WD25000AI

WD25000AI SERIES DISTRIBUTED DATA CENTER FABRIC SWITCHES

The WatchDog WD25000AI Series is a next-generation, high-performance AI data center Ethernet switch family designed to meet the demanding requirements of AI, HPC, cloud computing, and hyperscale data centers. Built on an advanced Distributed Data Center (DDC) architecture, the WD25000AI Series delivers high-density 800GE and 400GE Ethernet connectivity, enabling scalable, low-latency, and non-blocking network infrastructure.

WD9800-48CD8D front panel

WD25000AI-128EH

  • NCFN Distributed NCF Ethernet Switch Router with 128 × 800GE OSFP Fabric (SFI) Ports
WD9800-48CD8D rear panel

WD25000AI-36FH20EH

  • NCPN Distributed NCP Ethernet Switch Router with 36 × 400GE QSFP112 Service Ports + 20 × 800GE OSFP Fabric (SFI) Ports
WD9800-40B front panel

WD25000AI-18EH20EH

  • NCPN Distributed NCP Ethernet Switch Router with 18 × 800GE OSFP Service Ports + 20 × 800GE OSFP Fabric (SFI) Ports

Key Features

High-Density Connectivity with Scalable Expansion

The WD25000AI switch delivers ultra-high-density 800GE and 400GE connectivity with exceptional forwarding performance. It supports network deployments of up to 9,216 × 400GE ports or 4,608 × 800GE ports, making it ideal for large-scale data centers and AI computing networks that require non-blocking, high-bandwidth server access.

Built on the innovative NCF-NCP architecture, the WD25000AI series offers outstanding scalability and flexibility for evolving data center environments. In this architecture, NCP (Network Computing Processor) units connect directly to servers and storage resources, while the NCF (Network Convergence Fabric) functions as a high-performance backbone that seamlessly interconnects all NCP units.

As computing and storage demands increase, additional NCF or NCP units can be integrated to expand network capacity in a linear and efficient manner. This modular design enables seamless scalability without requiring major architectural changes, service interruptions, or application modifications, ensuring smooth network growth alongside business requirements.

WD-DDC Innovative Network Architecture

The WD25000AI utilizes WD-DDC (WatchDog Diversified Dynamic-Connectivity) technology, an innovative network architecture designed to enhance the flexibility and scalability of data center networks. Unlike traditional centralized chassis switch designs, WD-DDC adopts a distributed and decoupled architecture. By dividing conventional large network switches into smaller, independent modular components, known as box switches, network functions can be deployed in a decentralized manner. These box switches can function as line cards or switching fabric modules and can be distributed across multiple cabinets.

This approach improves heat dissipation and power consumption management while addressing limitations related to device upgrades and space expansion. At the same time, the architecture enables better decoupling of endpoint devices, such as NICs and GPUs, while maintaining high throughput and load efficiency.

Cell Switching

In traditional Box Ethernet switch networks, traffic forwarding typically relies on routing protocols combined with ECMP (Equal-Cost Multi-Path) to achieve a degree of load-balanced traffic forwarding. However, factors such as flow quantity, flow size, and hash polarization can influence hash results. For example, when forwarding elephant flows, load imbalance may occur, which can negatively impact performance. In addition, the traditional ECMP approach is based on hop-by-hop route forwarding and cannot effectively detect the available bandwidth of forwarding paths. As a result, packet loss is more likely when congestion occurs.

Based on the WD-DDC architecture, the WD25000AI series adopts a cell-based switching mechanism. By dividing packets into uniformly sized cell units and forwarding them individually, traffic can be distributed evenly across NCPs, significantly improving forwarding efficiency. The WD-DDC architecture also supports path awareness through the VOQ (Virtual Output Queuing) mechanism. Before forwarding packets, NCPs determine whether path bandwidth is available and perform load balancing only across available paths, enabling non-blocking balanced switching. While the traditional ECMP method is susceptible to packet loss during link congestion, the WD-DDC cell-switching mechanism proactively detects available path bandwidth and enables non-blocking, balanced traffic forwarding.

Software

SOFTWARE SPECIFICATIONS

Forwarding Mode, SDN controller
  • Cell-based switching
  • WD DC Controller
Data Center Features
  • RDMA lossless network RoCEv2
  • Buffer visualization
  • Intelligent traffic model identification and dynamic adjustment of AI ECN thresholds
  • PFC
  • ECN
Programmable
  • NETCONF
  • Openconfig
  • gRPC
  • Python
  • Ansible automated configuration
Traffic Monitoring
  • sFlow
  • Telemetry: gRPC, telemetry stream,MOD
  • RoCE packets visibility
IPv4 routing
  • Static route and default route
  • ECMP route and policy-based routing (PBR)
  • BGP
IPv6 routing
  • Static route and default route
  • ECMP route and policy-based routing (PBR)
  • BGP+
QoS
  • ACL, CAR, precedence remarking, and queues
  • Various queue scheduling algorithms such as SP, WRR, WFQ, SP+WRR, SP+WFQ
  • Layer 2 to Layer 4 packet filtering
  • Traffic classification based on source MAC, destination MAC, source IP (IPv4/IPv6), destination IP (IPv4/IPv6), port, and protocol
  • Traffic shaping (TS)
  • WRED, tail drop, and other congestion avoidance mechanisms
  • Port ingress and egress packet rate limiting
Management and Maintenance
  • Zero configuration auto-config and configuration rollback
  • Command line interface (cli) configuration
  • Configuration via console, telnet, SSH, etc.
  • RMON, SNMP v1/v2c/v3
  • Network management system
  • NETCONF and Python
  • Syslog and user operation log
  • Hierarchical alarms
  • Power, fan, and temperature alarm functions
  • Jumbo frame
  • NTP (network time protocol)
  • Ping, Tracert and other debugging information output
  • Uploading and downloading files via FTP, TFTP, USB, etc.
  • XMODEM protocol load upgrade (supported under boot)
Mirroring
  • Port and flow-based mirroring
  • Local and remote port mirroring (ERSPAN and ERSPANv3)
Security
  • User tiered management and password protection
  • Prevention of DoS, ARP, ICMP and other attacks
  • SSL
ARP
  • ARP, RARP, and gratuitous ARP
  • ARP Detection function
  • ARP anti-attack
  • ARP source suppression
Hardware

Hardware Specifications

Features WD25000AI-18E20E WD25000AI-36F20E WD25000AI-128E
Switching capacity 14.4T 14.4T 102.4T
Packet forwarding rate 5.4B 5.4B 21.43 B
Interfaces 18*OSFP
800G+20*OSFP 800G
36*QSFP-DD
400G+20*OSFP 800G
128*OSFP 800G
Console port RJ45 *1 RJ45 *1 RJ45 *1
Out-of-band management port RJ45 *1 RJ45 *1 RJ45 *1
USB port Type-C *1 Type-C *1 Type-C *1
Dimensions (mm)
(H × W × D, with package)
255×590×1135 255×590×1135 465×670×980
Dimensions (H × W × D, no package) (mm) 88.1×440×760 88.1×440×760 308.4×442×640
Weight (no package) (kg) 17.2kg 17.3kg 33.7kg
Weight (with package) (kg) 21.59kg 21.69kg 65.94kg
Height (RU) 2 2 7
Power supply 2 (1+1 Redundancy), AC 2 (1+1 Redundancy), AC 6 (3+3 Redundancy), AC
Fan 4 4 12
Typical Power Consumption 660W (Ports using cables)
1192W (Ports using typical optical transceiver modules)
694W (Ports using cables)
1233W (Ports using typical optical transceiver modules)
1300W (Ports using cables)
3092W (Ports using typical optical transceiver modules)
Static Power Consumption 322W 322W 680W
Maximum Power Consumption 2135W 2120W 5640W
Typical Heat Dissipation 2055.45 BTU/h 2055.45 BTU/h 3846.12 BTU/h
Maximum Heat Dissipation 6302.15 BTU/h 6131.54 BTU/h 16296.19 BTU/h
MTBF (year) 38.382 38.209 22.038
MTTR (h) 0.5 0.5 0.5
Availability 0.9999985 0.99999849 0.99999739
DRAM Memory 32G 32G 32G
NOR Flash 32MB*2 32MB*2 32MB*2
Nand Flash 64GB 64GB 64GB
SSD 240GB M.2 SATA
(Optional)
240GB M.2 SATA
(Optional)
240GB M.2 SATA
(Optional)
Operating temperature -5°C to 55°C
Storage temperature -40°C to 80°C
Operating humidity 5% to 95%, noncondensing