IBM HPC Transformation with AI

HPC Transformation with AI
COGNITIVE SYSTEMS
Ing. Florin Manaila
Senior Architect and Inventor
Cognitive Systems (Distributed Deep Learning and HPC)
IBM Systems Hardware Europe
Member of the IBM Academy of Technology (AoT)
March 24, 2020

Technical R&D today disruption
2Cognitive Systems Europe / March 24 / © 2020 IBM Corporation

Knowledge Discovery Pipeline

Infrastructure
Demands for AI
Equipped for volumes of data
Flexible storage for a
range of data demands
Versatile, power-efficient
data center accelerators
Advanced I/O for
minimal latency
Scalability and distributed
data center capability
Inference
Powerful data center
accelerators with coherence
Advanced I/O for high
bandwidth and low latency
Proven scalability
Training
Equipped for volumes of data
*** IBM and Business Partner Internal Use Only ***
Cognitive Systems Europe / March 24 / © 2020 IBM Corporation

Next-Generation
Infrastructure
Stack

6
AI Workflow

Distributed Deep Learning
Common options
7
SINGLE ACCELERATOR DATA PARALLEL MODEL PARALLEL DATA AND MODEL PARALLEL
1x Accelerator 4x Accelerators 4x Accelerators
4x n Accelerators
Longer Training Time Shorter Training Time
System1System2Systemn
System
Data
Data
DataDataDataData
DataDataData

Node 0
Data-Parallel Framework
Distributed Learning
Partition 0
GPU 0
GPU 1
GPU 2
GPU 3
Partition (0,0)
Partition (0,1)
Partition (0,2)
Partition (0,3)
Node 1
Partition 1
GPU 0
GPU 1
GPU 2
GPU 3
Partition (1,0)
Partition (1,1)
Partition (1,2)
Partition (1,3)
8
Large Dataset

Rearchitecting
the hardware
for AI

10
Experimentation Scaling Production
Architecture for large IBM HPC Cluster
Hardware overview for bare-metal / K8s

Architecture for large IBM HPC/AI Cluster
Hardware overview for bare-metal / K8s

Why Capable
and High BW Interconnects?
§ Each part of an application runs on the best
compute location
o But there are performance and
programmability challenges
o Desire a highly-capable interconnect between
PEs
§ Low-latency communication and high data
bandwidth
§ Fine-grained + bulk data transfers
§ Consistent, unified view of memory
§ Hardware cache coherence & atomic operations
PE Type A
(e.g. CPU)
PE Type B
(e.g. GPU)
Large, low-latency
Memory
Small, High-
bandwidth
Memory
Heterogeneous systems are attractive for efficient performance

13
Enterprise AI Hardware Portfolio Expansion
IBM Power AC922
TRAIN
Powering the Fastest Supercomputer
DATA
IBM Power IC922
INFERENCE
IBM Power IC922
Deploy AI into ProductionStorage Dense Server
§ Enterprise ready cloud deployment
with RH OpenShift and Power
Systems reliability
§ Superior I/O for data movement:
PCIe Gen 4
§ Superior price/performance
§ Best training platform with 4x
faster model iteration
§ ~6x data throughput with NVLink
to GPUs
§ Synergistic HW/SW offerings for
ease of use and leadership
performance
§ NVIDIA V100 SMX2 GPUs
§ Superior density and through-put to
inference accelerators
§ Open design for accelerator
flexibility
§ Deploy inference at scale with SW
capabilities leveraging superior IO
§ NVIDIA T4 GPUs
§ Upcoming: FPGAs and ASICs

Inference server details
14
Form Factor
§ 19” Rack 2U Server
POWER9 Processor
§ 2 dd2.3x P9 Nimbus chips (LaGrange pkg)
§ TDP : 225W
§ 12 (160W), 16, 20 cores (SMT <= 4)
Memory
§ Direct Attach Memory
§ 32 DDR4 ISDIMM Slots @2400 MHz (double
drop)
§ 16 DDR4 ISDIMMs @2667 MHz (single drop)
§ 16, 32, 64 GB RDIMMs
§ 2 TB Max memory
§ 340 GB/s peak memory BW (with 16x DIMMs)
10 Integrated I/O Slots – Standard PCIe Riser
§ 2 PCIe G3 x16 FHFL Slots
(Supports double-wide accelerator)
§ 2 PCIe G4 x16 LP Slots
§ 2 PCIe G3 x8 FHFL Slots (physically x16)
§ 2 PCIe G3 x8 FHHL Slots
§ 2 PCIe G3 x16 LP Slots

15
Internal Storage
§ Integrated Storage Controller = None
§ 24x 2.5” SAS/SATA
Native I/O
§ 2x USB 3.0 in rear
§ 2x 1G baseT (one shared mgmt) + 1x
1G dedicated IPMI
§ Serial port, VGA port
§ TPM2.0 via Nuvoton
NPCT650ABAWX included (for
Secure OS and trusted boot)
MTM (machine type – model)
• 9183-22X
Accelerators
Nvidia T4 Accelerator 16GB PCIe3 x16 LP
More to come
Networking
Mellanox MCX555A-ECAT 1-PORT EDR 100Gb IB CONNECTX-
5 GEN3 PCIe x16 CAPI CAPABLE
Mellanox MCX556A-ECAT 2-PORT EDR 100Gb IB CONNECTX-
5 GEN4 PCIe x16 CAPI CAPABLE
Mellanox MCX516A-CDAT 2-PORT 100Gb ROCE EN
CONNECTX-5 GEN4 PCIe x16
Mellanox MCX4121A-XCAT 2-PORT 10Gb NIC&ROCE SR/Cu
PCIe 3.0
Mellanox MCX4121A-ACAT 2-PORT 25/10Gb NIC&ROCE
SR/Cu PCIe 3.0
Marvell BCM957810A1008ICDM 2-PORT E'NET (2X10 10Gb),
PCIe Gen 2 X8
Marvell BCM957800A1006ICDM QUAD E'NET (2X1 + 2X10
10Gb), PCIe Gen 2 X8
Marvell BCM957800A1006ICDM QUAD E'NET (2X1 + 2X10
10Gb), PCIe Gen 2 X8
Broadcom BCM5719-4P 1Gb E'NET(UTP) 4-PORT ADPTR,
PCIE-x4
Fiber Channel
Broadcom LPe16002B-M6 2-PORT FIBER CHANNEL(16Gb/s),
PCIE3-8X
Broadcom LPe32002-M2 2-PORT FIBER CHANNEL(32Gb/s),
PCIE3-8X

16
8x SAS/SATA 8x SAS/SATA 8x SAS/SATA
9183-22X
OS Support – LE
§ RHEL 7.6-alt
RAS
§ Concurrent Maintenance disks
§ Redundant Hot plug Power
§ Redundant Hot plug fans
§ Customer Install and Repair
§ Simplified Op Panel
§ In-rack system service
BMC Service Processor
§ Aspeed AST2500
§ OpenBMC
Certifications
§ FCC Class A
§ ASHRAE A2 Environment
(10-35C)
§ Acoustics Datacenter 1A
HDD Drives
HDD; 600GB; 2.5"; 10k; SAS; 12Gb/s; 4Kn/512e; SED
HDD; 1200GB; 2.5"; 10k; SAS; 12Gb/s; 4Kn/512e; SED
HDD; 2400GB; 2.5"; 10k; SAS; 12Gb/s; 4Kn/512e; Non-SED
SSD Drives
SSD; 240GB; 2.5"; SATA; 6Gb/s; 1.4 DWPD; NonSED
CONTROLLERS
Broadcom (LSI) MegaRAID 9361-8i SAS3 Controller w/ 8
internal ports (2GB Cache) PCIe 3.0 x8 LP with cables
Broadcom 9300-8i PCIe gen3 x8 LP with cables
Broadcom 9305-16i PCIe gen3 x8 LP with cables

Feature List:
§ REST Management
§ IPMI
§ SSH based SOL
§ Power and Cooling
Management
§ Event Logs
§ Zeroconf discoverable
§ Sensors
Features In
Progress:
§ Full IPMI 2.0
Compliance with DCMI
§ Verified Boot
§ HTML5 Java Script Web
User Interface
§ BMC RAS
IBM is the
OpenBMC
Community Leader
§ Facebook
§ Google
§ IBM
§ Intel
§ Microsoft
§ OCP
17
OpenBMC is a free open
source management
software Linux distribution
§ Inventory
§ LED Management
§ Host Watchdog
§ Simulation
§ Code Update Support for
multiple BMC/BIOS
images
§ POWER On Chip
Controller (OCC) SupportCognitive Systems Europe / March 24 / © 2020 IBM Corporation

Next-Generation
Software Stack

What’s in the training of deep neural networks?
Neural network model
Billions of parameters
Gigabytes
Computation
Iterative gradient based search
Millions of iterations
Mainly matrix operations
Data
Millions of images, sentences
Terabytes
Workload characteristics: Both compute and data intensive!

20
AI Infrastructure Stack
ON-CLOUD and ON-PREM
Transform & Prep
Data (ETL)
Micro-Services / Applications
Governance AI
(Fairness, Explainable AI,
Model Health, Accuracy)
APIs
(external and in-house)
Machine & Deep Learning
Libraries & Frameworks
Distributed Computing
Data Lake & Data Stores
Segment Specific:
Finance, Retail, Healthcare,
Automotive
Speech, Vision,
NLP, Sentiment
TensorFlow, Caffe,
Pytorch
SparkML, Snap.ML
Spark, MPI
Hadoop HDFS,
NoSQL DBs,
Parallel File
System
Accelerated
Infrastructure

Watson ML Community Edition (WMLCE)
21

Watson ML Community Edition (WMLCE)
22

23
IBMWatsonMachineLearning
CommunityEdition
DockerContainers

24
IBMWatsonMachineLearning
CommunityEdition
UniversalBaseImages(UBI)

Thank you
26
Florin Manaila
—
florin.manaila@de.ibm.com
ibm.com

IBM HPC Transformation with AI

More Related Content

What's hot (20)

Similar to IBM HPC Transformation with AI (20)

More from Ganesan Narayanasamy (20)

Recently uploaded (20)

IBM HPC Transformation with AI