co-building an ai native base 21vianet and huawei jointly build a new generation of hyper-connected city computing power network base
2024-09-30
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
recently, 21vianet and huawei signed a strategic cooperation agreement in shanghai. through resource sharing and complementary advantages, the two parties will deepen strategic cooperation in the field of hyper-internet distributed computing power that is ai-native, chain-network integrated, and green computing power to create a new generation. ai native infrastructure accelerates the implementation of ai native applications. this cooperation will start an "intergenerational" transition from local area networks to the internet, to cloud networks, and even to the ultra-internet architecture for large-model applications.
as the world's leading ict solution provider, huawei will provide cutting-edge technological capabilities in areas such as intelligent computing wide-area backbone, distributed intelligent computing, and xsec security encryption, and will work with 21vianet on joint innovation, standards & ecological promotion, and engineering carry out in-depth cooperation in deployment and other aspects, jointly empower the "ai + digital economic complex" that serves hundreds of industries and thousands of enterprises, and jointly create a new generation of hyper-connected city distributed computing power base (computing network brain) platform.
(photo of the signing ceremony between century internet and huawei)
strong alliances to build ai native infrastructure
in september 2024, with the launch of openai’s “o1” model, the reasoning paradigm will usher in major changes. ai is accelerating towards intelligent agents, and computing shifts from pre-training to the inference stage, breaking through the bottleneck of the original ai model training scaling law to a certain extent. limited by the rising computing power holding and operating costs, ai development urgently needs innovative thinking and technology to break the situation. an effective solution is to efficiently integrate gpu, data and model resources scattered across various places, and by optimizing resource allocation and sharing, reduce the cost of using ai for enterprises and help the rapid implementation of ai native applications.
21vianet and huawei combined their own practices and summarized four major trends in enterprise ai application infrastructure: first, the need for multi-point non-blocking computing and connection architecture (east-west traffic); second, latency and bandwidth will have an important impact on overall productivity. it will have a huge impact; third, security will become more important, and the security framework will also undergo essential changes; fourth, the self-closed loop and regionalization characteristics of the business will make it no longer dependent on the large-scale internet.
under the new paradigm, traditional cloud network architecture has become "incompetent", and ai infrastructure is in urgent need of a revolutionary change. just in september this year, huawei cloud launched the ai-native cloud infrastructure architecture cloudmatrix to meet the large-scale, efficient, stable, elastic and flexible computing power needs of enterprises in the ai era. this time, huawei and 21vianet have joined forces to jointly build a distributed computing power base for hyper-connected cities to further improve the performance of ai native infrastructure.
jointly build an ai native base to empower hundreds of industries and thousands of enterprises
the hyper-connected city distributed computing power base is based on the decomposition of the native world and needs of ai, relying on 21vianet's long-term accumulation and practice in enterprise-oriented computing power, networks, platforms, services, etc., combined with huawei's leading hardware and overall solutions an integrated empowerment system built based on the plan and continuously iteratively upgraded.
1) ai-native ultra cross connects (ucc) urban computing bus architecture: build a campus all-optical network and a metro high-speed computing power bus. relying on huawei's leading dynamic slicing, encryption, and security gateway product solutions for integrated innovation, especially building a metropolitan computing power slice based on huawei's ipv6+ metropolitan computing power slicing or galaxy ai network, enabling any computing power entity within the metropolitan area , parks, buildings, enterprises, and individuals can access the bus at high speed and safely through nearby nodes. any enterprise, family, or individual can create a dynamic private network based on business needs, connect computing power, private domain data, and models to build a private computing space. as a new municipal infrastructure platform in the city, ucc will greatly reduce the marginal cost of enterprise connection and realize the intelligent upgrade of the city.
2) through the transformation of parks and buildings, and the direct connection of the operator's urban upf, 100g direct access to office desktops, 40g direct access to ai aboriginal homes, and smartphone access in 5g slicing mode are achieved. as the main place for enterprise business activities, traditional networks in campuses and buildings have many shortcomings. huawei's passive all-optical access solution is used to transform the campus all-optical minimalist network, move up the campus network control plane, and simplify daily routines. complexity of maintenance and management. at the same time, combined with the metropolitan ucc bus, the links from the campus to the computing power and the campus to 5g upf are opened to achieve high-speed and secure private domain networking.
3) one machine and two networks to maximize the use of gpu computing resources. the metropolitan ai computing power nodes are connected to the ucc bus and the internet at the same time, so that the computing power resources can not only meet the needs of private domain computing under the hyper-interconnected ucc architecture, but also can be scheduled through the neolink.ai computing power platform in leisure time to achieve sales on the internet platform make full use of gpu computing resources while ensuring security and isolation.
4) integrated management of multiple data centers in the metropolitan area and distributed computing resource pools. intelligent computing centers of different forms, such as hub, spoke, and edge, can achieve high-speed connections through the ucc computing bus. through resource pooling and scheduling, as well as long-distance rdma technology, multi-party computing resources within the metropolitan area can be efficiently integrated and utilized.
5) use edge intelligent gateways to achieve integrated innovation in protocols and security architecture on the access side. use independent public and private key addresses to achieve addressing and terminal verification, achieve "trustworthiness, manageability, and controllability" of all connections within the ucc city computing bus, build a new generation of secure native architecture, and provide b2b and b2c services consistent with traditional internet , c2c different new generation digital infrastructure.
6) one-stop model and data platform services. due to the continuous updating of basic models, the continuous import of new enterprise data and fine-tuning training, enterprise ai applications are always changing dynamically, and the model scale is increasing day by day. taking the llama3.1 405b model as an example, its complete model file has exceeded the tb level. through traditional the internet takes several days to download, but through ucc bus architecture and long-distance rdma technology, rapid updates can be achieved within minutes.
with the full arrival of the ai native era, the boundaries of ai technology continue to expand, and large models, as new technological driving forces, continue to spawn new application forms. the architecture and connotation of ai-native hyper-connected infrastructure also require gradual iteration and further innovation. 21vianet and huawei will work together to in-depth cooperation in the fields of standards, protocols, architectures, etc., build an open ecosystem and community, and jointly promote the development of ai native applications and ecosystems. in addition, the two parties will also jointly build a standardized empowerment system. through one-stop empowerment services such as technology, solutions, engineering, maintenance, and talent training, city partners can quickly build the city's ai native infrastructure and jointly open up ai. endless possibilities with native apps.
source: china daily.com
