algorithmic innovation.
Announced by Dr John Kelly at the Schwarzman School Kick
Off Feb 28. : ‘A slice of Summit’
Satori Specs:
• 2560 POWER9 Cores with NVLINK 2.0
• 256 NVIDIA V100 GPUs with NVLINK 2.0
• 64TB RAM DDR4 (8 channels)
• PCIe4.0 InfiniBand EDR Interconnect
• 2.9 PB of shared storage
• 1.46 PFLop @ 94 kW (#4 #Green500)
Satori lives at Mass Green HPC Center (MGHPCC)
Operational since MGHPCC Sept 30, 2019
MIT - IBM Watson AI Lab / version 1.0
https://researchcomputing.mit.edu/satori/home/

NICE: UK’s New ‘Northern Intensive
Computing Environment’
12
https://n8cir.org.uk/news/northern-intensive-computing-environment/
IBM POWER9 + NVIDIA NICE will comprise 32 IBM Power 9
dual-CPU nodes, each with 4
NVIDIA V100 GPUs and high
performance interconnect. This is
the same architecture as the US
government’s SUMMIT and
SIERRA supercomputers which
occupied the top two places in a
recently published list of the world’s
fastest supercomputers. An
additional 6 nodes will use T4 and
FPGA technology targeted towards
AI inference, to improve error
estimation and robust prediction.
This architecture supports memory
coherence between the GPU and
CPU and a hierarchy of
interconnects to allow effective
distributed GPU use, extending
problem sizes that can be tackled
beyond that of other GPU-
accelerated architectures.
Feb. 17, 2020— The N8 Centre of Excellence in Computationally Intensive
Research, N8 CIR, has been awarded £3.1m from the Engineering and Physical
Sciences Resources Council to establish a new Tier 2 computing facility in the
north of England. This investment will be matched by £5.3m from the eight
universities in the N8 Research Partnership which will fund operational costs and
dedicated research software engineering support.
Through the N8 CIR working in partnership with IBM, OCF and NVIDIA, the new
centre will contribute to EPSRCs’ and N8's goals through:
● Engaging with new and existing research communities
● Delivering a new computing architecture with new technologies
● Preparing the UK for new Tier-1 and Tier-0 architectures
● Accelerating the ‘time to science’ for a range of ‘hard’ problems
● Integration into the existing Tier-2 and National e-infrastructure
● Helping develop new computational skills for the RSEs and the researchers
● Developing new links with high-profile international supercomputing centres
● Enabling multi-disciplinary science through its unique architecture

IBM built
Supercomputer
Building Blocks

14

15
The Best Server for Enterprise AI
IBM® Power System™ Accelerated Compute Server (AC922)

16
AC922 System buses and components diagram
32 -140+GB/s
64GB/s
Fast link to exchange memory contents
between servers
Fast link to share
memory contents
with the GPUs

Large Memory
Support (LMS)
Distributed Deep
Learning (DDL)
Other Frameworks
(Snap ML)
IBM POWER Solutions
Unique advantages

18
Large Memory Support (LMS)
Objective: Overcome GPU Memory Limitations in DL Training. Increase the Batch Size
and/or increase the resolution of the features.
LMS enables processing of high definition images, large models, and higher batch
sizes that doesn’t fit in GPU memory today (Maximum GPU memory available in
Nvidia P100 and V100 GPUs is 16/32GB).
Available for
- Caffe
- TensorFlow
- Chainer
https://www.sysml.cc/doc/127.pdf
GPU RAM
System RAM
NVLink
v2.0
2 TB
16/32 GB
Accelerated
by
NVLink
Dataset

https://www.linkedin.com/pulse/deep-learning-high-resolution-images-large-models-sumit-gupta/
19

Performance results with TensorFlow Large Model Support v2
20
ResNet50
3D U-Net
TensorFlow Large Model Support in PowerAI 1.6 allows
training models with much higher resolution data.
Combining the large model support with the IBM Power
Systems AC922 server allows the training of these high
resolution models with low data rate overhead.
https://developer.ibm.com/linuxonpower/2019/05/17/performance-results-with-tensorflow-large-model-support-v2/
TF LMS v2
DeepLabV3+

More information
21
ü TensorFlow Large Model Support Code / Pull Request:
ü https://github.com/tensorflow/tensorflow/pull/19845/
ü TensorFlow Large Model Support Research Paper:
ü https://arxiv.org/pdf/1807.02037.pdf
ü TensorFlow Large Model Support Case Study:
ü https://developer.ibm.com/linuxonpower/2018/07/27/tensorflow-large-model-support-case-study-3d-image-segmentation/
ü IBM AC922 with NVLink 2.0 connections between CPU and GPU:
ü https://www.ibm.com/us-en/marketplace/power-systems-ac922

22
Distributed Deep Learning
Objective: Overcome the server boundaries of some DL frameworks.
How: Scaling. Using “ddlrun” applied to Topology aware distributed frameworks.
Our software does deep learning training fully synchronously with very low communication overhead.
The overall goal of ddlrun is to improve the user experience DDL users.
To this end the primary features of ddlrun are:
• Error Checking/Configuration Verification
• Automatic Rankfile generation
• Automatic mpirun option handling
Available for:
• Tensorflow
• IBM Caffe
• Torch
https://www.sysml.cc/doc/127.pdf
Good for:
• Speed
• Accuracy

Distributed Deep Learning (DDL) for Training phase
Using the Power of 100s of Servers
August 8, 2017
16 Days Down to 7 Hours: Near Ideal Scaling to 256 GPUs and Beyond
1 System 64 Systems
16 Days
7 Hours
ResNet-101, ImageNet-22K, Caffe with PowerAI DDL, Running on Minsky (S822Lc) Power System
58x Faster
https://www.ibm.com/blogs/research/2017/08/distributed-deep-learning/

SnapML v2
24
Our latest version of Snap ML adds high performance implementations of
Decision Trees and Random Forests. Our implementation of these algorithms
takes advantage of multiple CPU cores and threads (but so far do not take
advantage of GPUs). Click here for documentation on Snap ML.
https://medium.com/@sumitg_16893/snap-ml-2x-faster-machine-learning-than-scikit-learn-c3529a1a6172
Snap ML, a python-based machine learning framework that is designed to be a
high-performance machine learning software framework. Snap ML is bundled as
part of the WML Community Edition or WML CE (aka PowerAI) software
distribution that is available for free on Power systems.
The first release of Snap ML enabled GPU-acceleration of generalized linear
models (GLMs) and also enabled scaling these models to multiple GPUs and
multiple servers. GLMs are popular machine learning algorithms, which
include logistic regression, linear regression, ridge and lasso regression,
and support vector machines (SVMs).
http://www.cirrascale.com/ibmpower_snapml.php
Snap ML (PowerAI 1.6.0) supports:
•Generalized Linear models
• Logistic Regression
• Linear Regression
• Ridge Regression
• Lasso Regression
• Support Vector Machines
•Tree-based models
• Decision Trees
• Random Forest
•2H19 release will add support for
•Gradient Boosting Machines (GBMs)

Tera-scale Machine Learning Benchmark record with POWER9
25
Snap ML main features:
• Distributed training: It is data-parallel framework, enabling train on massive
datasets that exceed the memory capacity of a single machine.
• GPU acceleration: We take advantage of recent developments in
heterogeneous learning in order to enable GPU acceleration even if only a
small fraction of the data can indeed be stored in the accelerator memory.
• Sparse data structures: We employ some new optimizations for the
algorithms when applied to sparse data structures.
https://www.ibm.com/blogs/research/2018/03/machine-learning-benchmark/
Using an online advertising
dataset released by Criteo Labs
with over 4 billion training
examples, we train a logistic
regression classifier in 91.5
seconds. This training time is 46x
faster than the best result that
has been previously reported,
which used TensorFlow on
Google Cloud Platform to train
the same model in 70 minutes.

Benefits
summary of IBM
solutions

27 https://www.nextplatform.com/2018/08/28/ibm-power-chips-blur-the-lines-to-memory-and-accelerators/

28

Server virtualization security is critical for DB workloads since many
are run in virtual environments
• The PowerVM hypervisor has only had one
hypotetical reported security vulnerability and
provides the bullet-proof security that customers
demand for mission-critical workloads
• The VIOS, which is part of the overall
virtualization has had 0 reported security
vulnerabilities
• Dare to compare – search any security
tracking DB and compare Power against x86
1
reported hypothetical security
breeches on the PowerVM
hypervisor (in Dec 2020)
Power VM security
March 2021

NVIDIA T4
TPU2 = Tensor
Processing Unit
Accelerates neural
network computations
Training & Inference, from the core to the edge
INFERENCE
TRAINING
IBM
TrueNorth
Neuromorphic CMOS
integrated circuit
Nvidia Xavier ARM ML core
Nervana Neural Network
Processor (Intel)
Special purpose Processors
N
v
i
d
i
a
T
e
n
s
o
r
c
o
r
e
s
General purpose Processors
XILINX Alvep U200 FPGA
Havana GOYA
GRAPHCORE IPU

31
Open Architecture
Based in OPEN standards of the OpenPOWER
Foundation our servers gather the ingenuity of 340+
of the greatest IT companies all around the world.
- No lock-in to a single vendor, no licensing
- Under Linux Foundation umbrella
Consolidation
The Virtualization capabilities of our servers with
PowerVM exceed the granularity of resources and
flexibility for workload virtualization.
- We can divide a core in 20 virtual cores
- Plus we can divide one core in 8 SMT cores
- Plus VMs & containers, SDNs, SDS, CoD…
Artificial Intelligence
The NVLink connectivity with Nvidia GPUs provides
almost 10x data bandwidth vs x86.
- Use of NVIDIA V100 for train & inference.
- Future NVIDIA T4 for inference
- IBM Research developments (LMS,DDL,SnapML)
- …
RAS
Reliability against any internal & external issue.
- CPUs logical Hot-swap
- Memory logical Hot-swap
- Extra memory lanes
- Redundant everything (PSUs, disks…)
- …
Latency
The use of PVIe v4 plus OpenCAPI plus clever
design allows a Mellanox Infiniband PCIe card to be
shared between 2 sockets to minimize latency.
- RDMA
- Ideal for network traffic switching.
Input / Output Capabilities
- Use of PCIe V4 (32GB/s) since December 2018
vs PCIe V3 ( 16GB/s).
- 48 PCIe V4 channels.
- OpenCAPI support for coprocessors
- OpenCAPI 12x25GB/s devices (today).
- …
OpenPOWER benefits in a nutshell

IBM SW
solutions
leverage AI

33
Red Hat OpenShift
Cloud Pak for
Applications
Cloud Pak for
Data
Red Hat OpenShift
Cloud Pak for
Integration
Red Hat OpenShift Red Hat OpenShift
Cloud Pak for
Automation
Red Hat OpenShift
Cloud Pak for
Multicloud
Management
Node.js
React
Kitura.js
Jenkins Nginx
Spring
Istio
Swift
Open Liberty
WAS WAS ND
App Connect
Enterprise
API Connect DataPower
Event Streams MQ Aspera
Watson Voice
Gateway *
RabbitMQ Integration
Explorer
Business
Automation
Workflow
Operational
Decision
Management
Business
Automation
Insight
Business
Automation
Content Analyzer
File
Connect
Manager
+ Add-on *
Multicloud
Manager
Cloud
Application
Management
Cloud
Event
Manager
Cloud
Automation
Manager
Cloudant DB2 Data
Virtualization
Cognos Streams MongoDB
MariaDB* Redis* etcd
+ Add-on * (including Watson AI, …)
WAS Liberty
Transformation
Advisor
JBoss
Kabanero
Enterprise
RHOAR - OpenShift
Application Runtimes
*UrbanCode Deploy
*Developer Team Orch
*Dev Team Governance
+ Add-on * + Add-on *
Robotic
Process
Automation *
* PostgreSQL
* NetApp Persistent Storage
* Wand Taxonomies and Ontologies
* Knowis for Banking
* Lightbend Reactive Microservices
* Prolifics Prospecting Accelerator
IBM
Mobile
Foundations
Runs on choice of IBM Power
Systems Infrastructure-as-
a-Service (IaaS)
Bare-metal
Available
today
IBM Cloud Paks

IBM Cognitive Systems IBM Cloud Pak For Data: Base Platform vs Cartridges
ü Db2 Warehouse
ü Data Virtualization
ü Db2 Eventstore
ü IBM Streams
ü Watson Knowledge Catalog (including IGC)
ü IBM Regulatory Accelerator ( included in WKC)
ü Information Analyzer
ü Watson Studio (includes Data Refinery)
ü Watson Machine learning (includes AutoAI )
ü Watson OpenScale
ü Cognos Dashboards Embedded
ü Analytics Engine for Apache Spark
ü Open source governance
ü IBM Performance Server (only on CPD System)
ü Db2
ü DataStage
ü Cognos Analytics
ü Information Server
ü Watson Studio Premium
ü Watson Assistant
ü Watson Discovery
ü Watson API Kit
ü Watson Financial Crimes Insights
Base Platform Extensions
34
Services supported on Power Systems in YELLOW

Data Sources
35

Application
Container
Framework
SAP System
Operational
Data
ML Model
Repository
Train Data
Deployment
Training
Inference
Business Process
ML
SAP Hana as Data Source - SAP Hana and Artificial
Intelligence

“Cloud Pak for Data” is an integrated Data & AI Platform
Common Services
37

Cloud Pak for Data Key use cases
Operationalize Data
Science & AI
Manage your Data
Anywhere
Shift to Cloud Native
Data workloads Smarter Governance
e.g.; accelerate GDPR Compliance
Build, deploy, manage &
govern models & data at
scale to improve business
outcomes
e.g.; a. Customer Churn
b. Cross Sell / Up Sell
c. Predictive Maintenance
Shift to Cloud Native
a. Provision & scale Data & AI
services
b. Build once, deploy anywhere
– multi cloud support
c. Built in automation, scaling
& collaboration to increase
productivity
1. Manage all your
enterprise data regardless
of where it lives
(Data Virtualization)
2. Gain control & leverage
your real-time data for
analytics
(Fast data ingest & Streaming
analytics)
Governance to enable self
service analytics
Auto-discover meta data, manage
governance rules & policies,
enforce privacy etc. to mitigate
risk & ensure compliance
38

IBM Cloud pak
for Data DEMO

Login to the Cloud Pak for Data running in a POWER9 server
40

Dataset used for the AI experiment
41

Select winner model and create Notebook for accountability.
42

Put the model in production sharing access code and online quick test.
43

44
by
Your Innovation Platform!

Notice and disclaimers
ü Copyright © 2017 by International Business Machines Corporation (IBM). No part of this document may be reproduced or transmitted in any form without written permission from IBM.
ü U.S. Government Users Restricted Rights — use, duplication or disclosure restricted by GSA ADP Schedule Contract with IBM.
ü Information in these presentations (including information relating to products that have not yet been announced by IBM) has been reviewed for accuracy as of the date of initial publication and could
include unintentional technical or typographical errors. IBM shall have no responsibility to update this information. This document is distributed “as is” without any warranty, either express or
implied. In no event shall IBM be liable for any damage arising from the use of this information, including but not limited to, loss of data, business interruption, loss of profit or loss of
opportunity. IBM products and services are warranted according to the terms and conditions of the agreements under which they are provided.
ü IBM products are manufactured from new parts or new and used parts. In some cases, a product may not be new and may have been previously installed. Regardless, our warranty terms apply.”
ü Any statements regarding IBM's future direction, intent or product plans are subject to change or withdrawal without notice.
ü Performance data contained herein was generally obtained in a controlled, isolated environments. Customer examples are presented as illustrations of how those customers have used
IBM products and the results they may have achieved. Actual performance, cost, savings or other results in other operating environments may vary.
ü References in this document to IBM products, programs, or services does not imply that IBM intends to make such products, programs or services available in all countries in which IBM operates or
does business.
ü Workshops, sessions and associated materials may have been prepared by independent session speakers, and do not necessarily reflect the views of IBM. All materials and discussions are provided
for informational purposes only, and are neither intended to, nor shall constitute legal or other guidance or advice to any individual participant or their specific situation.
ü It is the customer’s responsibility to insure its own compliance with legal requirements and to obtain advice of competent legal counsel as to the identification and interpretation of any relevant laws
and regulatory requirements that may affect the customer’s business and any actions the customer may need to take to comply with such laws. IBM does not provide legal advice or represent or
warrant that its services or products will ensure that the customer is in compliance with any law.

Notice and disclaimers continued
Information concerning non-IBM products was obtained from the
suppliers of those products, their published announcements or
other publicly available sources. IBM has not tested those
products in connection with this publication and cannot confirm
the accuracy of performance, compatibility or any other claims
related to non-IBM products. Questions on the capabilities of
non-IBM products should be addressed to the suppliers of those
products. IBM does not warrant the quality of any third-party
products, or the ability of any such third-party products to
interoperate with IBM’s products. IBM expressly disclaims all
warranties, expressed or implied, including but not limited
to, the implied warranties of merchantability and fitness for
a particular, purpose.
The provision of the information contained herein is not intended
to, and does not, grant any right or license under any IBM
patents, copyrights, trademarks or other intellectual
property right.
IBM, the IBM logo, ibm.com, AIX, BigInsights, Bluemix, CICS,
Easy Tier, FlashCopy, FlashSystem, GDPS, GPFS,
Guardium, HyperSwap, IBM Cloud Managed Services, IBM
Elastic Storage, IBM FlashCore, IBM FlashSystem, IBM
MobileFirst, IBM Power Systems, IBM PureSystems, IBM
Spectrum, IBM Spectrum Accelerate, IBM Spectrum Archive,
IBM Spectrum Control, IBM Spectrum Protect, IBM Spectrum
Scale, IBM Spectrum Storage, IBM Spectrum Virtualize, IBM
Watson, IBM z Systems, IBM z13, IMS, InfoSphere, Linear
Tape File System, OMEGAMON, OpenPower, Parallel
Sysplex, Power, POWER, POWER4, POWER7, POWER8,
Power Series, Power Systems, Power Systems Software,
PowerHA, PowerLinux, PowerVM, PureApplica- tion, RACF,
Real-time Compression, Redbooks, RMF, SPSS, Storwize,
Symphony, SystemMirror, System Storage, Tivoli,
WebSphere, XIV, z Systems, z/OS, z/VM, z/VSE, zEnterprise
and zSecure are trademarks of International Business
Machines Corporation, registered in many jurisdictions
worldwide. Other product and service names might
be trademarks of IBM or other companies. A current list of
IBM trademarks is available on the Web at "Copyright and
trademark information" at:
www.ibm.com/legal/copytrade.shtml.
Linux is a registered trademark of Linus Torvalds in the United
States, other countries, or both. Java and all Java-based
trademarks and logos are trademarks or registered
trademarks of Oracle and/or its affiliates.

AI in Health Care using IBM Systems/OpenPOWER systems

Empfohlen

Empfohlen

Weitere ähnliche Inhalte

Was ist angesagt?

Was ist angesagt? (20)

Ähnlich wie AI in Health Care using IBM Systems/OpenPOWER systems

Ähnlich wie AI in Health Care using IBM Systems/OpenPOWER systems (20)

Mehr von Ganesan Narayanasamy

Mehr von Ganesan Narayanasamy (20)

Kürzlich hochgeladen

Kürzlich hochgeladen (20)

AI in Health Care using IBM Systems/OpenPOWER systems