Artificalintelligence at the edge - bernsteinresearch.com 2017/02_ARM.pdfUpfront licence fee –flexible licensing models ... Keypoint Features (SIFT/SURF/FAST/Harris) X X X X X. Convolutional

© 2017 Arm Limited

Ian Thornton, Head of Investor Relations16 October 2017

Artifical intelligence at the edge

© 2017 Arm Limited 2

Arm introduction

Global leader in technology licensing

• R&D outsourcing for semiconductor companies

Innovative business model

• Upfront licence fee – flexible licensing models

• Ongoing royalties on partner sales

• Technology reused across multiple applications

Long-term, secular growth markets >1,480 licencesGrowing by >100 every year

>460 potentialroyalty payers

17.7 bn Arm-based chips shipped in 2016

~15% CAGR over previous 5 years

© 2017 Arm Limited 3

Arm technology is everywhereFrom sensors to smartphones to servers

22 years

4 years

4 years

20171990 2013 2021

50 billionchips shipped

50 billionchips shipped

100 billionchips expected to ship

© 2017 Arm Limited 4

Distributing intelligence everywhere

Robotics

Home, surveillance & analyticsIoT VR/MR

Mobile DronesAutomotive

Shipping & logistics

© 2017 Arm Limited 5

Bandwidth Power and Cost

From the cloud to the edge

Privacy LatencySecurity

© 2017 Arm Limited 6

AI is different for different applications

Smart

Home

Self D

riving C

ars/A

DAS

General Purpose/

Personal

Robotics

HealthMobile

Device/C

amera

Secu

rity

Industrial

Robotics

Detection (B

inary Present/N

ot Present)

Classific

ation (F

ew/broad cla

sses)

Recogn

ition (M

any s

pecific)

Motion Detec

tion

Motion Esti

mation/Trac

king

Pose Esti

mation

Segm

entation

Depth Perceptio

n

Egomotio

n Estimati

on

Super R

esolutio

n

Noise Reducti

on

Content Generat

ion/Remova

l

X X X X X X Anomoly/Event Detection X XX X X X Behavior Detection/Prediction X X X XX X X X X X Object Understanding X X X XX X X X X X Scene Understanding X X X X X X

X X X Navigation (Scene + Motion) - - - - - - X XX X X X Facial and/or Emotion Understanding X X

X X X Augmented Reality X X X XX X X X Image Capture/Enhancement X X X X X X X

© 2017 Arm Limited 7

Detection (B

inary Present/N

ot Present)

Classific

ation (F

ew/broad cla

sses)

Recogn

ition (M

any s

pecific)

Motion Detec

tion

Motion Esti

mation/Trac

king

Pose Esti

mation

Segm

entation

Depth Perceptio

n

Egomotio

n

Noise Reducti

on

Content Generat

ion/Remova

l

Viola-Jones/Haar Cascade XImage Histogram X X X

Template Matching (OCR) X XHistogram of Oriented Gradients (HoG) X X

Eigenfaces/Eigenvectors X X XKeypoint Features (SIFT/SURF/FAST/Harris) X X X X X

Convolutional Neural Networks X X XGMM Background/Foreground Segmentation X X

Sparse Optical Flow (Lucas Kanade) X X XDense Optical Flow (Gunner Farneback) X X X X X

Stereo depth perception (Block Matching) XRegistration (Homography, Stereo Matching) X X X

Clustering (k-means, EM, mean-shift) XGraph Cuts X XWatershed X

Deconvolutional Neural Network XFilters (Gaussian Blur, Median) X

Image Averaging XInpainting X

Seam Carving X

Families of AI algorithms

© 2017 Arm Limited 8

The spectrum of CV and ML algorithmsLots of data, simple computation per data item, data parallelism

Smaller amounts of highly structured data, complex computation per item

Convolution Color space conversion

Affine transforms

Feature detectors (variable)Block matching Optical flow

Support vector machines

K-means clustering

Feature matching

Neural networks

Parallel workloads:ISPs, GPUs

Serial workloads: CPUs

In-between workloads: DSPs

fixed-function IP, FPGA

© 2017 Arm Limited 9

Heterogeneity is necessary

There is no one size fits all solution

Need multiple types of processors to handle this

Fixed function works when requirements known in advance, algorithm is well understood, and high performance needed

Programmable cores are essential

DSP DSP DSP

DSP DSP DSP

DSP DSP DSP

CPU big CPU big FPGA

CPU Little

CPU Little DSP

GPU GPU Acc

© 2017 Arm Limited 10

Arm’s approach to AI

Machine Learning applications

Domain-specific libraries + ML frameworks: TensorFlow, Caffe, MXNet, etc.

Tools and libraries from Arm

CPUs + GPUs + ISPs + CV engines DSP FPGA ACC

© 2017 Arm Limited 11

ARM DynamIQ – Multicore redefined

New single cluster design Advanced compute capabilities

Redesigned memory sub-systemGreater flexibility

© 2017 Arm Limited 12

Accelerating AI adoption everywhere

DynamIQ boosting AI/ML performance both on CPU and in system

Improved access to acceleration

Up to 10x quicker response for accelerators

Dedicated processor instructions and optimized libraries for AI

More than 50x AI performance boost on the CPU in the next 3-5 years

1313

Thank You!Danke!Merci!谢谢!ありがとう!Gracias!Kiitos!

© 2017 Arm Limited