driving performance beyond moore’s...

© 2018 Arm Limited

Ian SmytheOctober 2018

Driving Performance

Beyond Moore’s Law

2 © 2018 Arm Limited © 2018 Arm Limited 2

Innovation continues to drive growth and performance demands on our compute devices

5G transformationMulti-day battery life SecuritySmall to large screen

Untethered. Connected. Immersive.

3 © 2018 Arm Limited

Last year we announced new DynamIQ processors

>20%More mobile performance vs Cortex-A73

SameSustained

performance as Cortex-A73

+40%Infrastructure performance vs Cortex-A72

Performance leadership in mobile

Best possible power profile

Improved performance in infrastructure

Up to

2xmore

performance

Up to

15%better power

efficiency

Up to

10xmore

configurable

For advanced use cases

Higher sustainerperformance

Edge to cloud scalability


Arm Cortex-A portfolio

2008 – 2013 2017201620152014

Year of IP release, volume devices in the subsequent year

Arm big.LITTLE compatible

Cortex-A5xseries

Cortex-A3xseries

Cortex-A9

Well-established, mid-range processor

Cortex-A5/A7

Smallest and lowest power

Armv7-A

Cortex-A15/A17

Infrastructure performance;

mobile efficiency

Cortex-A57Proven

infrastructure performance

Cortex-A72For all

applications

Cortex-A35Smallest, lowest power Armv8-A

Cortex-A53Balanced

performance and efficiency

Cortex-A73For mobile and

consumer

Cortex-A32Smallest, lowest

power 32-bitArmv8-A

Cortex-A55Highest

efficiency mid-range processor

64/32-bit

Cortex-A75Ground-breaking performance for

all markets

64/32-bit

32-bit64/32-bit

64/32-bit64/32-bit64/32-bit

64/32-bit

32-bit

32-bit

32-bit

2018

Cortex-A7xseries

Armv8-AArmv7-A


Arm Cortex-A portfolio

2008 – 2013 2017201620152014

Year of IP release, volume devices in the subsequent year

Arm big.LITTLE compatible

Cortex-A5xseries

Cortex-A3xseries

Cortex-A9

Well-established, mid-range processor

Cortex-A5/A7

Smallest and lowest power

Armv7-A

Cortex-A15/A17

Infrastructure performance;

mobile efficiency

Cortex-A57Proven

infrastructure performance

Cortex-A72For all

applications

Cortex-A35Smallest, lowest power Armv8-A

Cortex-A53Balanced

performance and efficiency

Cortex-A73For mobile and

consumer

Cortex-A32Smallest, lowest

power 32-bitArmv8-A

Cortex-A55Highest

efficiency mid-range processor

64/32-bit

Cortex-A75Ground-breaking performance for

all markets

64/32-bit

32-bit64/32-bit

64/32-bit64/32-bit64/32-bit

64/32-bit

32-bit

32-bit

32-bit

2018

Cortex-A76Laptop-class performance

with smartphone efficiency

64/32-bit

Cortex-A7xseries

Armv8-AArmv7-A


Designing for user experience from small screens to large

Energy savings and power efficiency

Efficient scheduling

• Power efficient cores for low demand tasks and background services

• Fast switching between cores

Power optimization

• Finer-grained speed control

• Autonomous memory power management

• Fast power on/sleep/off management


Designing for user experience from small screens to large

System responsiveness

High single thread performance for fantastic user facing response

• App. launch and closing

• Web browsing

• Productivity applications

Scalability with area efficient octa-core solution

• High CPU availability for performance and responsiveness


Let’s take a closer look


Arm Cortex-A76 CPULaptop-class performance, smartphone experience

• Built from the ground up with new microarchitecture capabilities

• Built on innovative DynamIQ technology

• Battery life that can outlast your work day

Longer Battery Life

Better energy efficiency

Performance without compromise

IncreasingProductivity

Increased Machine Learning performance

Intelligent Computing


Cortex-A76: Performance efficiency - focus on the user

Cortex-A76 CPU is focused on performance and also performance efficiency

Performance efficiency - extract significantly more performance than any other microarchitecture at similar complexity

Requires intense focus on every aspect of the microarchitecture• More performance from every logic block

Focus on the end-user to enable sustained full-speed performance

• Yes, we also do well on benchmarks


Cortex-A76: Front-end

Front-end built to hide latency at high bandwidth

Multi-level branch-target caches

Hybrid indirect predictor - unparalleled prediction capability

Decode/Rename/Commit

Front-end

L1-ITLB

Bra

nch

p

red

icti

on Instruction Fetch

64K I-Cache


Cortex-A76: Decode/Rename/Commit

4-instruction/cycle, power-optimized decode

High-density decode/rename

Dispatch to out-of-order core and commit unit

Decode/Rename/Commit

Commit

DQ

De

cod

e

Re

gist

er r

en

ame

4-8

inst

ruct

ion

s/cy

cle

4 M

op

s/cy

cle

8 u

op

s/cy

cle

dis

pat

chDis

pat

ch

Execution core

L1 Data cache / MMU


Cortex-A76: Execution core

Uops dispatched to 120-entry issue queue capacity

Dual 128-bit ASIMD/FP execution pipelines

L1 Data cache / MMU

Execution core

IQ

IQ

IQ

Branch

ALU

ALU

ALU/MAC/DIV

FMUL/FADD/FDIV/ALU/IMAC

FMUL/FADD/ALU

Inte

ger

ASI

MD


Cortex-A76: Cache hierarchy and performance

Full cache hierarchy is co-optimized for latency and bandwidth

Sophisticated 4th generation prefetcher

256KB-512KB private L2 with 9-cycle LD-use

2M-4M DynamIQ L3 with 26-31 cycle LD-use

50% performance uplift from Cortex-A750

0.5

1

1.5

2

2.5

3

L1 cache L2 cache L3 cache DRAM

Memory hierarchy bandwidthCortex-A76 vs. Cortex-A75


Accelerating the performance curve in any workloads

Pushing the single-thread performance

• +25% more integer IPC than the Cortex-A75 CPU• +35% higher ASIMD/FP performance• +90% higher memory bandwidth

Boosting mobile experience

• +28% more Geekbench performance• +35% more Javascript performance

Enabling intelligence at the edge

• 3.9x more AI performance

Baseline IPC - frequency upside from here

1.58x1.79x

1.56x1.77x

2.44x

9.7x

SPECINT2K6 SPECFP2K6 Geekbenchv4

Javascript LMBenchmemcpy

GEMM lowp

Cortex-A73 Cortex-A75 Cortex-A76

IPC comparison - iso-process/-frequency


Cortex-A76 CPU delivers premium performance

Peak single-thread performance big.LITTLE performance 5W

Performance(relative scores based on AArch64 SpecInt2K6)

Cortex-A7316nm

Cortex-A7510nm

Cortex-A767nm

Configuration: Cortex-A73 – 2.45GHz, L1 64KB, L3 2MB: Cortex-A75 – 2.8 GHz, L1 64KB, L2 512KB, L3 2MB: Cortex-A76 – 3.3 GHz, L1 64KB, L2 512KB, L3 4MB

2xPerformance improvement

2.1x 1.9x


Cortex-A76

Building for the premium experience for advanced process

DynamIQ Shared Unit

CoreLink CCI-550

Cortex-A55

LPDDR4x

Memory SystemIntegrated TrustZone technology

DMC DMC

High-performance Cortex-A76 implementation3+ GHz in 7nm

Increasing Cortex-A55 CPU private L2 cache

Implementing 4MB L3 cache

Optimized memory system

Other IPs


It starts with an ecosystem

1919

Miix 630

NovaGo Always On, Always Connected PCs powered by Snapdragon

Envy x2

Pace of innovation*Requires network connection and will support up to 20 hours of battery life

Yoga

C630

And more…

835

Credit: Qualcomm Technologies, Inc.

Process Technology

Mobile leadership

2008 2009 2010 2011 2012 2015 2016 201720142013

45nm

32nm

22nm

14/16nm

14/16nm

40nm

28nm

14nm12nm

10nm 1st GenSnapdragon 835

10nm 2nd GenSnapdragon 850

14nmSnapdragon 820

20 nmSnapdragon 810

28 nmSnapdragon S4

45nmSnapdragon S1

Time

Pro

ce

ss N

od

e

2018


Performance

Performance forscenarios that matter

Balancing powerand performance

Sustained performancethat doesn’t throttle



Delivering on promises

1.4x 1.3x 1.2x 1.3x1.4x 1.4x

0

0.2

0.4

0.6

0.8

1

1.2

1.4

Speedometer Geekbenchsingle-core

Geekbenchmulti-core

WebXPRT 3 TouchXPRT16 MotionMark 1.0

Cortex-A75 based system performance(relative to Cortex-A73 system)

Improvement across all benchmarks

Over 25% minimum performance uplift

Source: Shrout Research, measured on Lenovo C630 and HP Envy x2 devices


1.3x

Web browsing battery life improvement (relative to Cortex-A73 system)

Extended battery life and thermal constraints on real-system

Running your apps longer Multi-day battery life

1.3x

Time improvement between charge(relative to Cortex-A73 system)

Source: Shrout Research, measured on Lenovo C630 and HP Envy x2 devices


The evolution of the always on, always connected pc

Image created by Arm based off Shrout Research data: download the full whitepaper at www.shroutresearch.com


The journey ahead


Client Compute CPU roadmap

Cortex-A767nm

202020192018

‘Deimos’7nm

‘Hercules’7nm and 5nm

2017

Cortex-A7510nm

Cortex-A7316nm


Perf

orm

ance

Path to compute performance leadership with efficiency

Cortex-A73Cortex-A15 Cortex-A57 Cortex-A72 Cortex-A75 Cortex-A76 Deimos Hercules

2.5xincrease

Arm Compute

Intel Core i5 U-series

Single-core performance estimates based on SPECINT2k6

Core i5-4300u22nm

Core i5-6300u14nm

Core i5-7300u14nm

28nm 20nm 16nm 16nm 10nm 7nm 7nm 5nm2013

A performance trajectory surpassing Moore’s law

Unmatched year-over-year Arm CPU performance gains


Expanding the mobile experience

• Innovation on mobile from small screens to large is changing the user experience and continues to push growth

• The new premium IP delivers Laptop-class performance

• Arm with its ecosystem is aligning itself to meet customer needs and get ready for 5G evolution for truly connected experiences

driving performance beyond moore’s...

Documents