DataCenter Network Products

Huawei DataCenter Network Products and Solution
Pre-sales Training
Page 0 Copyright © 2020 Huawei Technologies Co., Ltd.
Contents
• Trend and Strategy
• CloudEngine Switch Hardware Features
• CloudFabric Solutions Introduction
• Success Stories

Creating Opportunities Out of Data Center Challenges
Accelerated Cloud Computing Deployment Big Data
2014 2020
3.4 ZB 40 ZB
Fast Growth of SDN
2015 2019
US$2 billion US$ 18 billion
93% of enterprises surveyed have run IaaS 70% of SDN applications are deployed
applications or are assessing the IaaS model in data centers and enterprises

Exponential Growth of Traffic in the Tide of Big Data
Rapid expansion of data centers Requirement on Larger networks

with higher capacities
US$15 billion new

investments worldwide every
year
10 X more servers in 2020
•10 X larger data center networks

50 X growth in data •50 X higher network capacity
worldwide by 2020 •100 X faster server port speed

Exponential Growth of Traffic in the Tide of Big Data
vDC1 vDC2 vDCn
Network
Virtualization
Virtual
Data Center Trends Network
• Rapidly increasing
investments in
Virtual
servers and network
Computing
devices
• Low usage of
Server
hardware device Virtualization VM VM VM VM VM VM
resources Virtual
Storage
Storage
Virtualization Tenant1/Dept.1 Tenant2/Dept.2 TenantN/Dept.N
How to improve the usage of network resources?

Challenges How to flexibly deploy and migrate VMs?
in Network How to provision tenant networks on demand? How to isolate a large number of tenants?
How to implement smooth evolution from traditional networks to virtual networks while protecting initial
Virtualization investment?

Quick Cloud Service Update, But Slow DCN Deployment
200% growth of DC apps

every 4 years
Requirement Manual Service

analysis & network commissionin
planning: 4 to 8 configuration: g: 4 to 10
320 apps launched every day weeks 3 to 4 weeks weeks
DC network deployment
Servers/storage devices
deployed in 3 days 3 to 6 months
Slow DC network deployment hinders service innovation

Higher-Quality Networks Required for Cloud Data Centers
Users/Tenants: Require 24/7 non-stop

services X7
ISPs/ICPs: Network failures cause

customer and economic losses
Finance/Government customers:
… Require lossless, reliable, secure data
centers for efficient service operation 99.999%

Higher-Quality Networks Required for Cloud Data Centers
Elastic:
19.2 Next-generation core engine, providing the Virtualization: Comprehensive virtualization capabilities,
world's highest performance creating simple, efficient networks
• VS: 1:16 virtualization capability CSS: N:1 virtualization
• 1548 Tbps ultra-large switching capacity for smooth simplifies management.
evolution over the next 10 years • M-LAG: Reliable active-active system with dual management
and control planes.
• 19.2 Tbit/s high-speed high-density line card, supporting • VXLAN+BGP-EVPN: Intra- and inter-DC network
access of 27k 10GE servers virtualization.
• 24 GB ultra-large buffer to cope with traffic surge • VXLAN Mapping: Implement interconnection between
multiple DCs at Layer 2.
Agile:
` Full programmability for agile Quality: High quality, providing Easy O&M: Real-time health monitoring,
service provisioning best-in-class quality implementing minute-level fault locating
• Network programmability: SDN- • CLOS architecture, and VOQ-based

cell switching mode • Visible: gRPC-based, subsecond-level, and
based E2E data center network
high-precision data collection using the
using standard Network Service • Patented front-to-back airflow design
isolates cold and hot air channels subscription mode
Header (NSH) and Segment
• Clear: millisecond-level cache monitoring,
Agile Routing (SR) technologies. • Dynamic large buffer ensures non-
blocking switching providing high-precision monitoring data
• System programmability: The • Precise: network-wide path detection,
• Lossless Ethernet, constructing highly
VRP supports extensive APIs such achieving minute-level network fault
reliable HPC networks
as OPS to quickly meet service demarcation
• MACsec hardware encryption, secure
requirements. and reliable

Contents
• Success Stories

Portfolio of CloudEngine Series Switches (Mainstream Models)
Core switches Access switches

10GE TOR switch
CloudEngine 16800
CloudEngine 6881-48S6CQ
10GE + 25GE TOR switch (new model)
CloudEngine 6881E-48S6CQ
25GE TOR switch
10GE TOR switch
CloudEngine 16816 CloudEngine 16808 CloudEngine 16804 CloudEngine 6820-48S6CQ

(VXLAN not supported)

Portfolio of CloudEngine Series Switches
Core switches Access switches

10GE large-buffer
TOR switches
CE12800 Switches with
10GE TOR switches
flexible cards
CE6870-48S6CQ-EI
CE6870-48T6CQ-EI
CE8861-4C-EI CE6856-48T6Q-HI
CE6875-48S4CQ-EI
CE12816 CE12808 CE12804
100GE switches
CE12800S
25GE TOR switch GE TOR switches
CE8850-64CQ-EI
CE12808S CE12804S CE6865-48S8CQ-EI CE5880-48T6Q-EI

CE8850-32CQ-EI

Contents
• Modular Switch
• Fixed Switch
• Success Stories

CloudEngine 16800 Series Models and Cards
Item CloudEngine 16800
Large Capacity 19.2 Tbit/s per slot
AI Engine Embedded AI chip
Smooth Evolution 400GE smooth evolution
36*100GE 36*40GE 48*10GE
18*100GE 24*40GE

CloudEngine 16800 with AI-Powered MPUs and Embedded AI
Chips, Supporting AI Fabric Networking
Applicable to the
CE16804 and 16808
CE-MPUE-HALF
Embedded
AI chip
CE16804 CE16816 CE16808 Applicable to

the CE16816
CE-MPUE-FULL

CloudEngine16800: Orthogonal Architecture Without
Backplane Cabling
Long Backplane-
backplane free cabling
cabling
FIC
LPU
Fabric LPU
Traditional Switching CloudEngine 16800

Architecture Orthogonal Architecture
Orthogonal Architecture of LPUs and SFUs
• On core switches, the cabling between LPUs and SFUs greatly affects the slot bandwidth.
• In the traditional architecture, the length and rate of the backplane link affect the bandwidth and evolution
capability of the entire system.
• The CloudEngine 16800 uses an orthogonal architecture, which requires no backplane cabling. This
architecture greatly improves system bandwidth and scalability.

CloudEngine 16800 Patented Heat Dissipation Technology
Airflow Direction in a DC VC Phase Change Heat Dissipation

Industry-first Mixed-flow Fan
Equipment Room Technology
Heat dissipation fin
Back Front Front Front
Rack Rack Rack

Hot
AC
Cold Cold
Hot
Typical hot and cold airflow in a DC Chip

VC heat dissipation
substrate
• Strict front-to-back airflow design, which is suitable for cabinet arrangement and equipment room planning and reduces
power consumption of air conditioners in the equipment room
• VC phase change heat exchange technology, which uses evaporation-condensing phase change mechanism to quickly and
efficiently extend heat from the chip to the VC heat dissipation substrate, improving heat dissipation efficiency by four times
• Industry-first mixed-flow fan, provides the air volume three times higher than the industry average and has a unique
magnetic permeability motor and the mute defector ring to effectively reduce noise.

Smooth Evolution to 400GE, Supporting the Network Life Cycle
of Fourth-generation Servers
Server 10GE/25GE Server 25GE/100GE Server 100GE/200GE

Core high-density Core high-density Core high-density
40GE/100GE 100GE/400GE 400GE/800GE
25GE SerDes, compatibility 56GE SerDes, compatibility 112G SerDes,

with 10GE with 25GE compatibility with 56GE
S
S
S
10GE/40GE line 40GE/100GE line 100GE/400GE line
card card card
F
F
F
U
U
U
100GE line card 400GE line card
U
800GE line card
Smooth evolution, supporting different scenarios

and future evolution requirements

MPUs of the CloudEngine 16800
MPU Description
Half-width MPU, adapting to the CloudEngine
CE-MPUD-HALF
16804/CloudEngine 16808
CE-MPUD-FULL Full-width MPU, adapting to the CloudEngine 16816
• The CloudEngine 16804/CloudEngine 16808 uses half-width MPUs,

Half-width MPU of the CloudEngine and active and standby MPUs are installed side by side.
• The CloudEngine 16816 uses full-width MPUs, and the active and
standby MPUs are arranged vertically.
• HiSilicon CPU
• 16-core, single-core 1.8 GHz
• Memory: 8 GB
Full-width MPU of the
CloudEngine 16816 • Integrated AI chip
• 1588v2

SFUs of the CloudEngine 16800
SFU Performance
CE-SFU04G-G 8.4 Tbit/s
SFU04 CE-SFU04F-G 4.2 Tbit/s
CE-SFU08F-G 8.4 Tbit/s
SFU08
CE-SFU16F-G 16.8 Tbit/s
SFU16

Mapping Between Cards and SFUs of the CloudEngine 16800
Number of SFUs
Device Model Card SFU Required for Line-rate
Forwarding
36*100GE CE-SFU04G-G/ CE-SFU08G-G/CE-SFU16G-G 5
CE-SFU04F-G/ CE-SFU08F-G/CE-SFU16F-G 4
36*40GE
CE-SFU04G-G/ CE-SFU08G-G/CE-SFU16G-G 4
CloudEngine 16804/CloudEngine 48*10GE
18*100GE
24*40GE
Remarks: To use the preceding cards, you need to configure only six SFUs. The SFUs must be installed in slots 3, 4, 5, 6, 7, and 8.

High Performce Data Center Switch: CE12800
CE12800 Item CE12804S CE12808S CE12804 CE12808 CE12812 CE12816

Stable core switch for 10
years' network evolution
Switching
capacity 30/258 59/516 45/258 89/516 134/774 178/1,032
(Tbit/s)
Packet
forwarding 17,280 34,560 17,280 34,560 51,840 69,120
rate (Mpps)
Line card
• High switching capacity: 1032 Tbps 4 8 4 8 12 16
slots
• High-density 10GE/40GE/100GE
line cards
Architecture Clos architecture, cell switching, VoQ, and distributed large buffer
• 24 GB super-large buffer and 4M
FIB entries per line card
Switch
• 100G transceiver, 80 km super- 2 4 6
fabric units
long transmission distance

CE12800: High-Density 100G/40GE/10GE Line Cards
48*10GE optical +
36*100GE QSFP28 12*100GE QSFP28 24*40GE QSFP+ 2*40GE + 2*100GE
4*100GE CFP 36*40GE QSFP+ 48*GE Base-T
Port
100GE Line Card 40GE Line Card 10GE Line Card GE Line Card
Speed
CE- CE- CE- CE- CE- CE- CE- CE- CE- CE- CE- CE-
Card CE-
L04CF- L36CQ- L36CQ- L36CQ- L12CQ- L36LQ- L48XS- L48XS- L48XS- L48XS- L48XS- L48XS- CE-L48GT-EA
name L24LQ-FD
EF FD FG FD1 FD FD FDA FD FG FD1 EF ED
48*10GE
+ 2*40GE
36*100G 36*100G 36*100G 12*100G +
4*100GE E/36*40 E/36*40 E/36*40 E/12*40 36*40GE 2*100GE
24*40GE/9 48*10GE
Port /8*40GE/ GE/144* GE/144* GE/144* GE/48*2 /144*10 /40GE or 48*10GE 48*10GE 48*10GE 48*10GE 48*GE Base-T
6*10GE /25GE
40*10GE 25GE/14 25GE/14 25GE/14 5GE/48* GE 52*10GE
4*10GE 4*10GE 4*10GE 10GE +
8*25GE/
10GE

CE12800: Line Card Main Specification
Attribute 100GE Line Card 40GE Line Card 10GE Line Card GE Line Card
CE- CE- CE- CE- CE-
CE- CE- CE- CE- CE- CE- CE- CE-
Card name L36CQ- L36CQ- L12CQ- L48XS- L48XS- CE-L48GT-EA
L36CQ-FG L04CF-EF L36LQ-FD L24LQ-FD L48XS-FD L48XS-FG L48XS-ED L48XS-EF
FD FD1 FD FDA FD1
MAC 750K 256K 750K 256K
ARP ARP
extended extended
ARP 750K 1M 104K 750K 750K 750K 750K 750K 1M 750K 256K 128K
resource resource
mode:1M mode:1M
IPV4 FIB 380K 2M 750K 380K 1M 380K 380 380K 380K 4M 256K 256K 1M 32K
IPV6 FIB 128K 512K 140K 128K 512K 128K 128K 128K 128K 512K 128K 128K 512K 8K
IPv4
Multicast 32K 60K 32K 60K 60K 16K
routes
IPv6
Multicast 8K 2K
routes
VRF 16384 4096 16384 4096
MACSEC N N N N N N N N N N N N N N
ACL 120K 80K 80K 40K 20K 60K 40K 20K 20K 20K 20K 20K 20K 10K
Y(V2R5 Y(V2R5
VXLAN L3 GW Y and later and later Y Y Y Y Y Y Y Y Y Y Y
versions) versions)
Number of
Broastcast 32K 4K 32K 4K
Domains
Buffer 24G 16G 16G 8G 6GB 12G 8G 4G 1G 4G 4G 6GB 6GB 2GB

Contents
• Modular Switch
• Fixed Switch
• Success Stories

Recommended CE Series Switch Model — CloudEngine 6881
Parameter CloudEngine 6881-48S6CQ

48*10GE SFP+ and 6*100GE QSFP28 (Each
Port model QSFP28 port can be used as one 40GE QSFP+
port.)
Switching capacity 2.16 Tbit/s
Forwarding
performance 940 Mpps
Maximum number
16
of stacked switches
Buffer capacity 42 MB
FIBv4/FIBv6: 256K/80K
Performance
MAC: 256K
specifications
ARP: 256K
Four fan trays (one fan 1+1 power redundancy

 Diversified DC features: M-LAG, iStack, VXLAN, and BGP EVPN
module in each tray)  Hardware-based BFD
 Telemetry and ERSPAN enhancement
 Microsegmentation and NSH

CloudEngine 6881E-48S6CQ: A New TOR Switch with 10GE
+ 25GE Ports
 Rich data center features: M-LAG, iStack, VXLAN, and BGP-EVPN
CE6881E-48S6CQ panel view  Hardware-based BFD: minimum packet sending interval of 3.3 ms
 Telemetry and enhanced ERSPAN
Front view
Parameter CE6881E-48S6CQ
Twenty-four 25GE SFP28 ports, twenty-four
Port type 10GE SFP+ ports, and six 100GE QSFP28 ports
1-12: 10GE 13-36: 25GE 37-48: 10GE 6*100GE (compatible with 40GE)
Rear view Forwarding

940 Mpps
performance
Maximum number of
16
switches in a stack
Four fan modules (one fan in 1+1 power module
each fan module) redundancy
Performance
FIB (v4/v6): 256K/80K, MAC: 256K, ARP: 256K
specifications


48*25GE SFP28 and 6*100GE QSFP28 (Each
port.)
CloudEngine 6863-48S6CQ Switching capacity 3.6 Tbit/s

Forwarding performance
940 Mpps
Maximum number of
16
stacked switches
FIBv4/FIBv6: 256K/80K
Performance
MAC: 256K
specifications
ARP: 256K
Four fan trays (one fan module in 1+1 power

 Diversified DC features: M-LAG, iStack, VXLAN, and BGP
each tray) redundancy EVPNHardware-based BFD


48*10GE SFP+ and 6*100GE QSFP28 (Each
port)
Switching
2.16 Tbit/s
capacity
Forwarding
470 Mpps
performance
Maximum
number of 16
stacked switches
Four fan trays (one fan 1+1 power redundancy Performance

FIB (v4/v6): 256K/80K, MAC: 256K, ARP: 256K
module in each tray) specifications
 Diversified DC features: M-LAG, iStack,

 Not support VXLAN, and BGP EVPN

CE8861 with Flexible Cards
Interface side view Parameter CE8861-4C-EI

Interface type Flexible cards
2U Maximum number of
9
devices in a stack
Flexible cards, and five types of line cards with Forwarding
different rates 2030 Mpps
performance
Buffer 32 MB
16*40GE 24 x 10GE electrical interfaces FIB (v4/v6): 380K/256K
+ 2 x 100GE
Performance
MAC: 288K
specifications
ARP: 168K
24 x 25GE optical
8*100GE
interfaces + 2 x 100GE  25GE access switch with flexible cards, supporting VXLAN and
BGP EVPN
Double fan trays (two fan 1+1 power  4 card slots, 5 types of cards, flexible combination, building
Front panel view
modules in each tray) redundancy flexible and high-density access and aggregation (Cisco does
not provide models with flexible cards)
 Hardware BFD
 Telemetry, INT (IOAM), and ERSPAN enhancement
 Microsegmentation
 AI Fabric (dynamic ECN, fast CNP, VIQ, and DLB)

CE Series Switch Model — CE8850-64CQ-EI
Parameter CE8850-64CQ-EI
64*100GE QSFP28 ports (Each port can
Port model adapt to a 40GE QSFP+ port or be split into
four 25GE SFP28 ports)
Switching capacity 12.8 Tbps

Forwarding
4482 Mpps
CE8850-64CQ-EI performance
64*100GE/40GE QSFP28 ports Maximum number
16
of stacked switches
1588v2, burst traffic detection and statistics
Supported features collection capability, VIQ algorithm, and
fast ECN
 Diversified DC features: M-LAG, iStack, VXLAN, BGP-EVPN,

1588v2
 Hardware BFD
1+1 AC power
Three fan trays  Telemetry, and ERSPAN enhancement
redundancy

CE Series 100G Switch Model —CE8850-32CQ-EI
Parameter CE8850-32CQ-EI
32*100GE QSFP28 ports (Each port can
Port model adapt to a 40GE QSFP+ port or be split into
four 25GE SFP28 ports)
CE8850-32CQ-EI Switching capacity 6.44Tbps

32*100GE/40GE QSFP28+2*10G SFP+
Forwarding
3200Mpps
performance
Double fan trays (two fans in each tray)
Maximum number
16
of stacked switches
Buffer capacity 22MB
1+1 power redundancy  Diversified DC features: M-LAG, iStack, SVF, VXLAN, BGP-EVPN

CE Series Switch Model — CE6870
Port-side view Parameter CE6870-48S6CQ-EI/CE6870-48T6CQ-EI
Downlink: 48*10GE SFP+ ports, 48*10GE Base-T
ports
1U Uplink: 6*100GE QSFP28 ports (Each port can
Port model
adapt to a 40GE QSFP+ port or be split into four
48*10GE SFP+ 6*100GE QSFP28 25GE SFP28 ports.)
Switching Capacity 2.16 Tbps
1U Forwarding
720 Mpps
performance
48*10GE Base-T 6*100GE QSFP28 Maximum number
9
of stacked switches
Front panel view Buffer capacity 4 GB

Double fan trays (two fans in each tray) Performance MAC: 750K
specifications IPv4 FIB: 380K
 Diversified DC features: M-LAG, iStack, VXLAN, BGP-EVPN

 4 GB super-large buffer, suppressing flooding traffic on the
1+1 power redundancy network

Parameter CE6865-48S8CQ-EI
Downlink: 48*25GE SFP28 ports
Uplink: 8*100GE QSFP28 ports (Each
Port model port can adapt to a 40GE QSFP+ port or
be split into four 25GE SFP28 ports.)
CE6865-48S8CQ-EI Switching capacity 4.0 Tbps

48*25GE SFP 28 ports, 8*100GE/40GE QSFP28 ports Forwarding
2000 Mpps
performance
Maximum number
16
of stacked switches
 Diversified DC features: M-LAG, iStack, VXLAN, BGP-EVPN,

Two fan trays and 1588v2
 Hardware BFD
1+1 AC power redundancy  Telemetry, INT (IOAM), and ERSPAN enhancement
CE6865-48S8CQ-EI  Microsegmentation

Parameter CE5880-48T6Q-EI
Interface side view Downlink: 44*GE RJ45, 4*10GE RJ45
Uplink: 6*40GE QSFP+ (The first two 40GE
Interface type interfaces can be split into 10GE interfaces,
and the last four 40GE interfaces cannot be
split.)
1U
Maximum number of
16
devices in a stack
6*40GE QSFP+
44*GE RJ45+4*10GE RJ45 Switching capacity 648 Gbit/s
Forwarding
406 Mpps
performance
Buffer 16.5 MB
Front panel view FIB (v4/v6): 128K/64K
Double fan trays (two fan Performance
MAC: 176K
specifications
modules in each tray) ARP: 128K
 ToR switch with GE downlink and 40GE uplink,

supporting VXLAN and BGP EVPN
 Hardware BFD
 Accurate time synchronization: 1588v2
 Microsegmentation
1+1 power redundancy  Telemetry and ERSPAN enhancement

Contents
• Success Stories

Huawei CloudFabric DCN Solution Overview
Application Huawei CloudEngine:
layer Gartner Peer Insights Customers' Choice distinctions twice
Finance Internet Large enterprise Government

Cloud platform layer/Management • High SDN deployment efficiency
Intelligent
and control layer • 1-3-5 intelligent O&M
vCenter * 1-3-5: A fault is discovered within 1 minute,
Virtual located within 3 minutes, and rectified within 5
management Huawei CloudStack Autonomous
platform
minutes.
driving network
• Embedded AI chip + innovative

iLossless algorithm
Fabric Ultra-broadband • Ethernet with zero packet loss, 30%
network
computing power
CloudEngine 16800 Lossless

CloudEngine 16800
DCN switch with
large capacity
• High density
VM
• Gartner Peer Insights Customers' Ethernet with zero
VM VM Container BM Server Choice distinctions twice packet loss
VM

Contents
• iMaster NCE
• AI Fabric
• iMaster NCE-FabricInsight
• Software Features
• Network Architecture
• Success Stories
Autonomous Driving: Why Does a Cloud Network Require an
SDN Controller
Challenges: Cloud network requirements SDN: Why Is a Controller Required
① Northbound: Provides standard

Neutron interfaces to connect
to the orchestration layer.
1 Network capabilities need to be open to Service ② Southbound: Shields
orchestration layer differences at the device layer
services, and networks need to change and unifies service models.
with the cloud accordingly. ③ Operation and management: Provides
unified management and O&M of
physical and virtual networks.
④ Management Scale: Huawei single
cluster controller can manage 4200
switches.
2 Cloud network resources need to be
dynamically scheduled and elastically Fabric
allocated.
vSwitch vSwitch CE1800V CE1800V
VM VM VM VM
VM VM Hypervisor VM VM
VM VM VM VM
SDN architecture

iMaster NCE Key Changes 2-3-4
SDN automatic service deployment Unified data foundation Full-lifecycle management

AI-based intelligent analysis, Perception and location Optimization of simulation
prediction, and troubleshooting processing verification monitoring
Planning + construction +
Automation + Intelligence Manager + Controller + Analyzer
maintenance + optimization
2 3 4
Autonomous Driving
NMS Controller Analyzer Network system
Network =

iMaster NCE@ CloudFabric Data Center Network Solution
B2C B2B iMaster NCE-Fabric
Application
Provide simplified and automated DCN experience throughout the life cycle
layer
Industry DCI private Northbound abstraction of network resources and services, and southbound
Game WeChat Video VPC cloud line adaptation to different devices and networks
Big Data Automatic E2E network deployment, ‘0’ wait time

Cloud  Ultra-fast network provisioning: Graphical drag-and-drop
platform operations.
FusionSphere
layer/ FusionStage
Management  Fast container rollout: The container rollout rate is 10,000/min
and control OpenStack Pre-evaluate change risks, ‘0’ configuration error
layer
 Pre-event simulation: Evaluate the impact of the configuration
and resources to be changed on the network,
 Post-event verification: configuration plane verification for issues
such as underlay network connectivity, interfaces, and routes
Network
Layer Intelligent recovery of faults, ‘0’ service interruption
 Fault 1-3-5:75 Typical Faults, Perceiving in 1 Minute,
VM Physical Locating in 3 Minutes, and Recovery in 5 Minutes
VM VM Container
VM machine
 Network health: Comprehensively evaluate network health
based on service experience to proactively predict faults.

Contents
• iMaster NCE
• AI Fabric
• Success Stories
AI Fabric: The AI ECN Algorithm Facilitates Network Intelligence
and Implements Network Self-optimization
Software and hardware AI ECN algorithms,
AI ECN principles: Adaptive adjustment of the AI ECN solution: implementing network-wide traffic optimization in
ECN waterline
Traffic model the spine-leaf architecture
① Optimal
trained
waterline
(DQN
Traffic model algorithm) CloudEngine 16800 CloudEngine 16800
determination (CNN Security
classification policy
algorithm)
② Heuristic AI chip
Traffic model search
not trained
High waterline Low waterline H a r dwa r e A I ECN

Data Queue 1
collection Queue 2
Queue N Deliver the optimal

Forwarding chip
Current queue
depth ECN waterline
① Real-time awareness of network traffic models based on AI Software AI ECN

algorithms and adaptive adjustment of the ECN waterline
CloudEngine 8850 CloudEngine 6865 CloudEngine 8861
② AI models provide good generalization ability and heuristic
search to cope with unknown traffic models.

AI Fabric: 45% Performance, Three-Layer Cross-POD
Networking, and Three-Network Convergence
In the past: oriented at Now: DCN for the AI era, compatible
Effect
the single AI scenario with the cloud era
AI 1.0 AI 2.0
Algorithm AI ECN Algorithm
Dynamic ECN Algorithm
① Performance: 45% higher than the
+
upgrades
static ECN in the industry
Embedded AI chip AI algorithm
TOR
TOR + modular
Two-Layer Networking
PFC deadlock Three-Layer Networking ② Scale: single-POD networking
prevention and multi-POD networking
Single-POD
Multi-POD
Service Distributed
network storage Service network/Distributed storage ③ Three-network convergence: The
TCP/RoCE hybrid
service network, storage network,
transmission ROCE/TCP and HPC are converged on the
ROCE TCP Converged on Ethernet.
Separated one network
Ethernet

Contents
• iMaster NCE
• AI Fabric
• Success Stories
Trend: Traditional Manual O&M -> AI-Powered Intelligent O&M
Manual
packet Manual step-
Manual fault obtaining to by-step
6.48
identification locate faults fault isolation 30%
2.8 identified by
1.6 2.0 Abnormal
0.09 0.63 1.1 traditional O&M
traffic:
Media Healthcare Retail Manufa- Telec- Energy Finance 3.65% of
cturing om 70%
network-
unidentified by
wide traffic
traditional
Source: Network Computing, the Meta Group and O&M
Contingency Planning Research
Zero fault tolerance Difficult to detect faults Difficult locating

Loss of millions of US dollars 85%+ of network faults are Average duration for
per hour due to system detected until service complaints locating a fault: 76
downtime are received. minutes

iMaster NCE-FabricInsight : Inevitable on Data Center Networks
"1-3-5" Intelligent O&M Health check: Network Health
1 minute to detect faults, 3 minutes to locate faults, Assess network health based on service experience and
and 5 minutes to eliminate faults perform health checks periodically to prevent faults
Root cause analysis Fault prediction
Fault identification Fault rectification Capacity forecasting Health management
Faults occurred Possible faults

“1-3-5” Network Health
Troubleshooting
AI+ Knowledge-based Inference, Quickly Pinpointing Fault Root Causes
Collection Analysis Decision
Intelligent analysis engine
Knowledge-based BGP
flapping
OSPF
flapping Interface
inference engine flapping
Huawei 30+ years'

O&M expertise Router
ID
IS-IS
flapping
BFD
flapping
Manual
conflict
recovery
Anomaly detection
Continuous learning
and training Root cause analysis
based on real Knowledge
point 1
Knowledge
point 2
Knowledge
point 3
Knowledge
point 4 Risk prediction
site faults
Model application Intent-based loop
Data AI-powered Network closing
Holographic data anomaly object
of data centers cleansing
identification modeling
Service flow data/Telemetry data...
iMaster NCE-FabricInsight iMaster NCE-Fabric

Troubleshooting
Device Issues
Switch Function Fabric Loop-
Issue Dependencies
Version Closure Action
Switch Fault
Repeated Switch Restarts
13 types 36 26 types
Switch Management Channel Interruption
Switch LPU Fault
Repeated Switch LPU Fault
Switch MPU Fault
of monitored metrics of issues Repeated Switch MPU Fault
objects Switch SFU Fault
Repeated Switch SFU Fault
Switch Fan Fault
Switch Power Fault
Switch CPU Threshold Exceeded V200R005C00
or later Loop-closure actions
Switch Memory Threshold Exceeded vary according to
Value-added faults:
Switch MAC Entry Threshold Exceeded package of network 1. Warnings
Switch ARP Entry Threshold Exceeded health 2. Loop-closure
Telemetry | Syslog | ERSPAN Layer 2 Loop suggestions
3. Loop-closure plan
License Expried
Switch FIB4 Entry Threshold Exceeded
Switch SFU Forwarding Performance Insufficiency
Switch ACL Entry Threshold Exceeded
Switch Storage Space Threshold Exceeded
Switch FIB4 Entry Specifications Exceeded
Switch FIB6 Entry Specifications Exceeded
V200R019C00
Switch FIB6 Entry Threshold Exceeded
or later
Mismatch Between SFU Forwarding Performance
and LPU Line-Rate Forwarding Capability V200R019C10
or later
Switch ND Entry Threshold Exceeded

Troubleshooting
Network Issues
6 types
18 10 types Issue
Switch
Version
Function
Dependencies
Fabric Loop-
Closure Action
Intermittent Link Disconnection
of monitored metrics of issues
Unidirectional Link Connectivity
objects Fault on Network Side of a
Switch
Switch Physical Port Suspended
V200R005C00
Switch Port Error-Down
Loop-closure actions vary
Service Affected by Switch Port Value-added according to faults:
Congestion 1. Warnings
package of network
2. Loop-closure
Telemetry | Syslog Service Port Blocked by STP
health suggestions
3. Loop-closure plan
Suspected Optical Link Fault
Optical Module Type Mismatch
ARP Attack V200R019C10
ND Attack

Troubleshooting
Protocol Issues
3 types 8 6 types Issue Switch Version

Function
Dependencies
Fabric Loop-
Closure Action
Two Master Switches in

of monitored metrics of issues M-LAG
objects
Neighbor Relationship
Flapping Due to an
Incorrect Update Packet
Received by Switch
Loop-closure
OSPF Router ID Conflict actions vary
according to
OSPF Interface IP Address
Value-added faults:
Conflict V200R005C00
Config | Syslog or later
package of 1. Warnings
BGP Neighbor network health 2. Loop-closure
Relationship suggestions
Establishment Failure Due 3. Loop-closure
to Incorrect plan
Configurations
OSPF Neighbor
Relationship
Establishment Failure Due
to Incorrect
Configurations

Troubleshooting
Overlay Issues
3 types 10 4types
of monitored metrics of issues Function
Fabric Loop-
objects Issue Switch Version Closure
Dependencies
Action
Expired VXLAN License

Loop-closure
NVE Source IP Address Deleted actions vary
according to
Value-added faults:
V200R005C00
Config | Syslog | ERSPAN NVE-Associated Loopback package of 1. Warnings
or later
Interface Deleted network health 2. Loop-closure
suggestions
3. Loop-closure
NVE-Associated Loopback IP plan
Address Deleted

Troubleshooting
Comprehensive Evaluation Model, Intuitively Detecting

Network Quality
No anomaly on the
BD, VNI, and VRF resources and
Overlay virtual network plane
running status
that carries services
No anomaly of
OSPF/BGP peer protocols for network
Protocol M-LAG status
connection stability
Network Congestion and packet

Port traffic and Optical link
Network interconnection
error packets
Queue depth
status loss detection based on
port status the network link load
Whether physical
Hardware status: board, fan, Capacity: ARP, FIB, CPU and components are
Device
power supply, etc. MAC, etc. memory load normal or whether
resources overflow
Analyze 20+ types of monitored objects and 70+ indicators to intuitively display network-
wide experience quality

“1-3-5”
Troubleshooting Network Health
Network Health Evaluation, Facilitating DCN Quality Evaluation
Post-event manual inspection Pre-event risk prevention
Step 1
Health overview
Check the overall network health
metrics and trend based on five-layer
model.
Step 2
Multi-dimensional
detailed analysis
Analyze the network health
from the following dimensions
to determine the network
health trend: device, network,
protocol, service, and overlay.
Step 3
Professional report
interpretation
Summarize issues from each dimension and
periodically push reports of detection details,
facilitating identification of anomalies.

“1-3-5”
Real-Time or Periodic Push of Health Reports, Providing

References for Optimization
Network Overview Metric Details Report Details
Intuitively display the resource overview, Identify network quality issues based on Display reports in multiple dimensions,
load overview, and quality overview the five dimensions of the network identify abnormal monitoring items, and
across the entire network. health evaluation system. provide rectification and optimization
suggestions.
Provide comprehensive DCN network health evaluation based on a unique five-layer model, intelligently interpret the
Use Case
evaluation report, and push reports in real time or periodically, providing a basis for network fault rectification and
Benefits
optimization.

“1-3-5”
Use Case: Daily push of evaluation reports and proactive

network health awareness
Case: An O&M engineer at a site configures FabricInsight to automatically push the network health evaluation report at 08:00 every morning. After receiving the report on
the current day, the O&M engineer finds that the health score is lower than that of the previous day. Then, the O&M engineer detects and resolves the issue based on the
anomalies and rectification suggestions in the report. After the optimization, the health score becomes normal, and there is no open issue.
1. Click Health. On the health evaluation system page, click Report Setting in the upper right corner.
3. Immediate Sending: Select the start time and end time for sending reports,
2. Report Setting: Set the report sending time and notified and report receiving mode. The report is generated immediately. (Immediate
user group. Reports will be sent to the specified email Sending: The browser directly downloads the report; Send by email: The report
address at the specified interval. is sent to the specified email address immediately.)

Contents
• iMaster NCE
• AI Fabric
• Success Stories
VXLAN & BGP-EVPN MLAG SFC Microsegmentation
VXLAN & BGP-EVPN : Large Network Resource Pool,

Flexible Service Migration
RR RR Service A Service B Service A Service B
DC1 DC3
BGP EVPN
VXLAN
with EVPN
DC2
VNI VNI
MAC VTEP VTEP VTEP VTEP MAC
ARP ARP
... ...
Service A Service B
Scale-out within a DC L2/L3 interconnection between DCs

BGP-EVPN acts as the VXLAN control plane to:
• VXLAN-based EVPN enables L2 interconnection
• Trigger automatic VXLAN tunnel setup between VTEPs,
between DCs with different networking modes.
without manually configuring full-mesh tunnels.
• BGP EVPN advertises subnet routes so that L3 data
• Advertise host routes and MAC entries on the control plane
center interconnection can be established over VXLAN
to reduce flooding of unknown traffic and optimize packet
tunnels.
forwarding.

M-LAG: Millisecond-level Switchover Upon an Interface

Failure, and Dual-homing Access of Servers
Stacking M-LAG M-LAG Lite Ethernet Segment Identifier
Device models must be the Devices can use different When configurations of a server Dual-homing access
same, software versions of software versions, MAC fast are modified, BGP EVPN is used as the control
devices must be the same, reroute (FRR) is supported, two NICs of the server send ARP plane protocol and devices from
and devices of different series and the switchover time of an broadcast packets simultaneously. multiple vendors can be used.
and vendors cannot interwork. interface is reduced from 1s to
100 ms.
Dual-homing access Dual-homing access Dual-homing access Multi-homing access

 Control plane coupling and  Independent control plane and  Independent control plane, without  Independent control plane,
separate device upgrade
separate device upgrade service interruption during upgrade protocol-level coupling
 Proprietary protocols are used to  Proprietary protocols are used to  The peer-link is removed to save  The peer-link is removed to save
prevent loops, and links on the
prevent loops, and links on the link resources. link resources.
forwarding plane are aggregated.
forwarding plane are aggregated.  The link bandwidth usage increases  The link bandwidth usage
 The link bandwidth usage  The link bandwidth usage increases from 50% to 100%. increases from 50% to 100%.
increases from 50% to 100%.
from 50% to 100%.

SFC Solution: Standard Interconnection, Drag-and-Drop

Deployment, and Simplified and Efficient VAS Orchestration
SFP 1 Solution advantages:
VM2 VM5
 Two SFC modes: PBR and NSH
SFP 2
 NSH-based SFC: standard RFC featuring good interoperability,
Internet VM8
open ecosystem, good compatibility with third-party NSH devices
(interconnection with DPtech firewalls), and 20K SFC entries (two
Resource
pooling times higher than the industry)
deployment
Product selection for NSH-based SFC (IPv4 and IPv6):
 CloudEngine 6881, CloudEngine 6863, and CloudEngine 16800

Service Leaf
Product selection for PBR-based SFC (only IPv4):
OVS
VM
VM
VM  CloudEngine switches excluding the preceding models
VAS
resource
pool
Simplified deployment Efficient forwarding

Flexible Orchestration
Defining SFC in drag-and-drop The ACL consumption is
mode Full decoupling from the fabric reduced by more than half.

NSH-based SFC: Challenges to Networks Brought by Diversified

Data Center Security and Background of NSH-based SFC
Background: Why Is the NSH Usd Definition: NSH Background
1 Due to diversified policy deployment, ACLs become 1 NSH definition
scarce resources.
…
App 1 App 2 … App n
Switch Switch Switch Switch

2 The control plane configuration is complex
because of the chain deployment mode.
2 NSH encapsulation format
3 Security services are coupled with the physical VXLAN

topology, resulting in poor scalability.
VLAN

NSH-based SFC: The Forwarding Plane Based on Standard NSH

Encapsulation Saves ACL Resources
As-Is: Control plane solution To-Be: Forwarding plane solution
1. PBR-based SFC 2. NSH-aware SFC 3. NSH-unaware SFC

Single flow
ACL-1
Single flow Single flow
A ACL-1 ACL-1
Web App A A
Web App Web App
ACL-3
ACL ACL-2
VAS
resource
pool VAS VAS
Firewall LB IPS resource resource
pool pool
Firewall LB IPS Firewall LB IPS
• Number of consumed ACL rules: n+1 (n
indicates the hop counts of the VAS device.)
• Number of consumed ACL rules: 1 (irrelevant
• The VAS device consumes 4 ACL rules at 3 • Number of consumed ACL rules: n (n indicates
to the hop counts of the VAS device)
hops. the hop counts of the VAS device.)
• The VAS device consumes 1 ACL rule at 3 hops.
• Ecosystem: compatibility with reused VAS • The VAS device consumes 3 ACL rules at 3
• Ecosystem: The VAS device needs to support
devices hops.
NSH.
• Ecosystem: compatibility with reused VAS
devices
Microsegmentation: Network Devices Need to Securely

Isolate Traffic
Background: Why is microsegmentation introduced Principle: Finer granularity, wider dimensions,
and east-west security isolation
1 Traditional security is implemented
by configuring different service zones.
Microsegmentation
…
App 1 App 2 … App n Microsegmentation
• Conflict between cloud
sharing and security isolation
2 Traditional isolation methods

Discrete VM
have traffic detours. Organization
Subnet IP name/Cont OS type
name
address ainer
Spine
3 Due to diversified isolation policies,

ACLs become scarce resources.
Spine

Microsegmentation: From Network Segment-level to VM-level
Traditional: network segment-based Microsegmentation: discrete IP address-based

isolation, coarse-grained isolation, fine-grained and flexible
Spine Spine
Inter-subnet Inter-subnet
isolation isolation
× ×
Server leaf
Server leaf
V V V V V V V V V V V V V V V V V V V V V V V V
M M M M M M M M M M M M M M M M M M M M M M M M
Intra-subnet
Intra-subnet
isolation not Subnet 2
√ Subnet 1 isolation × Subnet 1 × Subnet 2
supported
Virus-infected VMs are
quickly isolated.
Industry average
Huawei
Network segment-based isolation: coarse-grained and Fine-grained isolation based on discrete IP addresses and VM
unable to implement fine-grained and flexible names, implementing fine-grained management and control
management and control
Microsegmentation Isolation, Implementing Precise Security Management and Control and Saving ACL Resources

Microsegmentation Provides East-West Traffic Isolation in a

Fine-grained Manner
IPv4 microsegmentation product models:
Interconnection with OpenStack Interconnection with  CloudEngine 6865, CloudEngine 8861
FusionSphere
IPv4 & IPv6 microsegmentation product models:
OpenStack
FusionSphere  CloudEngine 16800, CloudEngine 6881 and CloudEngine 6863
Unified isolation
Microsegmentation implements the zero-trust security model. It
(Secondary orchestration)
provides security isolation based on discrete IP addresses and VM
①North-south
names, and covers PMs. It can uniformly isolate traffic of VMs and
isolation
Border-Leaf
BMs.
Efficient forwarding
③ IPv6 microsegmentation Microsegmentation has a unique value in mutual access control

Spine
scenarios that have high forwarding efficiency and low security
②East-west requirements. There is no traffic bypassing problem, and the
isolation forwarding performance is not a bottleneck.
VTEP VTEP VTEP VTEP
Large specifications
OVS OVS OVS
VM VM VM
VM VM VM The mask length of the EPG member is not limited. Each EPG of
VM VM VM
BM
the industry supports a maximum of three mask lengths.
Server leaf Server leaf Server leaf
Server leaf

Contents
• iMaster NCE
• AI Fabric
• Success Stories
Network Architecture: CloudFabric SDN Baseline Networking
DC2 CloudFabric Product Selection

Stacking used if VXLAN 1. Server leaf:
Multi-active mapping is deployed M-LAG
Active-active used if VXLAN  GE access + 40GE uplink: CloudEngine 5880
Border mapping is not deployed Service  10GE access + 40GE/100GE uplink: CloudEngine 6881 (preferred), CloudEngine
leaf leaf 6856 (electrical)
Fabric 1. VAS device
in service  25GE access + 100GE uplink: CloudEngine 6865 (preferred) and CloudEngine
gateway 6863
mode
2. VAS device  Hybrid-rate access: CloudEngine 6881E (preferred), CloudEngine 8861
in bypass (preferred), and CloudEngine 16800
mode Spin
2. Spine:
e
 Core switch: CloudEngine 16800 (preferred), CloudEngine 12800
M-LAG  TOR switch: CloudEngine 8861 (preferred) , CloudEngine 8850

3. Border leaf:
Server
leaf  Core switch: CloudEngine 16800 (preferred), CloudEngine 12800
Scenario constraints:  TOR switch: CloudEngine 6881/6881E (preferred), CloudEngine 8861 (preferred),
(1) Layer 2 network: If the number of physical servers on the entire CloudEngine 6865, CloudEngine 6863, CloudEngine 6870
network is smaller than 200 or the number of VMs is smaller than 4. Service leaf:
6,000, you can combine border leaf nodes and spine nodes.  Core switch: CloudEngine 16800 (preferred), CloudEngine 12800,
(2) The network composed of fixed switches does not support the  TOR switch: CloudEngine 6881/6881E (preferred), CloudEngine 8861 (preferred),
scenario where border leaf nodes, service leaf nodes, and spine CloudEngine 6865, CloudEngine 6863, CloudEngine 6870
nodes are combined. 5. Fabric gateway (DCI Leaf)
(3) FabricInsight supported by the CloudEngine 16800, CloudEngine  Core switch: CloudEngine 16800; CloudEngine 12800
6881, and CloudEngine 6863 supports only IPv4 and does not  TOR switch: CloudEngine 6881(preferred), CloudEngine 6875 and CloudEngine
support IPv6. Overlay multicast is not supported. 6863

CloudFabric Easy: Standard Solution for Small- and
Medium-Sized DCs, Simplifying Pre-sales and Post-sales
Service requirements:
CloudFabric Easy Baseline Networking
1. Limited physical space and small

scale of the equipment room
iMaster NCE-
FabricInsight: standalone
deployment
Hosting or mini-sized equipment

Border leaf: CloudEngine 6863-48S6CQ
room with fewer than 30 cabinets Border leaf
2. Fixed services, no capacity

expansion
Spine: CloudEngine 8850
+ Server leaf
25GE/10GE optical access:
3. Standard networking,
10GE electrical access: ②
reducing delivery complexity CloudEngine 6856-48T6Q-HI
Interface
• Dual-mode NCE installation: BMs and VMs are supported.

• Pre-sales tools: solution DIY, baseline standardization, and one-click HLD export and quotation
• Post-sales channel-based: comprehensive documents, quick delivery through channels, and one-stop delivery

CloudFabric Single-Layer Networking: One Node Used as Five Device Roles in
the Spine-Leaf Architecture, Reducing Costs and Simplifying Management
Five-in-One Single-Layer Networking

Service
requirements:
Internet WAN Internet WAN Internet

② WAN
Border leaf Fabric gateway

Independent
egress
Five-in-one
Independent
① networking
Fabric gateway
VAS device
Independent Spine
Border leaf
management Service leaf
Spine
domain
VXLAN Server leaf
Server leaf Service leaf

VM VM
VM VM
VM VM
VM BM
DCs are scattered ① Five nodes (fabric gateway, border leaf node, service leaf node, spine node, and server leaf
and small in scale. node) can be integrated.
Five-in-one networking: CloudEngine product models: CloudEngine 6881/6863/16800 (The
② switches need to be connected to PE routers in the upstream direction.)

Contents
• Success Stories

Highly Recognized Performance
InterOP Awards Awards and Certifications
In 2013, the CloudEngine
12800 won the Best of Show
Award at Interop, which is the
highest exhibition in the IT
industry. Huawei is the first
Chinese provider that wins the
position.
Award of Excellent Award of the Most Award of Excellent

In 2016, the CE8860 and Product Trusted Competitive Product Product in Big Data
CE6851 won the Best of by CIO
Show Award at Interop.
Huawei's AI Fabric
Intelligent Lossless Data
Center Network Solution
Preferred Brand of
Takes Home Interop Tokyo Award of Annual
China SDN
Cloud Computing
SDN Best
Best of Show Award Excellent Technology and Network
Practice Award
Solution

CloudEngine Series Switches Serve 7800+ Global Customers
 The market share is No.1 in China and No.3 in the

 Over 32,000 CE12800 switches have been sold
world.
around the world, serving 7800+ customers in
 The global market share growth rate is No. 1 for
120+ countries.
four consecutive years.
Gartner Peer Insights

 2018 Approaching the DC SDN
Leaders Quadrant SDN hardware platform Customers’ Choice
 2017 Challenger leader for Data Center Networking

How to Get Huawei Document Resources
1: Enterprise Website
http://e.huawei.com/en/  Channel Partner Program

– To learn partnership policy
 Partnership
– To be a partner
 Material & Toolkit
– To find material and the toolkit
 Special Partner Zone
– ISV
2: Document Email 3: Document User Guide
Where can I find it

and give feedback?
1. What is Huawei...
[email protected]
Please confirm email address 2. How do I get it?
3. Where do I provide feedback?

How to Get Pre-sales Help
[email protected]
Huawei Experts Team
Partners
Support Center
 24/7 Pre-sales email and
telephone support
http://enterprise.huawei.co  Products and solutions
m/en/about/contact
Partners consulting
Please confirm addresses are still valid at e.huawei.com/en and channel sites

Thank You
www.huawei.com

DataCenter Network Products

Uploaded by

Copyright:

Available Formats

DataCenter Network Products

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

DataCenter Network Products

Uploaded by

Copyright:

Available Formats

What are some of the trends driving data center network transformation?

What are some of the trends driving data center network transformation?

What challenges does network virtualization pose for data centers?

What challenges does network virtualization pose for data centers?

Huawei DataCenter Network Products and Solution

• CloudEngine Switch Hardware Features

• CloudFabric Solutions Introduction

Page 1 Copyright © 2020 Huawei Technologies Co., Ltd.

Fast Growth of SDN

Page 2 Copyright © 2020 Huawei Technologies Co., Ltd.

Rapid expansion of data centers Requirement on Larger networks

US$15 billion new

10 X more servers in 2020

•10 X larger data center networks

Page 3 Copyright © 2020 Huawei Technologies Co., Ltd.

How to improve the usage of network resources?

Page 4 Copyright © 2020 Huawei Technologies Co., Ltd.

200% growth of DC apps

Requirement Manual Service

Slow DC network deployment hinders service innovation

Page 5 Copyright © 2020 Huawei Technologies Co., Ltd.

Users/Tenants: Require 24/7 non-stop

ISPs/ICPs: Network failures cause

Page 6 Copyright © 2020 Huawei Technologies Co., Ltd.

• Network programmability: SDN- • CLOS architecture, and VOQ-based

Page 7 Copyright © 2020 Huawei Technologies Co., Ltd.

• CloudEngine Switch Hardware Features

• CloudFabric Solutions Introduction

Page 8 Copyright © 2020 Huawei Technologies Co., Ltd.

Core switches Access switches

10GE + 25GE TOR switch (new model)

25GE TOR switch

10GE TOR switch

CloudEngine 16816 CloudEngine 16808 CloudEngine 16804 CloudEngine 6820-48S6CQ

Page 9 Copyright © 2020 Huawei Technologies Co., Ltd.

Core switches Access switches

CE12808S CE12804S CE6865-48S8CQ-EI CE5880-48T6Q-EI

Page 10 Copyright © 2020 Huawei Technologies Co., Ltd.

• CloudEngine Switch Hardware Features

• CloudFabric Solutions Introduction

Page 11 Copyright © 2020 Huawei Technologies Co., Ltd.

Item CloudEngine 16800

Large Capacity 19.2 Tbit/s per slot

AI Engine Embedded AI chip

Smooth Evolution 400GE smooth evolution

36*100GE 36*40GE 48*10GE

Page 12 Copyright © 2020 Huawei Technologies Co., Ltd.

CE16804 CE16816 CE16808 Applicable to

Page 13 Copyright © 2020 Huawei Technologies Co., Ltd.

Traditional Switching CloudEngine 16800

Page 14 Copyright © 2020 Huawei Technologies Co., Ltd.

Airflow Direction in a DC VC Phase Change Heat Dissipation

Back Front Front Front

Rack Rack Rack

Typical hot and cold airflow in a DC Chip

Page 15 Copyright © 2020 Huawei Technologies Co., Ltd.

Server 10GE/25GE Server 25GE/100GE Server 100GE/200GE

25GE SerDes, compatibility 56GE SerDes, compatibility 112G SerDes,

Smooth evolution, supporting different scenarios

Page 16 Copyright © 2020 Huawei Technologies Co., Ltd.

36100GE 3640GE 48*10GE

4100GE CFP 3640GE QSFP+ 48*GE Base-T