Shenzhen New Value Capital Management Co.,Ltd.

Information Details

Analysis of Three Architecture Requirements for Optical Modules

2023/08/24

The traditional three-layer network architecture consists of access layer switch, aggregation layer switch and core layer switch, the server accesses the network through the access layer switch, and the path of data traffic in the network is usually access layer switch - aggregation layer switch - core layer switch - aggregation layer switch - access layer switch. The disadvantage is that the bandwidth is layer-by-layer convergence, the network bandwidth of the core switch is much less than the sum of all bandwidth at each access layer switch, the core layer switch becomes a performance bottleneck, the east-west traffic delay is large, and the scalability is poor.

In view of the shortcomings of the traditional three-layer network architecture, the fat tree architecture is proposed.

Fat tree architecture

At present, Fat-Tree is the industry's widely recognized technology for implementing non-blocking networks. The basic idea is to use a large number of low-performance switches to build a large-scale non-blocking network, and for any communication mode, there is always a path for their communication bandwidth to reach the bandwidth of the network card. Another benefit of Fat-Tree is that it uses the same switches for all of them, which allows us to use inexpensive switches throughout our data center network architecture. The fat tree architecture uses homogeneous switches, which can achieve no convergence in the bandwidth of the communication path, and at the same time, for the traffic inside the same pod, the communication distance can be achieved by 3 hops, improving the communication efficiency of east-west traffic.

Spine-Leaf network architecture

In order to further improve the east-west flow efficiency, a leaf ridge structure was proposed.

Spine-Leaf network architecture, also known as distributed core network, is also known as Fabric network architecture because this network architecture is derived from the Switch Fabric inside the switch, which also belongs to the CLOS network model. Spine-Leaf network architectures have proven to provide high-bandwidth, low-latency, non-blocking server-to-server connectivity. The leaf spine architecture only has two layer switches, leaf layer and spine layer, which reduces the number of communication hops and can achieve no bandwidth convergence.

Assuming that the number of ports of each switch is k, the number of switches, servers, and links of the three architectures is as follows:

Architecture	Leaf ridges	Fat tree	Traditional three-tiered
Number of switches	3k/2	5k^3/4	3k^2/4+k/2
Number of servers	k^2/2	k^3/4	k^3/4
Number of links	k^2	k^3	5k^3/8+k^2/2

It can be seen that when the number of ports is fixed, the fat tree architecture has the largest number of links, the largest network scale, and the demand for optical modules and optical engines is more than that of traditional three-layer networks, although the leaf spine architecture has fewer links than the traditional three-layer network, but because the leaf spine architecture uses homogeneous switches to form a bandwidth-free convergence network, the bandwidth of its link optical module is higher than that of the traditional three-layer network, so the required speed of the optical engine optical module is also higher.

The investment research team of Xinhenglida believes that when the needs of downstream application manufacturers reach a certain bone point, and the traditional communication model can no longer meet the needs of the company or cannot solve the needs of specific scenarios, changes must be made. And this change is only an architecture, not limited to a certain model or a certain manufacturer's switch. The huge flashpoint of the optical module comes from the development of artificial intelligence, and there are many new formats. Traditional data center, its transformation and upgrading is a gradual and continuous growth process, it is not like a sudden peak, suddenly gone. The number of applications of the leaf ridge architecture to the entire high-end optical module is dozens of times that of the traditional architecture. High rate is the inevitable trend of the future development of optical modules, with the evolution of optical modules to 400G, 800G and even 1.6T and other high speeds, the optical fiber transmission speed of Tb/s may become the bottleneck of optical communication transmission rate, and the ultra-high transmission rate of silicon photonics integration technology can break this bottleneck and realize Pb/s transmission. At the same time, due to the low price of silicon materials and the mature application in the semiconductor process, it can greatly reduce the procurement cost and integration technology difficulty of optical modules, and break through the cost limitation of traditional optical modules. Up to now, the research and development and industrialization of silicon photonics integration technology are mainly concentrated in the upstream silicon photonics chip manufacturing in the optical module industry chain, led by foreign enterprises represented by Intel and Luxtera, and the localization rate is low.

A Brief Introduction to SEEG

The future of AIGC has come, where has it come from, and where has it gone?