All categories

An In-Depth Look at InfiniBand Technology: NDR(400G)


Release time:2026-03-26


I. What is NDR?

NDR (Next Data Rate) is the seventh generation rate standard of InfiniBand technology. Its full name is "Next Generation Data Rate". It is the mainstream high-speed interconnection technology for AI and high-performance computing (HPC) at present (in 2026). 

It is the new generation of InfiniBand standard following SDR, DDR, QDR, FDR, EDR and HDR. It is specifically designed to meet the massive data transmission requirements in scenarios such as large model training and supercomputing.

   

II. Key Technical Specifications

Feature Specification
Single Port Rate 400 Gbps (4 × 100 Gbps lanes)
Sub-rate Version NDR200 (200 Gbps, 2 × 100 Gbps lanes)
Modulation Format PAM4 (4-level Pulse Amplitude Modulation)
Encoding Scheme 64b/66b
Physical Interface OSFP (Octal Small Form-factor Pluggable), supporting dual-port multiplexing
Typical Application Bandwidth A single server can achieve a total access bandwidth of 3.2 Tbps via 8 × 400G links

 

III. Key Improvements Over Previous-Generation HDR

Compared to the previous-generation HDR (200 Gbps), NDR delivers a significant performance leap:

Doubled Bandwidth: The per-port data rate has increased from 200 Gbps to 400 Gbps, directly doubling data throughput.

Increased Port Density: By adopting OSFP interfaces, each physical port can support two NDR links, tripling the switch’s port density.

Expanded System Capacity: Taking the NVIDIA Quantum-2 platform as an example, total system capacity has increased fivefold, supporting larger-scale GPU clusters.

Enhanced AI Acceleration: The performance of the switch’s built-in AI acceleration engine has increased 32-fold, optimizing collective communication operations such as NCCL.

Translated with DeepL.com (free version)

 

IV. Key Application Scenarios

NDR InfiniBand has become the “standard” technology for building modern AI computing centers and supercomputing centers, and is primarily used in the following areas:

Large-scale model training: Enables high-speed parameter synchronization and gradient aggregation across thousands of GPUs, reducing training cycles from months to weeks.

High-Performance Computing (HPC): Enables real-time processing and analysis of terabyte-scale data in fields such as weather forecasting, genetic sequencing, and materials science.

GPU Direct RDMA: Allows GPUs to directly access remote GPU memory, significantly reducing data copy overhead and improving computational efficiency.

High-Frequency Trading: Provides microsecond-level latency for order transmission and matching in financial markets.

 

V. Typical Network Design Architecture

List of typical architectures built on NDR InfiniBand:

Core Equipment: NVIDIA Quantum-2 MQM9790 switches and ConnectX-7 network interface cards that support NDR.

Link Speed:

Server Side: 400 Gbps NDR OSFP SR4 optical modules.

Spine-Leaf Interconnect: 800 Gbps OSFP 2x SR4 optical modules (aggregated from two 400 Gbps NDR links).

Design Objective: To provide an AI cluster consisting of 256 GPU servers with extreme bandwidth, ultra-low latency, and linear scalability through NDR technology.

 

VI. Future-Oriented Evolution

NDR is not the end of the road for InfiniBand technology. The next-generation XDR (800 Gbps) standard has already been widely adopted in commercial settings and will further increase per-port throughput to meet the demands of future large-scale AI models and supercomputing applications.

Submit
%{tishi_zhanwei}%

Leave Message

If you have already experienced our product, please let us know your true feelings. Your satisfaction is our driving force for progress, while your suggestions are our valuable asset for continuous improvement.

Contact Us

Contact Us

Email: alen@wisdom-c.com

中企跨境-全域组件 制作前进入CSS配置样式

在线客服添加返回顶部

右侧在线客服样式 1,2,3 1

图片alt标题设置: NexusWise Cloud Computing Co., Ltd.

表单验证提示文本: Content cannot be empty!

循环体没有内容时: Sorry,no matching items were found.

CSS / JS 文件放置地

Welcome to leave an online message, we will contact you promptly

%{tishi_zhanwei}%