The vigorous development of the artificial intelligence industry has promoted the rapid growth of computing power scale, the continuous optimization of computing power structure, and the development trend of diversified computing power has become increasingly prominent. Recently, New H3C Group, a subsidiary of Tsinghua Unigroup, released its intelligent computing network solution at the 2024 media and analyst communication conference with the theme of "× AI" (multiplying AI), which will give full play to the multiplier effect of "computing power × connection" and support the release of multiple computing power with standardized connections. H3C will build an open network that fully meets the needs of heterogeneous computing power through the optimal optimization and cooperation of computing power and connection technologies, and provides the optimal choice for computing power connections of different scales in intelligent computing centers.
Diversified computing power has become the mainstream, and the value of open networks has become prominent
The popularity of AI large models has driven a surge in the demand for various special computing power chips, the proportion of intelligent computing power has gradually increased, and the multi-heterogeneous computing system has become the mainstream model. In the actual scenario, the computing power system has formed a huge ecosystem, involving a large number of computing units within the information interaction, the network is not only the link connecting the computing power units, but also determines the efficiency and stability of computing power scheduling and data circulation. In this regard, New H3C Group believes that solving the interconnection problem between heterogeneous components such as CPUs, GPUs, network cards, and optical modules, and creating open, decoupled, and flexibly expanded network connections is the key to building a multi-integrated intelligent computing system. By decoupling the network platform from the intelligent computing platform, the advantages of various fields in the computing ecosystem can be fully leveraged, resource sharing and efficient collaboration can be realized, and customers can enjoy advanced AI intelligent computing platforms, excellent network equipment, and high-quality connection media. In addition, users can take advantage of the open standard of Ethernet to gradually build large-scale intelligent computing clusters, achieve seamless interconnection with existing facilities, and flexibly expand and upgrade according to business needs.
Intelligent Computing Network Solution Explore an open network that opens up heterogeneous computing power
In order to meet the more stringent needs of intelligent computing, New H3C Group has explored new intelligent computing network solutions, which meet the network construction needs of intelligent computing centers in different scenarios and scales with flexible and diverse networking methods and full-scenario network optimization technologies, and comprehensively enhance the carrying capacity of the network for multiple heterogeneous computing power.
- The industry's most comprehensive product layout supports full-model networking: The construction of intelligent computing networks attaches great importance to openness, deployability, and scalability, and requires diverse product forms and supports open protocols. H3C has switch products that support 200G/400G/800G port densities and rich forms, and supports a variety of flexible networking architectures such as single-frame single-layer, box-box two-layer, box-box two-layer, box-box two-layer, etc., providing an open, compatible, scalable, and stable network environment and end-to-end heterogeneous interconnection guarantee.
- Global load balancing brings extreme bandwidth utilization: Traditional load balancing technology is difficult to adapt to the requirements of high traffic congestion sensitivity, low latency, and high throughput in AIGC cluster training, which can easily lead to uneven load balancing and decreased throughput of the entire network, affecting training efficiency. H3C proposes a combination of load balancing technologies, including SprayLink device-network convergence, LBN&DLB, FGLB global load balancing, and distributed decoupled chassis DDC architecture, which can improve network bandwidth utilization to 95% and realize all-scenario intelligent computing network optimization.
- The self-healing technology of the data plane achieves microsecond-level fault convergence: Network devices are usually separated from the control plane, and when a fault occurs, the control plane refreshes the entries and recalculates the paths, and then delivers them to the data plane to achieve fault convergence. The time consumed by this processing method has a huge impact on the intelligent computing scenario. In response to the requirements of remote link load and fault detection and real-time traffic adjustment in intelligent computing scenarios, H3C has launched DPSH data plane self-healing technology, which supports rapid traffic switching after local or remote link Down, and the entire traffic switching cycle is reduced from milliseconds to microseconds, and the user side is not aware of link faults.
Born for AI computing power scenarios Computing power cluster switches improve the overall availability of intelligent computing networks
In order to further improve the overall availability of the intelligent computing network, H3C Group has simultaneously launched the H3C S12500 AI series of computing cluster core switches based on the DDC architecture (Disaggregated Distributed Chassis), aiming to provide users with a more scalable, easier O&M management, and more cost-effective distributed decoupled chassis solution.
As a product designed for AI computing power scenarios, the H3C S12500 AI series has the advantages of cell-level load balancing, native losslessness, and ultra-large-scale. It realizes GPU decoupling based on cell switching, achieves the best load balancing effect for any traffic model, ensures 100% lossless transmission, and can support up to 32K (400G) GPU cards, getting rid of the upper limit of traditional modular device port capacity. Relying on its strong ecological decoupling capabilities and excellent computing network performance, the H3C S12500 AI series can build a lossless network with zero natural packet loss, provide automatic deployment and NCF and NCP ad hoc network capabilities.
In addition, in the process of building heterogeneous computing power networks, H3C will continue to promote the standardization of internal and external GPU connections in servers, realize heterogeneous GPU intelligent computing clusters, reduce the cost of computing power deployment and application, and open up intelligent computing islands through the standardization of software ecology, so as to promote resource sharing and common prosperity of the industry.
The network is the carrier of the digital economy, computing power is the engine of the digital economy, and the mutual collaboration between the network and computing power will better promote the vigorous development of the digital economy. Facing the computing power needs and challenges in the AIGC era, New H3C Group will uphold the concept of "intensive cultivation and pragmatism, endowing wisdom to the times", and make every effort to build a high-quality intelligent computing network with ultra-high bandwidth, ultra-low latency, and ultra-high reliability, injecting strong momentum into the development of digital intelligence in all walks of life.