12/06 2024 509
The grand blueprint of 'Digital China' is woven together by the digitization and intellectualization of various industries. Many enterprises often find it not as simple as expected when promoting technological upgrades in business scenarios, encountering issues such as expensive computing power, difficult network upgrades, slow data transmission, and inaccessible applications, causing their digital and intelligent dreams to stall.
The upgrade towards digital and intellectualization is highly dependent on the connectivity of an advanced network. To meet industrial demands, the field of data communications has adopted the 'New-Generation Internet (Net5.5G)' as the next-generation internet and its development path in recent years. However, having a direction alone is insufficient; to better guide the practical implementation of the new-generation Internet and accelerate industrial cohesion, a crucial element is needed—standards.
We understand that standardization is crucial in network construction, allowing different vendors to develop products and focus on innovation using the same set of indicators and language, thereby ensuring consistency in network performance. For example, the freezing of 4G/5G wireless communication standards significantly accelerated the network construction process. So, when will data centers, parks, and other scenarios establish their standards to drive network upgrades?
Recently, we have witnessed the dawn of standardization for the new-generation Internet in intelligent computing data centers and parks.
Not long ago, at the Second Network Innovation and Development Conference hosted by the Network Innovation and Development Alliance (NIDA), with the assistance of the Computational Networking Committee of the China Institute of Communications and the Global IPv6 Forum, and organized by the Industrial and Standard Innovation Service Center of Co-Entropy, NIDA, together with multiple industry partners, released two key network construction standards: the 'High-Quality 10Gbps Campus Network Technology Development Research Report' and the 'Technical Requirements for the Construction of Intelligent Computing Data Center Networks'. These standards provide guidance for building the next-generation Internet and promote the introduction of industry standards.
Today, let's discuss how these two network construction standards can unite the data communications industry and strengthen its intelligent development.
Why is the construction of the new-generation Internet in data centers and parks important? Imagine the journey new technologies like large models, cloud computing, AIoT, IoT, and XR take from prestigious research laboratories and vendor conferences to actual application in various industries. What network challenges must they overcome?
Firstly, AI, cloud services, autonomous driving, and other applications require immense computing resources. Data centers, large and small across the country, serve as the heart of computing power. For tens of thousands of computing cards to pump out more resources, the interconnection between cards within clusters requires network upgrades. Sometimes, a large model requires multiple thousand-card clusters to train in parallel, necessitating the reconstruction of networks between clusters.
Next, the computing power pumped out is transmitted through the wide-area network to business scenarios, and the campus network closest to the enterprise must also be upgraded simultaneously. Network upgrades in data centers and parks are led by different industry entities, leading to uneven upgrade progress and inconsistent industry understanding, which poses significant obstacles for enterprises adopting digital and intelligent technologies.
To support industrial intelligence and make enterprises faster and stronger in the intelligent era, more powerful computing cores and campus networks are needed.
However, upgrading networks faces issues such as inconsistent standards, significant conceptual disputes, and numerous solutions. The lack of standards has adverse effects on both the field of data communications and the industrial intelligent upgrade:
1. Lack of consensus. Different ICT vendors adopt their solutions, explore separately, and make trial-and-error attempts, leading to high opportunity and time costs, which slow down industrial progress.
2. Poor experience. Different definitions, concepts, and architectures in various solutions lead to incompatible product systems, affecting implementation effects and even reducing customer enthusiasm for intelligent computing and AI technologies.
3. Waste of resources. Vendors operate independently, and the lack of collaboration within the industrial chain leads to a significant waste of R&D resources, which is detrimental to the long-term development of domestic AI.
Considering these factors, only by establishing network construction standards can the data communications industry achieve maximum consensus and pave a fast track for industrial intelligence.
NIDA has integrated various forces from industry, academia, and research, recently introducing two major standards for data center networks and campus networks.
Intelligent computing data centers serve as the computational hub of the intelligent world and are moving towards clusters with over ten thousand cards. Reports indicate that xAI has already established a cluster of one hundred thousand GPU cards and plans to double its scale in the future. Meta also announced plans to purchase 350,000 N cards for cluster construction.
However, clusters with ten thousand, three hundred thousand, or even five hundred thousand cards are not simple stacks of computing cards; instead, they require tens of thousands of cards to operate efficiently like a 'supercomputer.' Interconnections between cards and between clusters require a high-performance network as a fundamental support.
How should this network be constructed? The 'Technical Requirements for the Construction of Intelligent Computing Data Center Networks' standard released this time explains multiple networking architectures.
The first is the three-tier CLOS architecture, currently the mainstream intelligent computing network architecture. Compared to the two-tier CLOS architecture, it requires more optical modules due to the added Core layer switch, resulting in suboptimal latency, construction costs, and relatively high energy consumption.
The second is the Groupwise Dragonfly+ architecture. The Groupwise Dragonfly+ direct connection architecture mentioned in the report can support clustering of over ten thousand cards, balancing networking scale and power consumption control.
The three-tier CLOS architecture is well-known, but what is the Groupwise Dragonfly+ architecture? Specifically, it flattens the architecture by interconnecting the Spine layers between PODs based on the two-tier CLOS architecture.
Firstly, compared to the two-tier CLOS architecture, it significantly increases the networking scale, supporting the construction of AI intelligent computing data center networks with clusters exceeding one hundred thousand cards in the future.
Secondly, compared to the three-tier CLOS architecture, it significantly reduces the number of devices and optical modules, lowering costs and power consumption under the same cluster scale. Taking a 128K-card cluster as an example, using the same 128 400GE port switch for networking, the Groupwise Dragonfly+ architecture requires 1024 fewer devices than the three-tier CLOS architecture, reducing the overall network energy consumption by over 20%.
Thirdly, compared to the traditional Dragonfly+ architecture, the Groupwise Dragonfly+ architecture avoids cross-group traffic routing through other group devices, simplifying routing complexity and improving system efficiency.
In the future, various networking architectures mentioned in the report are expected to be implemented in intelligent computing data centers. The construction of data center networks with over ten thousand cards can adopt not only the three-tier CLOS architecture but also the Groupwise Dragonfly+ architecture as a new option.
With the release of the report, a consensus on the networking standards for clusters with over ten thousand cards will accelerate, promoting the official introduction of industry standards and accelerating the construction of ultra-large-scale intelligent computing data centers in China, truly realizing network-enhanced computing power.
Accelerating the industrialization of intelligent computing through standardized network construction enables the 'computing core' to deliver high computing power, injecting robust momentum into industrial intelligence.
If data center networks connect the 'computing core,' campus networks connect people, terminals, and applications, serving as the 'neural hub' of the campus, ensuring smooth campus operations. It is also the network level we interact with and rely on most frequently in daily life.
The 'High-Quality 10Gbps Campus Network Technology Development Research Report' states that 'campuses cover most work and production scenarios, with over 80% of GDP and over 90% of innovations generated within them.' Therefore, industrial digital and intelligent transformation relies on campus digital and intelligent transformation, and network upgrades play a unique strategic role in campus digital and intelligent transformation. With the rise of new businesses like AI, cloud services, remote mobile offices, and high-definition XR, campus networks must accommodate an increasing number of digital and intelligent applications with larger data transmission volumes and more complex data structures, evolving from gigabit to 10-gigabit.
How should 10-gigabit campus networks be constructed? The Broadband Development Alliance, NIDA, and WAA jointly released the 'High-Quality 10Gbps Campus Network Technology Development Research Report,' providing detailed technical requirements.
Specifically, high-quality 10-gigabit campus network construction solutions should possess six capabilities: 10-gigabit ultra-broadband, deterministic reliability, experience assurance, intelligent operation and maintenance, security protection, and green and low-carbon. These indicators might seem numerous, but remember one key feature: the primary difference between the new-generation Internet campus network and traditional campus networks is the shift from 'connectivity-centric' to 'experience-centric.'
The next-generation campus network not only focuses on successful construction and connectivity but also on providing a superior user experience. From a user experience perspective, the following characteristics of 10-gigabit campus networks are crucial:
First, greater bandwidth. Campus users want to access information anytime and anywhere. Cloud-based business intelligence requires real-time transmission of vast amounts of data. High-definition video conferencing for smoother mobile offices, smooth and reliable autonomous AGV vehicles, AIoT devices, etc., all require a powerful wireless network.
Second, stronger experience assurance. The report mentions the need to support application visibility, assurance, quality inspection, location, and VIP experience assurance capabilities to enhance business and user experiences. At the conference, Zhu Wen, Deputy Director of the Information Center at Peking Union Medical College Hospital, also shared insights from the industry, stating, 'The medical industry provides 7x24-hour uninterrupted services, with hospital operations relying on information systems and networks. Ensuring extreme reliability is particularly important.' A new-generation Internet for medical campuses can provide deterministic assurance for critical network services required by patients and medical staff.
Third, stronger security protection capabilities. Changes like remote work, business cloudification, and massive wireless terminal access are gradually erasing the physical boundaries of campus networks, with devices varying in security protection capabilities. In the next-generation network era dominated by the new-generation Internet, campus networks will adopt advanced technologies including MACsec security certification, a security protection system based on the zero-trust concept, exit security protection technology, and network slicing to enhance overall security protection.
Fourth, green energy efficiency. Responding to the global carbon neutrality goal and focusing on sustainable campus development, campus networks also become greener through new device-level energy-saving technologies such as low-power devices, intelligent variable-speed fans, and automatically shut-off lasers; network-level energy-saving technologies such as network architecture optimization, on-demand resource allocation, and load balancing; and AI-enhanced system-level energy-saving technologies like intelligent energy management platforms and AI clustering intelligent energy saving.
These tangible values are being experienced by various industries and users through the new-generation Internet in campuses, and the ICT industry is well-prepared for the implementation of technologies and solutions.
In wireless technology, Wi-Fi 7 is a prominent feature of new-generation campus networks. Industry vendors have already released Wi-Fi 7 series products, offering 2-3 times the bandwidth of Wi-Fi 6, delivering the latest wireless experience in various scenarios such as indoor, outdoor, and IoT. On the application side, the adoption of wireless terminals like Wi-Fi 7 mobile phones is gradually accelerating. The 10-gigabit campus wireless Wi-Fi technology solution mentioned in the report should possess capabilities like zero-blind-spot coverage, intelligent antenna optimization, and intelligent roaming switching, aligning well with current trends.
In wired technology, the report mentions that the 10-gigabit campus wired networking solution includes classic Ethernet and Ethernet all-optical networks, with Ethernet all-optical networks suitable for new campus network scenarios. In room-intensive scenarios like education and healthcare, vendors are actively introducing high-quality Ethernet all-optical network solutions.
Taking smart classrooms as an example, products using Ethernet all-optical technology can enable 10GE all-optical access to classrooms, ensuring smooth multimedia teaching with 4K/VR without lag. In fire safety, passive solutions are also available to enhance campus safety. Combined with network management software, full-network visibility and optimization can be achieved, with AI intelligently analyzing issues, reducing the network operation and maintenance burden on schools.
It is evident from the above industrial practices that without the support of high-quality 10-gigabit networks, the digital and intelligent visions of many campuses will remain illusory.
The value of the report lies in accelerating the formation of standards through consensus building, thereby driving the construction of high-quality 10-gigabit campus networks onto a fast track.
Seizing the growth opportunities of the digital economy, accelerating the construction of 'Digital China,' and promoting digital and intelligent upgrades across industries have become urgent needs of our era.
However, the new-generation Internet cannot be built overnight and requires a step-by-step approach, paving a standardized fast track.
The release of the two reports signifies that the construction of the new-generation Internet in data centers and parks has taken three crucial steps:
The first step is consensus building. By releasing authoritative reports, the direction and goals of network evolution are clarified, not only providing theoretical guidance but also indicating a feasible technical system, offering clear guidance for industrial development and laying a solid foundation for the construction of the new-generation Internet.
The second step is industrial implementation. Only by transforming advanced technical solutions into practical industrial applications can we avoid them becoming castles in the air. Industrial implementation is a crucial aspect of building the new-generation Internet. Technology companies have made positive contributions in this regard, providing robust technical and product support for the construction of the new-generation Internet through independent R&D and innovation.
The third step is standard establishment. Unified technical standards and specifications ensure the interconnectivity between different vendors and devices, improving network compatibility and reliability, and safeguarding the long-term development of the new-generation Internet. Currently, standardization is accelerating onto a fast track.
Strengthen intelligent computing, reinforce the park, and when intelligent computing data centers and gigabit parks make a leap on a new-quality Internet, industrial digital intelligence will become higher, faster, and stronger.