Qualcomm Centriq™ Arm-based Servers for Edge Computing

World’s First 10nm Server Processor

Chaitali Sengupta, PhD

Sr Director, Technology

Qualcomm Datacenter Technologies, Inc.

March 27, 2018 Open Networking Summit, North America

Qualcomm Centriq is a product of Qualcomm Datacenter Technologies, Inc

2

What is “Edge”

. . .

EDGE

Centralized CloudDevices / Premises Edge Cloud

Cloud Service Providers |

Datacenters

◦ > 100 ms latency

◦ 5-10 per operator or cloud service

provider

◦ 100s-1000s of server racks per site

Edge Cloud | Cloudlets |

Edge Gateways

◦ 5-20 ms latency

◦ Few server racks per site

Smartphones | Connected Cars |

Drones | IoT | Enterprise | Homes

◦ Customer Devices:

<2 ms latency for millions of devices

◦ Customer Premises:

<5 ms latency for thousands of devices

3

Computation Needs: Location & Purpose

. . .

EDGE

Centralized CloudDevices / Premises Edge Cloud

• Bandwidth / Backhaul Traffic

• Latency

• Compute performance needed

• Power/Thermal Envelope

• Privacy & Security

• Reliability

Parameters that drive compute locality – ie. Device / Edge / Centralized Cloud

4

Edge computing: Growing set of use cases

End Markets. Applications. Requirements.

End users / Subscribers

Enterprises

IoT

Automotive

Others

Augmented reality

Autonomous action

Location and context based

user engagement

Cognitive assistance

Connected / Self-driving vehicles

Video orchestration in events/venues • Responsive user experience

Requires low and deterministic

latency end to end

• Interactive

High transaction rate between

device and service

• Privacy/security concerns

Requiring local processing

• Data and compute heavy

More cost effective to bring

compute close(er) to the end user

5

Centralized CloudDevices / Premises

. . .

Edge Cloud

Edge computing vs. Cloud computing: What is the difference?

Edge Compute Characteristics Edge Compute Infrastructural Needs

Both virtual machines and

containers / Re-use NFV

infrastructure / Micro-services /

micro-services chaining

Use cases:

Heavy dependence on computer

vision, machine learning

Multiple access: Mobile/Fixed,

public/private/enterprise

Real time access to radio

network info

Location awareness

Ultra low latency including

network and compute delay

Distributed data handling

Radio access network interfaces

Orchestration/Management &

Run-time telemetry

CPU processing offload:

FPGA, GPU, NPU, Smart NIC, ML

accelerators

6

QualcommCentriq™

2400

Qualcomm® Falkor™ CPU

5th-Generation Custom Core Design / ARMv8-Compliant

Highly Integrated Server SoCSingle Chip Platform-level Solution /

ARM SBSA Level 3 Compliant /

60 MB L3 cache /

32 Lanes PCIe Gen3

High core countUp to 48 cores / 2.6 GHz all cores peak

frequency

Qualcomm Centriq is a product of Qualcomm Datacenter Technologies, Inc

Purpose-built for Edge and Centralized Cloud

Qualcomm Centriq & Falkor are a product of Qualcomm Datacenter Technologies, Inc

• Network Function Virtualization infrastructure support via OPNFV and leading

partners to enable components such as OpenStack, DPDK, etc.

• Micro services, Containers, Virtualization support (e.g. KVM, Docker)

Software ecosystem to enable

NFV and Edge Computing

7

Why the Centriq™ 2400 Server Processor is a good fit for Cloudlets / Edge

• High thread density and high

performance per thread at lower power

– large number of VM-s and

containers running independent

processing for radio and edge

application for multiple

bearers/users/services

• Thread isolation and predictable

latency

• Quality of service features to ensure

resources are allocated fairly and no

one service hogs them

– Isolation between multiple

bearers/users/services

– Service each user in real time

13.9 12.8 13.613.7 13.8 14.1

~Parity 8% better 4% better

Qualcomm Centriq 2460

120W TDP

vs.

Intel Xeon Platinum 8180

205W TDP*

Qualcomm Centriq

2452 120W TDP

vs.

Intel Xeon Gold

6152 140W TDP*

Qualcomm Centriq 2434

110W TDP

vs.

Intel Xeon Silver 4116

85W TDP*

SP

EC

int®

_ra

te2

00

6 E

stim

ate

pe

r T

hre

ad

Intel Xeon

Qualcomm

Centriq

Performance per thread leadership vs. top end Intel Xeon

Throughput performance leadership at same thread count

Qualcomm Centriq is a product of Qualcomm Datacenter Technologies, Inc

8

Why the Centriq™ 2400 Server Processor is a good fit for Cloudlets / Edge

• Performance per watt and low

power leadership

– suitable for limited power

budget at edge cloud

– Cloudflare:

“Although it has a TDP of

120W, during my tests it

never went above 89W (for

the go benchmark). In

comparison Skylake and

Broadwell both went over

160W, while the TDP of the

two CPUs is 170W.”

3.8 4.0 3.9

5.5 5.3 5.1

SP

EC

int®

_ra

te2

00

6 E

stim

ate

pe

r T

DP

W

45% better 32% better 31% better

Intel Xeon

Qualcomm

Centriq

Qualcomm Centriq 2460 120W TDP

vs.

Intel Xeon Platinum 8180 205W TDP*

Qualcomm Centriq 2452 120W TDP

vs.

Intel Xeon Gold 6152 140W TDP*

Qualcomm Centriq 2434 110W TDP

vs.

Intel Xeon Silver 4116 85W TDP*

12,403

16,31115,393

161165

72

0

50

100

150

200

0

6,000

12,000

18,000

Broadwell Skylake Centriq

Pow

er c

on

sum

pti

on

(w

)

Req

ues

ts/

seco

nd

NGINX test data CPU powerNGINX

Equivalent performance: Qualcomm Centriq 46-core comparable to two Skylake 12-core processors

Performance per Watt leadership vs. top end Intel Xeon

https://blog.cloudflare.com/arm-takes-wing/ : Centriq “managed to get 214 requests/watt vs the Skylake’s 99 requests/watt and Broadwell’s 77”.

Qualcomm Centriq is a product of Qualcomm Datacenter Technologies, Inc

9

Why the Centriq™ 2400 Server Processor is a good fit for Cloudlets / Edge

• More performance per CPU$

and TCO leadership

– Cost efficiency is essential

for edge computing to be

viable

• Analysis by Cloudflare

– Improved server density

per rack at same power

>> cost efficiency

Qualcomm Centriq 2460 120W TDP

vs.

Intel Xeon Platinum 8180 205W TDP*

Qualcomm Centriq 2452 120W TDP

vs.

Intel Xeon Gold 6152 140W TDP*

Qualcomm Centriq 2434 110W TDP

vs.

Intel Xeon Silver 4116 85W TDP*

12,403

16,31115,393

161165

72

0

50

100

150

200

0

6,000

12,000

18,000

Broadwell Skylake Centriq

Pow

er c

on

sum

pti

on

(w

)

Req

ues

ts/

seco

nd

NGINX test data CPU powerNGINX

Equivalent performance: Qualcomm Centriq 46-core comparable to two Skylake 12-core processors

Performance per CPU $ vs. top end Intel Xeon Platinum, Gold, and Silver

0.080.15

0.330.33

0.46

0.64

2xbetter

SP

EC

int®

_ra

te2

00

6 E

stim

ate

pe

r C

PU

$

Intel Xeon

Qualcomm

Centriq

3xbetter

4xbetter

40servers

60servers

Skylake Centriq

Improved density

Can fit 60 Qualcomm servers per cabinet using same power as 40 Intel Skylake servers

Qualcomm Centriq is a product of Qualcomm Datacenter Technologies, Inc

10

Enabling Edge Computing on Qualcomm Centriq™ Arm-based Servers

• Data Plane performance is critical for most

Edge Computing use cases

• Network scaling from 10 to 40 to 100GigE

consumes progressively more CPU resources

• As a result of data plane processing optimization and the

higher core count of Centriq:

– Centriq™ 2400 offers head room for significantly more

additional compute at same networking line rate

performance as comparable Intel Skylake

Edge-specific

Optimizations

Packet processing, Containers,

Networking acceleration

Leverage NFV

Infrastructure

OPNFV Release D, E

OpenStack Nova, Neutron, Keystone

Virtualization &

Containers

KVM, Kubernetes, Docker

Data Plane

Optimizations

DPDK, OVS, FD.IO/VPP

NIC/SmartNIC

HW & SW

Optimizations

Mellanox, Netronome

* Based on QDT internal benchmarking of Centriq™ 2400 with Intel Skylake Gold 6152

Ecosystem Components

1111

Edge Computing Use Case on Centriq™: Cloud Gaming

48 Cores.

More Cloud Apps. More Instances Serviced.

Qualcomm Centriq™ 2400 Server Platform

OS: Ubuntu

LXD + AnBox

Android Runtime

Gaming

App

Container

Android Runtime

Gaming

App

Container

32-bit to 64-bit Binary Translator

• Game logic runs on

server

• Real time rendering

in the client

(Android, iOS, any

other OS)

• Client takes user

input in real time

Qualcomm Centriq is a product of Qualcomm Datacenter Technologies, Inc

• Exceptional compute density for a high

number of game instances per CPU socket

• Full gaming experience live at >60fps

• Real time response with low latency

• Low CPU utilization

1212

Call for Action.

As edge computing popularizes, edge computing

SW infrastructure will proliferate.

We (the ecosystem) need to work together to ensure it all gets ported to and optimized

on Arm!

• http://www.etsi.org/technologies-clusters/technologies/multi-

access-edge-computing

• https://www.openstack.org/edge-computing/

• https://www.openfogconsortium.org/

• http://openedgecomputing.org/

• https://www.akraino.org/

• http://www.telecominfraproject.com/project-groups-2/access-

projects/edge-computing/

• … …

Enabling Edge Computing on Arm via Industry Initiatives

Nothing in these materials is an offer to sell any of the

components or devices referenced herein.

©2018 Qualcomm Technologies, Inc. and/or its affiliated

companies. All Rights Reserved.

Qualcomm is a trademark of Qualcomm Incorporated,

registered in the United States and other countries. Other

products and brand names may be trademarks or registered

trademarks of their respective owners.

References in this presentation to “Qualcomm” may mean Qualcomm

Incorporated, Qualcomm Technologies, Inc., and/or other subsidiaries

or business units within the Qualcomm corporate structure, as

applicable. Qualcomm Incorporated includes Qualcomm’s licensing

business, QTL, and the vast majority of its patent portfolio. Qualcomm

Technologies, Inc., a wholly-owned subsidiary of Qualcomm

Incorporated, operates, along with its subsidiaries, substantially all of

Qualcomm’s engineering, research and development functions, and

substantially all of its product and services businesses, including its

semiconductor business, QCT.

Follow us on:

For more information, visit us at:

www.qualcomm.com & www.qualcomm.com/blog

Thank you