Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem...

17
0 © 2017 FUJITSU ISC2017 – Exhibitor Forum Connect. Challenge. Inspire. HPC Universe Expansion ISC2017 – Exhibitor Forum

Transcript of Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem...

Page 1: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

0 © 2017 FUJITSUISC2017 – Exhibitor Forum

Connect. Challenge. Inspire.

HPC UniverseExpansion

ISC2017 – Exhibitor Forum

Page 2: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

1 © 2017 FUJITSUISC2017 – Exhibitor Forum

1975…

SOLAR 16-65:• 750ns for 16bits ADD instruction• 64KBytes main memory• 2.5MBytes disk storage

Page 3: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

2 © 2017 FUJITSUISC2017 – Exhibitor Forum

1979 !

ICL PERQ

« smalltalk » environment

Multi-windows manager

Mouse

Text editor

shell

Page 4: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

3 © 2017 FUJITSUISC2017 – Exhibitor Forum

2017 ?

SAMSUNG Galaxy S6

0.788 Gflops (Linpack) multi-threaded

Page 5: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

4 © 2017 FUJITSUISC2017 – Exhibitor Forum

The scope of HPC is much larger than most think

Weather, Climate

Government Labs

Product Design & Behaviour

National Security

Academic Research

Oil/Gas Exploration and Extraction

New Hardware+Software for Deep Learning• HPC for Training: Exaflops per session, Computationally intensive,

GPUs/Phis/FPGAs

• Tactical HPC for Inference: data centers, self driving cars, smart phones, IoT, robots etc.

HPC in the Cloud• Virtualized and/or bare metal, Public/private/hybrid, On-prem/cloud

ecosystems, Making HPC more like clouds

New big data applications running in non-traditional HPC environments

• Finance or cyber security sectors: supercomputing-based threat analytics service on subscription basis

• Business analytics growing (forced) into HPC

Compute/Data-intensity creates challenges solved only by HPC architecture and software

Visible HPC Hidden HPC

Page 6: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

5 © 2017 FUJITSUISC2017 – Exhibitor Forum

How does HPC work? … Parallelism

Early supercomputers first exploited vectorisation to multiply performance

Including Fujitsu’s own VPP systems

Followed by further parallelism from multiple CPUs

Increasingly, greater parallelism is available within a range of processors and systems, accelerating a broader of applications beyond the primarily scientific

Black-Scholes, Monte Carlo in Financial Services

Convolutional Neural Networks in Deep Learning

Parallel file systems for Data Analytics

And HPC scales to solve the largest problems

Efficiently handle the throughput to match data volume

Maximise performance levels even as model/analytical complexity increases

Page 7: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

6 © 2017 FUJITSUISC2017 – Exhibitor Forum

Parallelism gains in Intel® Xeon Phi™ Processors

Latest Intel many-core processor available in PRIMERGY CX600

Page 8: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

7 © 2017 FUJITSUISC2017 – Exhibitor Forum

2003

MapReduce:

Simplified Data Processing on Large Clusters

Jeff Dean, Sanjay Ghemawat

Google, Inc.

Page 9: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

8 © 2017 FUJITSUISC2017 – Exhibitor Forum

High Performance Data Analysis (HPDA)

Big Data needs HPC capabilitiesPlatform to process Big Data with HPC technologies

HPDA solution Customer benefits

1

2

3

Fastest processing/transformationof large volume data

Real-time analysisto extract invisible insight from the data

Accelerated deep-learning technologyby GPU computation

Implementation & data management service

Analysis support service in collaboration with

data scientists

Data gathering

Data storeModeling/Analysis

Page 10: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

9 © 2017 FUJITSUISC2017 – Exhibitor Forum

MODELING & SIMULATION

Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud

New commercial users SMEs Digitalisation across

organisations

ADVANCED ANALYTICS

Existing HPC users Intelligence community, FSI Data-driven science/

engineering (e.g., biology) Knowledge discovery ML/DL, cognitive, AI

New commercial users Fraud/anomaly detection Business intelligence Affinity marketing Personalized medicine

Convergence of Analytics and HPC

Drivers: Competition, Complexity, Time Fraud and anomaly detectionIdentifying harmful or potentially harmful patterns and causes using graph analysis, semantic analysis, or other high performance analytics techniques.

MarketingPromote products or services using complex algorithms to discern potential customers' demographics, buying preferences and habits.

Business intelligenceUses HPDA to identify opportunities to advance the market position and competitiveness of businesses, by better understanding themselves, their competitors, and the evolving dynamics of the markets they participate in.

Other Commercial HPDAAn example of such a high-potential workload is the use of HPDA to manage large IT infrastructures, ranging from on premise data centers to public clouds and Internet-of-Things (IoT) infrastructures.

Source: IDC,2016

Page 11: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

10 © 2017 FUJITSUISC2017 – Exhibitor Forum

2014, the « caffe » break

@article{jia2014caffe, Author = {Jia, Yangqing and

Shelhamer, Evan and Donahue, Jeff and Karayev,

Sergey and Long, Jonathan and Girshick, Ross and

Guadarrama, Sergio and Darrell, Trevor}, Journal =

{arXiv preprint arXiv:1408.5093}, Title = {Caffe:

Convolutional Architecture for Fast Feature

Embedding}, Year = {2014} }

Deep Learning “democratization”• Caffe supports many different types of deep learning

architectures geared towards image classification and image segmentation.

• It supports CNN, RCNN, LSTM and fully connected neural network designs.

• Caffe supports GPU based accleration using CuDNN of Nvidia.

Page 12: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

11 © 2017 FUJITSUISC2017 – Exhibitor Forum

Deep Learning is now reaching viable accuracy

Courtesy of Nervana

Continuing Challenges

ARTIFICIAL INTELLIGENCEA program that can sense, reasons, act and adapt

MACHINE LEARNINGAlgorithms whose performance improve when exposed to more data over time

DEEP LEARNINGMulti-layered neural networks

learn from vast amounts of data

Large compute requirements for training

Performance that scales with data

Calculation of increasingly complex models

Page 13: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

12 © 2017 FUJITSUISC2017 – Exhibitor Forum

And HPC is fundamental to Deep Learning

Convolutional Neural Networks (CNNs) represent a significant class of Deep Learning (DL) algorithms today, and the de facto tool for visual understanding

Majority of the computations in CNNs can be formulated as Matrix-Matrix multiplications**

Optimisation approach is identical to simulation/iterative solvers in conventional HPC applications

Convolution and AllReduce are other main algorithmic kernels –also accelerated using HPC principles

All achieve high performance on many-core accelerators (GPU, Phi), with gains in specific kernels offered by dedicated units

Intel® Xeon Phi™ Processor

Knights Mill

Intel® Xeon Processor

Skylake

Lake Crest

Intel® Xeon® Processor + FPGA

Image Identity

** For more: https://svail.github.io/DeepBench/

Intel® Lake Crest Deep neural network processor

Accelerating Deep Learning Applications Convolutional Neural Network

Page 14: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

13 © 2017 FUJITSUISC2017 – Exhibitor Forum

Fujitsu High Performance AI Ecosystem

Page 15: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

14

Invest more than 50 Millions on Digital Transformation in France

A joint research on Deep-learning technology with Inria

Center of Excellence focus on AI located in Polytechnique

CoE R&D

Collaboration with France’s leading technology companies

Ecosystem

Page 16: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

15 © 2017 FUJITSUISC2017 – Exhibitor Forum

AI

Cloud

Parallel Processing

GPU

HPC

Deep Learning

Machine Learning

Image Analytics / Search

Encoding / Decoding

Imagingtech.

Video Surveillance

View of HPC in 2017 and Beyond

Page 17: Connect. Challenge. Inspire. - Fujitsu · MODELING & SIMULATION Existing HPC users Larger problem sizes Higher resolution Iterative methods EP jobs to the cloud New commercial users

16 © 2017 FUJITSUISC2017 – Exhibitor Forum