Accelerating Data Centers: FS Unveils Next-Gen 400G Solutions

As large-scale data centers transition to faster and more scalable infrastructures and with the rapid adoption of hyperscale cloud infrastructures and services, existing 100G networks fall short in meeting current demands. As the next-generation mainstream port technology, 400G significantly increases network bandwidth, enhances link utilization, and assists operators, OTT providers, and other clients in effectively managing unprecedented data traffic growth.

To meet the demand for higher data rates, FS has been actively developing a series of 400G products, including 400G switches, optical modules, cables, and network adapters.

FS 400G Switches

The emergence of 400G data center switches has facilitated the transition from 100G to 400G in data centers, providing flexibility for building large-scale leaf and spine designs while reducing the total number of network devices. This reduction can save costs and decrease power consumption. Whether it’s the powerful N9510-64D or the versatile N9550 series, FS 400G data center switches can deliver the performance and flexibility required for today’s data-intensive applications.

Of particular note is that, as open network switches, the N8550 and N9550 series switches can enhance flexibility by freely choosing preferred operating systems. They are designed to meet customer requirements by providing comprehensive support for L3 features, SONiC and Broadcom chips, and data center functionalities. Additionally, FS offers PicOS-based open network switch operating system solutions, which provide a more flexible, programmable, and scalable network operating system (NOS) at a lower total cost of ownership (TCO).

FS 400G Transceivers

FS offers two different types of packaging for its 400G transceivers: QSFP-DD and OSFP, developed to support 400G with performance as their hallmark. Additionally, FS provides CFP2 DCO transceivers for coherent transmission at various rates (100G/200G/400G) in DWDM applications. Moreover, FS has developed InfiniBand cables and transceivers to enhance the performance of HPC networks, meeting the requirements for high bandwidth, low latency, and highly reliable connections.

FS conducts rigorous testing on its 400G optical modules using advanced analytical equipment, including TX/RX testing, temperature measurement, rate testing, and spectrometer evaluation tests, to ensure the performance and compatibility of the optical modules.

FS 400G Cables

When planning 400G Ethernet cabling or connection schemes, it’s essential to choose devices with low insertion loss and good return loss to meet the performance requirements of high-density data center links. FS offers various wiring options, including DAC/AOC cables and breakout cables. FS DAC/AOC breakout cables provide three connection types to meet high-density requirements for standard and combination connector configurations: 4x100G, 2x200G, and 8x50G. Their low insertion loss and ultra-low crosstalk effectively enhance transmission performance, while their high bend flexibility offers cost-effective solutions for short links.

FS 400G Network Adapters

FS 400G network adapters utilize the industry-leading ConnectX-7 series cards. The ConnectX-7 VPI card offers a 400Gb/s port for InfiniBand, ultra-low latency, and delivers between 330 to 3.7 billion messages per second, enabling top performance and flexibility to meet the growing demands of data center applications. In addition to all existing innovative features from previous versions, the ConnectX-7 card also provides numerous enhanced functionalities to further boost performance and scalability.

FS 400G Networking Soluitons

To maximize the utilization of the 400G product series, FS offers comprehensive 400G network solutions, such as solutions tailored for upgrading from 100G to high-density 400G data centers. These solutions provide diverse and adaptable networking options customized for cloud data centers and AI applications. They are designed to tackle the continuous increase in data center traffic and the growing need for high-bandwidth solutions in extensive 400G data center networks.

For more information about FS 400G products, please read FS 400G Product Family Introduction.

How FS Can Help

Register for an FS account now, choose from our range of 400G products and solutions tailored to your needs, and effortlessly upgrade your network.

Exploring FS 100G EDR InfiniBand Solutions: Powering HPC and AI

In the realm of high-speed processing and complex workloads, InfiniBand is pivotal for HPC, AI, and hyperscale clouds. This article explores FS’s 100G EDR InfiniBand solution, emphasizing the deployment of QSFP28 EDR transceivers and cables to boost network performance.

What are the InfiniBand HDR 100G Cables and Transceivers

InfiniBand EDR 100G Active AOC Cables

The NVIDIA InfiniBand MFA1A00-E001, an active optical cable based on Class 1 FDA Laser, is designed for InfiniBand 100Gb/s EDR systems. With lengths ranging from 1m to 100m, these cables offer predictable latency, consuming a max of 3.5W, and enhancing airflow in high-speed HPC environments.

InfiniBand EDR 100G Passive Copper Cables

The NVIDIA InfiniBand MCP1600-E001E30 is available in lengths of 0.5m to 3m. With four high-speed copper pairs supporting up to 25Gb/s, it offers efficient short-haul connectivity. Featuring EEPROM on each QSFP28 port, it enhances host system communication, enabling higher port bandwidth, density, and configurability while reducing power demand in data centers.

InfiniBand EDR 100G Optical Modules

The 100Gb EDR optical modules, packaged in QSFP28 form factor with LC duplex or MTP/MPO-12 connectors, are suitable for both EDR InfiniBand and 100G Ethernet. They can be categorized into QSFP28 SR4, QSEP28 PSM4, QSFP28 CWDM4, and QSFP28 LR4 based on transmission distance requirements.

100Gb InfiniBand EDR System Scenario Applications

InfiniBand has gained widespread adoption in data centers, artificial intelligence, and other domains, primarily employing the spine-leaf architecture. In data centers, transceivers and cables play a pivotal role in two key scenarios: Data Center to User and Data Center Interconnects.

For more on application scenarios, please read 100G InfiniBand EDR Solution.

Conclusion

Amidst the evolving landscape of 100G InfiniBand EDR, FS’s solution emerges as mature and robust. Offering high bandwidth, low latency, and reduced power consumption, it enables higher port density and configurability at a lower cost. Tailored for large-scale data centers, HPC, AI, and future network expansion, customers can choose products based on application needs, transmission distance, and deployment. FS 100G EDR InfiniBand solution meets the escalating demands of modern computational workloads.

Navigating Optimal GPU-Module Ratios: Decoding the Future of Network Architecture

The market’s diverse methods for calculating the optical module-to-GPU ratio lead to discrepancies due to varying network structures. The precise number of optical modules required hinges on critical factors such as network card models, switch models, and the scalable unit count.

Network Card Model

The primary models are ConnectX-6 (200Gb/s, for A100) and ConnectX-7 (400Gb/s, for H100), with the upcoming ConnectX-8 800Gb/s slated for release in 2024.

Switch Model

MQM 9700 switches (64 channels of 400Gb/s) and MQM8700 switches (40 channels of 200Gb/s) are the main types, affecting optical module needs based on transmission rates.

Number of Units (Scalable Unit)

Smaller quantities use a two-tier structure, while larger quantities employ a three-tier structure, as seen in H100 and A100 SuperPODs.

  • H100 SuperPOD: Each unit consists of 32 nodes (DGX H100servers) and supports a maximum of 4 units to form a cluster, using a two-layer switching architecture.
  • A100 SuperPOD: Each unit consists of 20 nodes (DGX A100 servers) and supports a maximum of 7 units to form a cluster. If the number of units exceeds 5, a three-layer switching architecture is required.

Optical Module Demand Under Four Network Configurations

Projected shipments of H100 and A100 GPUs in 2023 and 2024 indicate substantial optical module demands, with a significant market expansion forecasted. The following are four application scenarios:

  • A100+ConnectX6+MQM8700 Three-layer Network: Ratio 1:6, all using 200G optical modules.
  • A100+ConnectX6+MQM9700 Two-layer Network: 1:0.75 of 800G optical modules + 1:1 of 200G optical modules.
  • H100+ConnectX7+MQM9700 Two-layer Network: 1:1.5 of 800G optical modules + 1:1 of 400G optical modules.
  • H100+ConnectX8 (yet to be released)+MQM9700 Three-layer Network: Ratio 1:6, all using 800G optical modules.

For detailed calculations regarding each scenario, you can click on this article to learn more.

Conclusion

As technology progresses, the networking industry anticipates the rise of high-speed solutions like 400G multimode optical modules. FS offers optical modules from 1G to 800G, catering to evolving network demands.

Register for an FS account, select products that suit your needs, and FS will tailor an exclusive solution for you to achieve network upgrades.

Revolutionizing Data Center Networking: From Traditional to Advanced Architectures

As businesses upgrade their data centers, they’re transitioning from traditional 2-layer network architectures to more advanced 3-layer routing frameworks. Protocols like OSPF and BGP are increasingly used to manage connectivity and maintain network reliability. However, certain applications, especially those related to virtualization, HPC, and storage, still rely on 2-layer network connectivity due to their specific requirements.

VXLAN Overlay Network Virtualization

In today’s fast-paced digital environment, applications are evolving to transcend physical hardware and networking constraints. An ideal networking solution offers scalability, seamless migration, and robust reliability within a 2-layer framework. VXLAN tunneling technology has emerged as a key enabler, constructing a virtual 2-layer network on top of the existing 3-layer infrastructure. Control plane protocols like EVPN synchronize network states and tables, fulfilling contemporary business networking requirements.

Network virtualization divides a single physical network into distinct virtual networks, optimizing resource use across data center infrastructure. VXLAN, utilizing standard overlay tunneling encapsulation, extends the control plane using the BGP protocol for better compatibility and flexibility. VXLAN provides a larger namespace for network isolation across the 3-layer network, supporting up to 16 million networks. EVPN disseminates layer 2 MAC and layer 3 IP information, enabling communication between VNIs and supporting both centralized and distributed deployment models.

For enhanced flexibility, this project utilizes a distributed gateway setup, supporting agile execution and deployment processes. Equal-Cost Multipath (ECMP) routing and other methodologies optimize resource utilization and offer protection from single node failures.

RoCE over EVPN-VXLAN

RoCE technology facilitates efficient data transfer between servers, reducing CPU overhead and network latency. Integrating RoCE with EVPN-VXLAN enables high-throughput, low-latency network transmission in high-performance data center environments, enhancing scalability. Network virtualization divides physical resources into virtual networks tailored to distinct business needs, allowing for agile resource management and rapid service deployment.

Simplified network planning, deployment, and operations are essential for managing large-scale networks efficiently. Unnumbered BGP eliminates the need for complex IP address schemes, improving efficiency and reducing operational risks. Real-time fault detection tools like WJH provide deep network insights, enabling quick resolution of network challenges.

Conclusion

Essentially, recent advancements in data center networking focus on simplifying network design, deployment, and management. Deploying technological solutions such as Unnumbered BGP eliminates the need for complex IP address schemes, reducing setup errors and boosting productivity. Tools like WJH enable immediate fault detection, providing valuable network insights and enabling quick resolution of network issues. The evolution of data center infrastructures is moving towards distributed and interconnected multi-data center configurations, requiring faster network connections and improving overall service quality for users.

For detailed information on EVPN-VXLAN and RoCE, you can read: Optimizing Data Center Networks: Harnessing the Power of EVPN-VXLAN, RoCE, and Advanced Routing Strategies.

HPC and Future Networks: Architectures, Technologies, and Innovations

High-Performance Computing (HPC) has become a crucial tool for solving complex problems and pushing the boundaries of scientific research, artificial intelligence, and various other applications. However, efficient operation of HPC systems requires specialized infrastructure and support. HPC has emerged as an indispensable tool across various domains, capable of addressing complex challenges and driving innovation in fields such as science, meteorology, finance, healthcare, and artificial intelligence.

Understanding the importance of data centers in supporting HPC is essential, as knowing the three fundamental components—compute, storage, and networking—that constitute high-performance computing systems is crucial.

Facilities in High-Performance Computing

Intensive computations in HPC environments generate substantial heat, necessitating advanced cooling solutions. Efficient cooling prevents overheating, ensuring system stability and prolonging hardware lifespan. Supporting HPC, data centers employ cutting-edge cooling facilities, including liquid cooling systems and precision air conditioning. Moreover, data center architects explore innovative cooling technologies like immersion cooling, submerging servers in special liquids for effective heat dissipation.

Success in HPC data centers relies on a range of specialized equipment tailored to meet the unique demands of high-performance computing. Key components include data center switches, server network cards, high-speed optical modules, DAC and AOC cables, and power supplies.

The Growing Demand for Network Infrastructure in High-Performance Computing

With revolutionary technologies like 5G, big data, the Internet of Things (IoT), and artificial intelligence (AI) permeating various aspects of society, the trajectory towards an intelligent, digitized society over the next two to three decades is inevitable. Data center computing power has become a powerful driving force, shifting focus from resource scale to computational scale.

To meet the ever-growing demand for computing power, high-performance computing (HPC) has become a top priority, especially as computational cluster scales expand from the petascale to the exascale. This shift imposes increasingly higher demands on interconnect network performance, marking a clear trend of deep integration between computation and networking. HPC introduces different network performance requirements in three typical scenarios: loosely coupled computing scenarios, tightly coupled scenarios, and data-intensive computing scenarios.

In summary, high-performance computing (HPC) imposes stringent requirements on network throughput and latency. To meet these demands, the industry widely adopts Remote Direct Memory Access (RDMA) as an alternative to the TCP protocol to reduce latency and maximize CPU utilization on servers. Despite its advantages, the sensitivity of RDMA to network packet loss highlights the importance of lossless networks.

The Evolution of High-Performance Computing Networks

Traditional data center networks have historically adopted a multi-hop symmetric architecture based on Ethernet technology, relying on the TCP/IP protocol stack for transmission. However, despite over 30 years of development, Remote Direct Memory Access (RDMA) technology has gradually replaced TCP/IP, becoming the preferred protocol for HPC networks. Additionally, the choice of RDMA network layer protocols has evolved from expensive lossless networks based on the InfiniBand (IB) protocol to intelligent lossless networks based on Ethernet.

From TCP to RDMA

In traditional data centers, Ethernet technology and the TCP/IP protocol stack have been the norm for building multi-hop symmetric network architectures. However, due to two main limitations—latency issues and CPU utilization—the TCP/IP network is no longer sufficient to meet the demands of high-performance computing. To address these challenges, RDMA functionality has been introduced at the server side. RDMA is a direct memory access technology that enables data transfer directly between computer memories without involving the operating system, thus bypassing time-consuming processor operations. This approach achieves high bandwidth, low latency, and low resource utilization.

From IB to RoCE

RDMA enables direct data read and write between applications and network cards. RDMA’s zero-copy mechanism allows the receiving end to read data directly from the sending end’s memory, significantly reducing CPU burden and improving CPU efficiency. Currently, there are three choices for RDMA network layer protocols: InfiniBand, iWARP (Internet Wide Area RDMA Protocol), and RoCE (RDMA over Converged Ethernet). Although RoCE offers many advantages, its sensitivity to packet loss requires support from lossless Ethernet. This evolution of HPC networks reflects a continuous pursuit of enhanced performance, efficiency, and interoperability.

Enterprise Innovative Solution: Designing High-Performance Data Center Networks

The architecture of data center networks has evolved from the traditional core-aggregation-access model to the modern Spine-Leaf design. This approach fully utilizes network interconnection bandwidth, reduces multi-layer convergence rates, and is easy to scale. When traffic bottlenecks occur, horizontal expansion can be achieved by increasing uplink links and reducing convergence ratios, minimizing the impact on bandwidth expansion. Overlay networks utilize EVPN-VXLAN technology to achieve flexible network deployment and resource allocation.

This solution draws on the design experience of internet data center networks, adopting the Spine-Leaf architecture and EVPN-VXLAN technology to provide a versatile and scalable network infrastructure for upper-layer services. Production and office networks are isolated by domain firewalls and connected to office buildings, labs, and regional center exits. The core switches of the production network provide up to 1.6Tb/s of inter-POD communication bandwidth and 160G of high-speed network egress capacity, with each POD’s internal horizontal network capacity reaching 24Tb, ensuring minimal packet loss. The building wiring is planned based on the Spine-Leaf architecture, with each POD’s switches interconnected using 100G links and deployed in TOR mode. The overall network structure is more streamlined, improving cable deployment and management efficiency.

Future-Oriented Equipment Selection

When envisioning and building data center networks, careful consideration of technological advancements, industry trends, and operational costs over the next five years is crucial. The choice of network switches plays a vital role in the overall design of data center networks. Traditional large-scale network designs often opt for chassis-based equipment to enhance the overall capacity of the network system, but scalability is limited.

Therefore, for the network equipment selection of this project, NVIDIA strongly advocates for adopting a modular switch network architecture. This strategic approach facilitates rapid familiarization by maintenance teams. Additionally, it provides operational flexibility for future network architecture adjustments, equipment reuse, and maintenance replacements.

In response to the ongoing trend of business transformation and the surge in demand for big data, most data center network designs adopt the mature Spine-Leaf architecture, coupled with EVPN-VXLAN technology to achieve efficient network virtualization. This architectural approach ensures convenient high-bandwidth, low-latency network traffic, laying the foundation for scalability and flexibility.

How FS Can Help

FS is a professional provider of communication and high-speed network system solutions for network, data center, and telecommunications customers. Leveraging NVIDIA® InfiniBand switches, 100G/200G/400G/800G InfiniBand transceivers, and NVIDIA® InfiniBand adapters, FS offers customers a comprehensive set of solutions based on InfiniBand and lossless Ethernet (RoCE). These solutions meet diverse application requirements, enabling users to accelerate their businesses and enhance performance. For more information, please visit FS.COM.

The Future Network: Unlocking the Potential of Training Super-Large-Scale AI Models

From Transformers to the widespread adoption of ChatGPT in 2023, people have come to realize that increasing model parameters can enhance performance, aligning with the scaling law of parameters and performance. Particularly, when the parameter scale exceeds trillions, the language comprehension, logical reasoning, and problem-solving capabilities of large AI models improve rapidly.

To meet the demands of efficient distributed computing in large-scale training clusters, the training process of AI models typically involves various parallel computing modes such as data parallelism, pipeline parallelism, and tensor parallelism. In these parallel modes, collective communication operations among multiple computing devices become crucial. Therefore, designing efficient cluster networking schemes in large-scale training clusters of AI models is key to reducing communication overhead and improving the effective computation-to-communication time ratio of GPUs.

Challenges in Scaling GPU Networks for Efficient Training of Ultra-Large AI Models

The computing demands of artificial intelligence applications are experiencing exponential growth, with model sizes continuously expanding, necessitating significant computational power and high memory requirements. Appropriate parallelization methods such as data, pipeline, and tensor parallelism have become key to improving training efficiency. Training extra-large models requires clusters containing thousands of GPUs, utilizing high-performance GPUs and RDMA protocols to achieve throughputs of 100 to 400 Gbps. Specifically, achieving high-performance interconnection among thousands of GPUs poses several challenges in terms of network scalability:

  • Challenges encountered in large-scale RDMA networks, such as head-of-line blocking and PFC deadlock storms.
  • Network performance optimization, including more effective congestion control and load balancing techniques.
  • Issues with NIC connectivity, as individual hosts are subject to hardware performance limitations. Addressing how to establish thousands of RDMA QP connections.
  • Selection of network topology, considering whether to adopt traditional Fat Tree structures or reference high-performance computing network topologies like Torus or Dragonfly.

Optimizing GPU Communication for Efficient AI Model Training Across Machines

In AI large-scale model training, GPU communication within and across machines generates significant data. With billions of model parameters, communication from parallelism can reach hundreds of GB. Efficient completion relies on GPU communication bandwidth within machines. GPUs should support high-speed protocols to reduce CPU memory copying. PCIe bus bandwidth determines if network card bandwidth is fully utilized. For example, with PCIe 3.0 (16 lanes = 16GB/s), if inter-machine communication has 200Gbps bandwidth, network performance may not be fully utilized.

Crucial Factors in AI Large-Scale Model Training Efficiency

In data communication, network latency comprises two components: static latency and dynamic latency. Static latency includes data serialization, device forwarding, and electro-optical transmission delays, determined by the forwarding chip’s capacity and transmission distance, representing a constant value when network topology and data volume are fixed. In contrast, dynamic latency significantly affects network performance, including queuing delays within switches and delays caused by packet loss and retransmission typically due to network congestion. Besides latency, network fluctuations introduce latency jitter, affecting training efficiency.

Critical for Computational Power in Large-Scale AI Model Training

Cluster computing power is crucial for AI model training speed. Network system reliability forms the foundation of cluster stability. Network failures disrupt computing node connections, impairing overall computing capability. Performance fluctuations may decrease resource utilization. Fault-tolerant replacement or elastic expansion may be necessary to address failed nodes during training tasks. Additionally, unexpected network failures can lead to communication library timeouts, severely impacting efficiency. Therefore, obtaining detailed throughput, packet loss, and other information is vital for fault detection.

The Role of Automated Deployment and Fault Detection in Large-Scale AI Clusters

The establishment of intelligent lossless networks often relies on RDMA protocols and congestion control mechanisms, accompanied by a variety of complex configurations. Any misconfiguration of these parameters can potentially impact network performance and lead to unforeseen issues. Therefore, efficient and automated deployment can effectively enhance the reliability and efficiency of large-scale model cluster systems.

Similarly, in complex architectural and configuration scenarios, timely and accurate fault localization during business operations is crucial for ensuring overall business efficiency. Automated fault detection aids in quickly identifying issues, notifying management accurately, and reducing costs associated with issue identification. It can swiftly identify root causes and provide corresponding solutions.

Large-scale AI models have specific requirements in terms of scale, bandwidth, stability, latency/jitter, and automation capabilities. However, there still exists a technological gap in current data center network configurations to fully meet these requirements.

Al Intelligent Computing Center Network Architecture Design Practice

Traditional cloud data center networks prioritize north-south traffic, leading to congestion, high latency, and bandwidth constraints for east-west traffic. For intelligent computing scenarios, it’s recommended to build dedicated high-performance networks to accommodate workloads, meeting high-bandwidth, low-latency, and lossless requirements.

Based on current mature commercial switches, it is recommended to consider different models of InfiniBand/RoCE switches and the supported GPU scale to set the following specifications for physical network architecture:

Standard: Based on InfiniBand HDR switches, a dual-layer Fat-Tree network architecture supports up to 800 GPU cards per cluster.

Large-scale: Based on 128-port 100G Ethernet switches, a RoCE dual-layer Fat-Tree network architecture supports up to 8192 GPU cards per cluster.

Extra-large: Based on InfiniBand HDR switches, an InfiniBand three-layer Fat-Tree network architecture supports up to 16000 GPU cards per cluster.

Extra-extra-large: Based on InfiniBand Quantum-2 switches or equivalent Ethernet data center switches, adopting a three-layer Fat-Tree network architecture supports up to 100000 GPU cards per cluster.

In addition, high-speed network connections are crucial for ensuring efficient data transmission and processing.

How FS Can Help

FS provides high-quality connectivity products to meet the demands of AI model network deployment. The FS product portfolio includes (200G, 400G) InfiniBand switches, data center switches (10G, 40G, 100G, 400G) network cards, and (10/25G, 40G, 50/56G, 100G) optical modules, accelerating AI model training and inference processes. Optical modules offer high bandwidth, low latency, and low error rates, enhancing data center network capabilities for faster and more efficient AI computing. For more information, please visit the FS website.

Enhancing Data Center Networks with InfiniBand Solutions

With the rapid growth of data centers driven by expansive models, cloud computing, and big data analytics, there is an increasing demand for high-speed data transfer and low-latency communication. In this complex network ecosystem, InfiniBand (IB) technology has become a market leader, playing a vital role in addressing the challenges posed by the training and deployment of expansive models. Constructing high-speed networks within data centers requires essential components such as high-rate network cards, optical modules, switches, and advanced network interconnect technologies.

NVIDIA Quantum™-2 InfiniBand Switch

When selecting switches, NVIDIA’s QM9700 and QM9790 series stand out as the most advanced devices. Built on NVIDIA Quantum-2 architecture, they offer 64 NDR 400Gb/s InfiniBand ports within a standard 1U chassis. This breakthrough translates to an individual switch providing a total bidirectional bandwidth of 51.2 terabits per second (Tb/s), along with an unprecedented handling capacity exceeding 66.5 billion packets per second (BPPS).

The NVIDIA Quantum-2 InfiniBand switches extend beyond their NDR high-speed data transfer capabilities, incorporating extensive throughput, on-chip compute processing, advanced intelligent acceleration features, adaptability, and sturdy construction. These attributes establish them as the quintessential selections for sectors involving high-performance computing (HPC), artificial intelligence, and expansive cloud-based infrastructures. Additionally, the integration of NDR switches helps minimize overall expenses and complexity, propelling the progression and evolution of data center network technologies.

It can be said that NVIDIA Quantum-2 InfiniBand switches not only feature high-speed NDR data transfer capabilities but also integrate extensive throughput, on-chip compute processing, advanced intelligent acceleration features, and robust structure. These attributes make them a typical choice in the realm of High-Performance Computing (HPC), Artificial Intelligence, and a wide range of cloud-based infrastructure applications. Moreover, the integration of NDR switches helps minimize overall cost and complexity, thereby promoting the development of data center network technology.

Also Check- Revolutionizing Data Center Networks: 800G Optical Modules and NDR Switches | FS Community

ConnectX®-7 InfiniBand Card

The NVIDIA ConnectX®-7 InfiniBand network card (HCA) ASIC delivers a staggering data throughput of 400Gb/s, supporting 16 lanes of PCIe 5.0 or PCIe 4.0 host interface. Utilizing advanced SerDes technology with 100Gb/s per lane, the 400Gb/s InfiniBand is achieved through OSFP connectors on both the switch and HCA ports. The OSFP connector on the switch supports two 400Gb/s InfiniBand ports or 200Gb/s InfiniBand ports, while the network card HCA features one 400Gb/s InfiniBand port. The product range includes active and passive copper cables, transceivers, and MPO fiber cables. Notably, despite both using OSFP packaging, there are differences in physical dimensions, with the switch-side OSFP module equipped with heat fins for cooling.

OSFP 800G Optical Transceiver

The OSFP-800G SR8 Module is designed for use in 800Gb/s 2xNDR InfiniBand systems throughput up to 30m over OM3 or 50m over OM4 multimode fiber (MMF) using a wavelength of 850nm via dual MTP/MPO-12 connectors. The dual-port design is a key innovation that incorporates two internal transceiver engines, fully unleashing the potential of the switch. This allows the 32 physical interfaces to provide up to 64 400G NDR interfaces. This high-density and higgh-bandwidth design enables data centers to meet the growing network demands and requirements of applications such as high-performance computing artificial intelligence, and cloud infrastructure.

FS’s OSFP-800G SR8 Module offers superior performance and dependability, offering strong optical interconnection options for data centers. This module empowers data centers to harness the full performance capabilities of the QM9700/9790 series switch, supporting the transmission of data with both high bandwidth and low latency.

NDR Optical Connection Solution

Addressing the NDR optical connection challenge, the NDR switch ports utilize OSFP with eight channels per interface, each employing 100Gb/s SerDes. This allows for three mainstream connection speed options: 800G to 800G, 800G to 2X400G, and 800G to 4X200G. Additionally, each channel supports downgrade from 100Gb/s to 50Gb/s, facilitating interoperability with previous-generation HDR devices. The 400G NDR series cables and transceivers offer diverse product choices for configuring network switch and adapter systems, focusing on data center lengths of up to 500 meters to accelerate AI computing systems. The various connector types, including passive copper cables (DAC), active optical cables (AOC), and optical modules with jumpers, cater to different transmission distances and bandwidth requirements, ensuring low latency and an extremely low bit error rate for high-bandwidth AI and accelerated computing applications. Please see the article Infiniband NDR OSFP Solution for deployment details from FS community.

RoCE Technology for Data Transmission in HPC Networks

RDMA (Remote Direct Memory Access) enables direct data transfer between devices in a network, and RoCE (RDMA over Converged Ethernet) is a leading implementation of this technology. improves data transmission with high speed and low latency, making it ideal for high-performance computing and cloud environments.

Definition

As a type of RDMA, RoCE is a network protocol defined in the InfiniBand Trade Association (IBTA) standard, allowing RDMA over converged Ethernet network. Shortly, it can be regarded as the application of RDMA technology in hyper-converged data centers, cloud, storage, and virtualized environments. It possesses all the benefits of RDMA technology and the familiarity of Ethernet. If you want to understand RoCE in depth, you can read this article RDMA over Converged Ethernet Guide | FS Community.

Types

Generally, there are two RDMA over Converged Ethernet versions: RoCE v1 and RoCE v2. It depends on the network adapter or card used.

RoCE v1

Retaining the interface, transport layer, and network layer of InfiniBand (IB), the RoCE protocol substitutes the link layer and physical layer of IB with the link layer and network layer of Ethernet. In the link-layer data frame of a RoCE data packet, the Ethertype field value is specified by IEEE as 0x8915, unmistakably identifying it as a RoCE data packet. However, due to the RoCE protocol’s non-adoption of the Ethernet network layer, RoCE data packets lack an IP field. Consequently, routing at the network layer is unfeasible for RoCE data packets, restricting their transmission to routing within a Layer 2 network.

ROCE v2

Introducing substantial enhancements, the RoCE v2 protocol builds upon the RoCE protocol’s foundation. RoCEv2 transforms the InfiniBand (IB) network layer utilized by the RoCE protocol by incorporating the Ethernet network layer and a transport layer employing the UDP protocol. It harnesses the DSCP and ECN fields within the IP datagram of the Ethernet network layer for implementing congestion control. This enables RoCE v2 protocol packets to undergo routing, ensuring superior scalability. As RoCEv2 fully supersedes the original RoCE protocol, references to the RoCE protocol generally denote the RoCE v2 protocol, unless explicitly specified as the first generation of RoCE.

Also Check- An In-Depth Guide to RoCE v2 Network | FS Community

InfiniBand vs. RoCE

In comparison to InfiniBand, RoCE presents the advantages of increased versatility and relatively lower costs. It not only serves to construct high-performance RDMA networks but also finds utility in traditional Ethernet networks. However, configuring parameters such as Headroom, PFC (Priority-based Flow Control), and ECN (Explicit Congestion Notification) on switches can pose complexity. In extensive deployments, especially those featuring numerous network cards, the overall throughput performance of RoCE networks may exhibit a slight decrease compared to InfiniBand networks.

In actual business scenarios, there are major differences between the two in terms of business performance, scale, operation and maintenance. For detailed comparison, please refer to this article InfiniBand vs. RoCE: How to choose a network for AI data center from FS community.

Benefits

RDMA over Converged Ethernet ensures low-latency and high-performance data transmission by providing direct memory access through the network interface. This technology minimizes CPU involvement, optimizing bandwidth and scalability as it enables access to remote switch or server memory without consuming CPU cycles. The zero-copy feature facilitates efficient data transfer to and from remote buffers, contributing to improved latency and throughput with RoCE. Notably, RoCE eliminates the need for new equipment or Ethernet infrastructure replacement, leading to substantial cost savings for companies dealing with massive data volumes.

How FS Can Help

In the fast-evolving landscape of AI data center networks, selecting the right solution is paramount. Drawing on a skilled technical team and vast experience in diverse application scenarios, FS utilizes RoCE to tackle the formidable challenges encountered in High-Performance Computing (HPC). FS offers a range of products, including NVIDIA® InfiniBand Switches, 100G/200G/400G/800G InfiniBand transceivers and NVIDIA® InfiniBand Adapters, establishing itself as a professional provider of communication and high-speed network system solutions for networks, data centers, and telecom clients. Take action now – register for more information and experience our products through a Free Product Trial.

Revolutionize High-Performance Computing with RDMA

To address the efficiency challenges of rapidly growing data storage and retrieval within data centers, the use of Ethernet-converged distributed storage networks is becoming increasingly popular. However, in storage networks where data flows are mainly characterized by large flows, packet loss caused by congestion will reduce data transmission efficiency and aggravate congestion. In order to solve this series of problems, RDMA technology emerged as the times require.

What is RDMA?

RDMA (Remote Direct Memory Access) is an advanced technology designed to reduce the latency associated with server-side data processing during network transfers. Allowing user-level applications to directly read from and write to remote memory without involving the CPU in multiple memory copies, RDMA bypasses the kernel and writes data directly to the network card. This achieves high throughput, ultra-low latency, and minimal CPU overhead. Presently, RDMA’s transport protocol over Ethernet is RoCEv2 (RDMA over Converged Ethernet v2). RoCEv2, a connectionless protocol based on UDP (User Datagram Protocol), is faster and consumes fewer CPU resources compared to the connection-oriented TCP (Transmission Control Protocol).

Building Lossless Network with RDMA

The RDMA networks achieve lossless transmission through the deployment of PFC and ECN functionalities. PFC technology controls RDMA-specific queue traffic on the link, applying backpressure to upstream devices during congestion at the switch’s ingress port. With ECN technology, end-to-end congestion control is achieved by marking packets during congestion at the egress port, prompting the sending end to reduce its transmission rate.

Optimal network performance is achieved by adjusting buffer thresholds for ECN and PFC, ensuring faster triggering of ECN than PFC. This allows the network to maintain full-speed data forwarding while actively reducing the server’s transmission rate to address congestion.

Accelerating Cluster Performance with GPU Direct-RDMA

The traditional TCP network heavily relies on CPU processing for packet management, often struggling to fully utilize available bandwidth. Therefore, in AI environments, RDMA has become an indispensable network transfer technology, particularly during large-scale cluster training. It surpasses high-performance network transfers in user space data stored in CPU memory and contributes to GPU transfers within GPU clusters across multiple servers. And the Direct-RDMA technology is a key component in optimizing HPC/AI performance, and NVIDIA enhances the performance of GPU clusters by supporting the function of GPU Direct-RDMA.

Streamlining RDMA Product Selection

In building high-performance RDMA networks, essential elements like RDMA adapters and powerful servers are necessary, but success also hinges on critical components such as high-speed optical modules, switches, and optical cables. As a leading provider of high-speed data transmission solutions, FS offers a diverse range of top-quality products, including high-performance switches, 200/400/800G optical modules, smart network cards, and more. These are precisely designed to meet the stringent requirements of low-latency and high-speed data transmission.

InfiniBand: Powering High-Performance Data Centers

Driven by the booming development of cloud computing and big data, InfiniBand has become a key technology and plays a vital role at the core of the data center. But what exactly is InfiniBand technology? What attributes contribute to its widespread adoption? The following guide will answer your questions.

What is InfiniBand?

InfiniBand is an open industrial standard that defines a high-speed network for interconnecting servers, storage devices, and more. It leverages point-to-point bidirectional links to enable seamless communication between processors located on different servers. It is compatible with various operating systems such as Linux, Windows, and ESXi.

InfiniBand Network Fabric

InfiniBand, built on a channel-based fabric, comprises key components like HCA (Host Channel Adapter), TCA (Target Channel Adapter), InfiniBand links (connecting channels, ranging from cables to fibers, and even on-board links), and InfiniBand switches and routers (integral for networking). Channel adapters, particularly HCA and TCA, are pivotal in forming InfiniBand channels, ensuring security and adherence to Quality of Service (QoS) levels for transmissions.

InfiniBand vs Ethernet

InfiniBand was developed to address data transmission bottlenecks in high-performance computing clusters. The primary differences with Ethernet lie in bandwidth, latency, network reliability, and more.

High Bandwidth and Low Latency

InfiniBand provides higher bandwidth and lower latency, meeting the performance demands of large-scale data transfer and real-time communication applications.

RDMA Support

InfiniBand supports Remote Direct Memory Access (RDMA), enabling direct data transfer between node memories. This reduces CPU overhead and improves transfer efficiency.

Scalability

InfiniBand Fabric allows for easy scalability by connecting a large number of nodes and supporting high-density server layouts. Additional InfiniBand switches and cables can expand network scale and bandwidth capacity.

High Reliability

InfiniBand Fabric incorporates redundant designs and fault isolation mechanisms, enhancing network availability and fault tolerance. Alternate paths maintain network connectivity in case of node or connection failures.

Conclusion

The InfiniBand network has undergone rapid iterations, progressing from SDR 10Gbps, DDR 20Gbps, QDR 40Gbps, FDR56Gbps, EDR 100Gbps, and now to HDR 200Gbps and NDR 400Gbps/800Gbps InfiniBand. For those considering the implementation of InfiniBand products in their high-performance data centers, further details are available from FS.com.