800G Optical Module Deployment

In recent years, driven by the growing demand for artificial intelligence (AI) computing, the demand for GPUs and other computing hardware has skyrocketed dramatically. This is evident in the fact that NVIDIA's stock price has risen by more than 200% just this year. Additionally, due to the high-speed data transmission required for AI training in data centers, the demand for AI server clusters has also surged sharply. As data center infrastructure continues to expand, the use of optical modules has grown exponentially. Currently, 200G and 400G optical modules have been deployed on a large scale, and 800G optical modules have begun to enter the mass production and introduction phase.

Why are 800G optical modules needed?

Data transmission within data centers is mainly divided into two directions: north-south traffic and east-west traffic. North-south traffic refers to data transmission between external users and internal servers outside the data center, while east-west traffic refers to data transmission between data centers and between servers within the data center. With the AI large model market ignited by ChatGPT, the proportion of east-west traffic and internal server traffic in data centers has increased.

Updating Data Center Topology

As the scale of data center computing and east-west traffic continues to expand, the network architecture of data centers is also constantly changing. In the traditional three-tier topology, data exchange between servers needs to go through access switches, aggregation switches, and core switches. This puts a huge workload on the aggregation switches and core switches.

Three-tier topology: If the traditional three-tier topology is adopted, further expanding the scale of the server cluster will require deploying high-performance equipment at the core layer and aggregation layer, which will significantly increase equipment costs.

Two-tier Spine-Leaf topology: The Leaf-Spine topology flattens the traditional three-tier architecture into a two-tier architecture. In this setup, the leaf switch (similar to the access layer switch in the traditional three-tier architecture) is directly connected to the servers. The spine switch, equivalent to the core layer switch, is directly connected to the leaf switches, with each spine switch connected to all leaf switches.

AI Computing and 800G Optical Modules

Compared with the traditional three-tier topology, the spine-leaf structure requires a large number of ports. Therefore, both servers and switches require more optical modules for fiber optic communication. For widespread AI training applications using GPUs, in NVIDIA's DGX H100 servers, which integrate 8 H100 GPUs, the demand for computing and storage networks is equivalent to about 12 800G optical modules and 18 400G optical modules. More information can be found in this article: Introduction to NVIDIA DGX H100.

This indicates that under the spine-leaf architecture, the demand for high-speed optical modules in data centers is growing exponentially. With the expansion of data center scale, especially the demand for AI large model training, and the increasing demand for higher transmission rates from GPU servers, the adoption rate of 800G optical modules is accelerating.

Accelerating the Deployment of 800G Optical Modules

Against the backdrop of progress in AI computing, the demand for 800G modules is steadily rising, and leading global manufacturers are stepping up deployment efforts.

For example, Star Optics Company has started to increase the production of 800G optical modules in the first half of 2023, and has been steadily delivering 800G products matched with GPUs to major overseas customers.

At OFC 2023, Hisense Broadband launched the industry's first 800G QSFP-DD BiDi SR4.2 optical module product and showcased a comprehensive series of 800G optical module products based on PAM4 signaling, including 800G QSFP-DD SR8, 800G OSFP SR8, 800G OSFP 2xFR4/DR8, and 800G QSFP-DD 2xFR4/DR8.

Summary

With the certainty of GPU order-driven demand, the mass shipment phase of 800G optical modules will start from the second half of this year. As an important part of AI computing power, 800G optical modules will see accelerated market growth and deployment speed driven by the continuous expansion of data center scale and the sustained growth of AI training demand.