Intelligent computing center photoelectric hybrid high-speed interconnection system and method

Abstract

The invention discloses an intelligent calculation center photoelectric hybrid high-speed interconnection system and method, and aims to solve the problems of construction cost and energy consumption of a high-speed interconnection network in a high-performance large-scale intelligent calculation center. The method comprises the following steps: configuring an optical switch (OCS) and an electric switch with proper scales and specifications, and connecting a network interface of a GPU to the optical switch; the method comprises the following steps of: dynamically establishing or disconnecting optical path connections between GPUs, between the GPUs and the electric switches and between the electric switches through the OCS controller by connecting ports of the electric switches to ports of the optical switches to form a photoelectric hybrid interconnection structure, so as to obtain a required photoelectric hybrid interconnection topological structure. The optical path connections between the GPUs, between the GPUs and the electric switches and between the electric switches are dynamically established or disconnected through the OCS controller. Through the photoelectric hybrid interconnection network, double flexibility of light path direct connection between the GPUs and electric exchange connection is realized, construction cost and energy consumption are effectively reduced, light path connection in the photoelectric hybrid interconnection structure can be dynamically adjusted, and network performance and resource utilization rate are optimized.

Description

Technical Field

[0001] This invention relates to the field of supercomputer interconnection network communication technology, and in particular to a high-speed optoelectronic hybrid interconnection system and method for intelligent computing centers. Background Technology

[0002] High-performance, large-scale intelligent computing centers are the infrastructure supporting the training and inference of large models. As model sizes continue to increase, intelligent computing centers not only need to be equipped with a greater number of GPUs (Graphics Processing Units), but also require the construction of correspondingly expanded high-speed interconnect networks to connect these GPU resources. This trend directly leads to a significant increase in construction and maintenance costs. Therefore, designing and implementing a more efficient and energy-saving interconnect network architecture for intelligent computing centers is particularly urgent and important for promoting the continuous advancement of artificial intelligence technology, especially generative artificial intelligence technology based on large models.

[0003] On the other hand, the network traffic generated by the large model training process in intelligent computing centers exhibits significant temporal and spatial sparsity. Traditional Spine / Leaf or fat tree structures often face high redundancy issues when building such interconnected networks, leading to high construction costs.

[0004] Therefore, exploring and adopting more efficient and energy-saving network architectures to reduce unnecessary redundancy and lower costs has become an important research direction in the current development of intelligent computing centers. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a hybrid optoelectronic high-speed interconnect system and method for intelligent computing centers, placing high-speed optical switching at the center of the high-speed interconnect network. By reconstructing the high-speed optical switching section, a flexible interconnect topology is formed, and pooled utilization of high-speed electrical switching resources is achieved.

[0006] The technical solution of the present invention is as follows:

[0007] On the one hand, the present invention provides a high-speed optoelectronic hybrid interconnection system for intelligent computing centers, characterized in that it includes:

[0008] An optical circuit switch (OCS) provides optical switching functionality, connecting the GPU and the electrical switch.

[0009] An electrical switch, used to provide electrical switching functionality, has each port connected to a port of the optical switch;

[0010] Multiple GPUs, whose network interfaces are connected to the optical switch, enable direct communication with the optical switch;

[0011] The OCS controller is used to control the optical switches to reconfigure by issuing commands, establishing or disconnecting optical path connections between GPUs, between GPUs and electrical switches, and between electrical switches, thereby obtaining the desired optoelectronic hybrid interconnect topology.

[0012] Furthermore, it also includes connectors for connecting the network interface of the GPU, the port of the electrical switch, and the port of the optical switch to form a hybrid optoelectronic interconnect structure.

[0013] Furthermore, the OCS controller also includes a traffic monitoring module, which is used to monitor the network traffic of the intelligent computing center and dynamically adjust the optical path connections in the optoelectronic hybrid interconnect structure based on the traffic monitoring results.

[0014] On the other hand, the present invention also provides a method for high-speed optoelectronic hybrid interconnection in intelligent computing centers, characterized in that the method includes the following steps:

[0015] Step S1. Configure at least one optical switch (OCS) and electrical switch, the specifications and quantity of which are adapted to the scale and requirements of the intelligent computing center.

[0016] Step S2. Connect the network interfaces of all GPUs in the intelligent computing center to the optical switch via optical fiber or optical interface card to form an optical switching layer;

[0017] Step S3. Connect each port of the electrical switch to the corresponding port of the optical switch via a cable or optical interface card to form a hybrid optoelectronic interconnection structure;

[0018] Step S4. Based on model training or requirements, configure the optical switches through the OCS controller to dynamically establish or disconnect optical path connections between GPUs, between GPUs and electrical switches, and between electrical switches, thereby obtaining the required optoelectronic hybrid interconnect topology.

[0019] Furthermore, the optoelectronic hybrid interconnect structure can support at least one of the following interconnect topologies: Spine / Leaf topology; Fat-Tree topology; Rail-Only topology; TPU v4 topology.

[0020] Furthermore, the OCS controller communicates with the optical switch via a dedicated communication link or network protocol, and issues configuration commands to control the optical switch to reconstruct its internal connections.

[0021] Furthermore, the method also includes the following steps:

[0022] Step 5. Monitor the network traffic of the intelligent computing center to identify traffic hotspots and sparse areas needed for model training;

[0023] Step 6. Based on the traffic monitoring results, dynamically adjust the optical path connections in the optoelectronic hybrid interconnect structure to optimize network performance and resource utilization.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] Configure optical switches (OCS) and electrical switches of appropriate size and specifications, connect the network interface of the GPU to the optical switch, and connect the ports of the electrical switch to the ports of the optical switch to form a hybrid optoelectronic interconnect structure. The OCS controller dynamically establishes or disconnects optical path connections between GPUs, between GPUs and electrical switches, and between electrical switches according to the needs of model training or inference, thereby obtaining the desired optoelectronic hybrid interconnect topology.

[0026] The optoelectronic hybrid interconnect network achieves dual flexibility, enabling both direct optical connections and electrically switched connections between GPUs, effectively reducing construction costs and energy consumption. Furthermore, to meet the diverse training needs of large models, the optical path connections within the optoelectronic hybrid interconnect structure can be dynamically adjusted, optimizing network performance and resource utilization.

[0027] This invention also achieves full pooling of electrical switching resources, allowing the use of switches in the electrical switching resource pool on demand, further reducing the cost and energy consumption of the electrical switching component. Attached Figure Description

[0028] Figure 1 Intelligent computing center optoelectronic hybrid high-speed interconnection network

[0029] Figure 2 Connection between GPU and optical switch

[0030] Figure 3 Connection between electrical switches and optical switches

[0031] Figure 4 Different connection methods are achieved through OCS. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the scope of protection of the present invention.

[0033] Please see Figure 1-3 , Figure 1This is a schematic diagram of the optoelectronic hybrid high-speed interconnect network of the intelligent computing center of the present invention. As shown in the figure, port 301 of the electrical switch 103 and network interface 202 of the GPU 101 are both connected to the OCS 102, forming an optoelectronic hybrid interconnect structure 104 that includes optical switching, electrical switching, and GPU. The OCS, as the core component, is located at the center of the architecture, serving as the management and coordination center to ensure efficient communication and data exchange between units, forming a highly efficient network structure, particularly suitable for high-performance computing and data center applications requiring massive parallel processing capabilities. This design not only improves data processing efficiency but also enhances the system's flexibility and scalability.

[0034] Figure 2 The diagram illustrates the connection method between the GPU and OCS. Figure 3 The diagram illustrates the connection method between the OCS and the electrical switch. For example... Figure 4 As shown, GPUs can be directly connected via connection (401) provided by OCS. GPUs can be connected to an electrical switch via connection (402) provided by OCS. Electrical switches can be connected via connection (403) provided by OCS. GPUs can be connected to an electrical switch via connections (404 and 405) provided by OCS, and the connection is established through the electrical switch. Through connections (402, 403, 404, 405) provided by OCS, GPU3 can also be connected to GPU4 and GPU5 across two electrical switches. Other topologies can be combined... Figure 4 The connection is obtained from the middle.

[0035] Figure 4 The internal OCS connections shown (i.e., 401, 402, 403, 404, 405) are established by issuing commands to control OCS102 to reconfigure via communication connection 407 between OCS controller 406 and OCS102.

[0036] Figure 4 The specific implementation of the internal connections of the OCS shown (i.e., 401, 402, 403, 404, 405) is determined by the specific OCS technology. Different OCS technologies do not affect the interconnection method of this patent.

[0037] The optoelectronic switching interconnect network in this embodiment can achieve direct optical connections between GPUs or interconnections between GPUs via electrical switching. It can implement different interconnect topologies, including Spine / Leaf, Fat-Tree, Rail-Only, and TPU v4, to meet different needs of large model training. It can achieve full pooling of electrical switching resources, thereby effectively reducing the cost of the electrical switching component. Switches in the electrical switching resource pool can be used on demand, thereby reducing the energy consumption of electrical switching.

Claims

1. A high-speed optoelectronic hybrid interconnection system for intelligent computing centers, characterized in that, include: At least one optical switch is provided to provide optical switching functionality and connect the GPU and the electrical switch; An electrical switch, used to provide electrical switching functionality, has each port connected to a port of the optical switch; Multiple GPUs, whose network interfaces are connected to the optical switch, enable direct communication with the optical switch; The OCS controller is used to control the optical switches to reconfigure by issuing commands, establishing or disconnecting optical path connections between GPUs, between GPUs and electrical switches, and between electrical switches, thereby obtaining the desired optoelectronic hybrid interconnect topology.

2. The intelligent computing center optoelectronic hybrid high-speed interconnection system according to claim 1, characterized in that, It also includes connectors for connecting the network interface of the GPU, the port of the electrical switch and the port of the optical switch to form a hybrid optoelectronic interconnect structure.

3. The intelligent computing center optoelectronic hybrid high-speed interconnection system according to claim 1, characterized in that, The OCS controller also includes a traffic monitoring module, which monitors the network traffic of the intelligent computing center and dynamically adjusts the optical path connections in the optoelectronic hybrid interconnect structure based on the traffic monitoring results.

4. A method for high-speed optoelectronic hybrid interconnection in intelligent computing centers, characterized in that, The method includes the following steps: Step S1. Configure at least one optical switch (OCS) and electrical switch, the specifications and quantity of which are adapted to the scale and needs of the intelligent computing center. Step S2. Connect the network interfaces of all GPUs in the intelligent computing center to the optical switch via optical fiber or optical interface card to form an optical switching layer; Step S3. Connect each port of the electrical switch to the corresponding port of the optical switch via a cable or optical interface card to form a hybrid optoelectronic interconnection structure; Step S4. Based on model training or requirements, configure the optical switches through the OCS controller to dynamically establish or disconnect optical path connections between GPUs, between GPUs and electrical switches, and between electrical switches, thereby obtaining the required optoelectronic hybrid interconnect topology.

5. The method for high-speed optoelectronic hybrid interconnection of intelligent computing centers according to claim 4, characterized in that, The optoelectronic hybrid interconnect structure can support at least one of the following interconnect topologies: Spine / Leaf topology; Fat-Tree topology; Rail-Only topology; TPU v4 topology.

6. The method for high-speed optoelectronic hybrid interconnection of intelligent computing centers according to claim 4, characterized in that, The OCS controller communicates with the optical switch via a dedicated communication link or network protocol, and issues configuration commands to control the optical switch to reconstruct its internal connections.

7. The method for high-speed optoelectronic hybrid interconnection of intelligent computing centers according to claim 4, characterized in that, The method also includes the following steps: Step 5. Monitor the network traffic of the intelligent computing center to identify traffic hotspots and sparse areas needed for model training; Step 6. Based on the traffic monitoring results, dynamically adjust the optical path connections in the optoelectronic hybrid interconnect structure to optimize network performance and resource utilization.

Images