With the exponential growth in AI, ML, and LLM, the surge in data traffic is pushing traditional network architectures to their limits, they need high bandwidth, high throughput, especially in large-scale AI training and real-time intelligent applications. High-performance GPUs are often deployed in large AI clusters to handle multiple complex model training tasks simultaneously, increasing traffic complexity and dynamic load on the network. This necessitates flexible and scalable network architectures to handle evolving traffic demands and ensure balanced load distribution.
To meet these requirements, NADDOD introduces the N9500 series 51.2T Ethernet/RoCE-based AI data center switches. These switches leverage high-speed 400G and 800G network technologies to significantly reduce latency and enhance data transmission efficiency. Designed for AI model training, large-scale parallel computing, and HPC, they are optimized to support AI-driven business growth. Additionally, the N9500 series is compatible with several NOS options, including Open Network Linux (ONL) — the open-source OCP reference NOS — and comes preloaded with SONiC OS.
NADDOD N9500-128QC: 128-Port 400G QSFP112 Ethernet AI Switch
The NADDOD N9500-128QC is engineered with advanced hardware architecture, featuring the Broadcom Tomahawk 5 (BCM78900) chipset, and provides 128 400GE ports with line-rate forwarding, redundant hot-swappable power supplies, and fans. It supports RALB and AILB load balancing technologies to optimize traffic bandwidth for AI applications, reducing AI training time. Typical deployments support up to 8K 400G ports in Layer 2 networking and up to 32K 400G ports in Layer 3 networking. The switch also offers one-click RoCE configuration for simplified RDMA network deployment.
NADDOD N9500-64OC: 64-Port 800G OSFP Ethernet AI Switch
The NADDOD N9500-64OC is a next-generation, high-performance, high-density box switch designed for AI, big data, HPC, and distributed storage applications. It provides up to 64 800G OSFP ports and leverages network flow control technologies such as PFC/ECN and MMU optimization to build a lossless, low-latency RDMA-based network, meeting the networking needs of AI, machine learning, HPC, and distributed storage scenarios.
Core Features of NADDOD N9500 Series 51.2T Ethernet AI Switches
Large Scale Networking
AI training requires large-scale GPU clusters and distributed parallel computing. NADDOD's N9500 series, based on a multi-rail interconnect architecture, supports flexible expansion with Broadcom Tomahawk 5 chip-based switches (64-port 800G OSFP / 128-port 400G QSFP112), scaling up to 32K GPU clusters.
- Supports clusters of up to 4,096 GPU servers (32,768 CX7 NICs) with 512 ToR, 512 Leaf, and 256 Spine switches.
- Each POD contains 64 GPU servers and up to 64 PODs, with 8 ToR switches per POD.
- 1:1 uplink-to-downlink traffic convergence ratio between ToR and Leaf layers, and Leaf and Spine layers.
- Inter-switch connections use NADDOD 800G OSFP or 400G QSFP112 DAC cables, optimizing cost and stability while reducing overall cluster power consumption.
- Switch-to-NIC connections use 800G OSFP 2xSR4 and 400G QSFP112 SR4 modules, offering flexible connectivity solutions.
RoCE Lossless Technology
RoCE (RDMA over Converged Ethernet) is an RDMA technology based on the Ethernet protocol, available in two versions: RoCEv1 and RoCEv2. RoCEv2 offers enhanced features and performance, comparable to InfiniBand in dynamic latency, which is crucial for AI applications. Its standardized, open, and cost-effective nature makes it ideal for high-bandwidth, low-latency scenarios like HPC, distributed storage, and AI. NADDOD N9500 series switches support one-click RoCE configuration, enabling efficient and lossless RDMA communication over Ethernet with low-latency transmission for AI networks.
RALB and AILB Load Balancing Technologies
In AI computing networks, a notable traffic pattern is observed: fewer flows from different sources converging on the same destination but with larger data sizes. Traditional load balancing like ECMP is less effective here due to long-tail latency caused by uneven hash distribution in large flows, impacting AI training efficiency.
To overcome this, NADDOD innovates multi-path traffic scheduling technologies, RALB (Remote Adaptive Load Balancing) and AILB (Artificial Intelligence Scene Load Balancing), providing intelligent resource allocation tailored for different environments, improving bandwidth and reducing AI training times.
- RALB dynamically senses link quality to publish symmetrical routing and execute global dynamic load balancing per packet, enabling congestion-free operation with near-maximum throughput. It adapts to both local and remote link quality changes, providing precise traffic scheduling at the TOR layer to achieve global load distribution, enhancing network bandwidth utilization.
- AILB leverages predictable traffic patterns between GPUs and the 1:1 convergence characteristic between leaf and spine switches. It automatically generates globally balanced paths for all network cards connected to the leaf switch, supporting automatic path switching during multi-task path conflicts and congestion.
Cognitive Routing
Cognitive routing is a dynamic routing technology based on AI and machine learning, designed to optimize data transmission paths through self-learning and real-time decision-making. Unlike traditional routing protocols, cognitive routing adapts to changes in network conditions, traffic patterns, link status, and user needs, adjusting routing strategies in real-time for improved network performance and smarter traffic management.
- Efficient Network Resource Utilization: Real-time monitoring and learning improve bandwidth usage, reduce congestion, and minimize resource waste.
- Adaptive Capability: Quickly adjusts to changes in complex, dynamic network environments, ensuring service quality and stability.
- Intelligent Decision-Making: Uses AI algorithms for more complex and precise routing decisions, enhancing network responsiveness and overall performance.
Built on the industry-leading Broadcom Tomahawk 5 (51.2T) chip, the NADDOD N9500 series delivers high-density 400G/800G ports with robust networking capabilities and integrated software-hardware optimization. Key benefits include high performance, efficiency, utilization, and cost-effectiveness—RoCE networks offer approximately 50%+ cost savings compared to InfiniBand at similar scales. Alongside compatible optical transceivers and cables, this series provides an integrated intelligent computing network solution, forming a stable, high-performance, and scalable foundation to meet high-performance computing needs. Contact our experts to find the best fit for your requirements.


- 1Three Core Broadcom Optical Interconnect Technologies
- 2A Comprehensive Guide to 800G Optical Transceivers
- 3Silicon Photonics for AI - Balancing Cost, Power, and Reliability
- 4NVIDIA GB300 Deep Dive: Performance Breakthroughs vs GB200, Liquid Cooling Innovations, and Copper Interconnect Advancements.
- 5Blackwell Ultra - Powering the AI Reasoning Revolution