AI Computing Accelerates the Deployment of 800G Optical Transceiver

In recent years, the demand for GPU and other computational hardware, driven by the growing need for AI Computing, has surged dramatically. This is evident in the impressive over 200% increase in NVIDIA's stock price this year alone. Additionally, as AI training in data centers requires high-speed data transmission, the demand for AI server clusters has also risen sharply. With the expansion of data center infrastructure, there has been an exponential growth in the usage of optical modules. At present, 200G and 400G transceivers have been deployed on a large scale, and 800G optical transceivers have begun to enter the mass production and introduction stage.

Why Do We Need 800G Optical Transceiver?

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

Updating Data Center Topology Architecture

Data center computing scale and east-west traffic continue to expand, and data center network architecture is also constantly changing. In the traditional three-layer topology, data exchange between servers needs to go through access switches, aggregation switches, and core switches. This puts huge work pressure on aggregation switches and core switches.

3-tier Topology: If the traditional three-tier topology is used to further expand the scale of server clusters, then high-performance devices need to be deployed in the core and aggregation layers, leading to a significant increase in equipment costs.

2-tier Spine-Leaf Topology: The Leaf-Spine topology flattens the traditional three-tier architecture into a two-tier architecture. In this setup, the leaf switches, akin to access layer switches in the traditional three-tier architecture, directly connect to the servers. The spine switches, comparable to core layer switches, directly connect to the leaf switches, with each spine switch connecting to all leaf switches.

AI Computing and 800G Optical Transceivers

Compared to the traditional three-layer topology architecture, the spine-leaf architecture requires a significantly larger number of ports. Consequently, both servers and switches require more optical modules for fiber optic communication. For extensive AI training applications utilizing GPUs, in NVIDIA's DGX H100 servers, which integrate 8 H100 GPUs, the demand for compute and storage networking corresponds to approximately 12 800G optical transceivers and 18 400G optical transceivers. More information can be found in this post: Introduction to NVIDIA DGX H100.

This indicates that the number of high-speed optical modules required in data centers under leaf-spine architecture increases exponentially. As data centers scale up, especially with the demand for AI large model training, and the increased need for higher transmission rates in GPU servers, the adoption pace of 800G optical modules is accelerating.

Accelerated Deployment of 800G Optical Transceivers

Against the backdrop of AI computing advancement, the demand for 800G modules is steadily rising, with leading global manufacturers intensifying their deployment efforts.

Interstellar Optics, for instance, has begun ramping up the production of its 800G optical modules in the first half of 2023, delivering steadily to key overseas clients alongside GPU-matched 800G products.

FS provides 800G optical transceivers of different transmission types, such as 800G QSFP-DD SR8, 800G OSFP SR8 and 800G OSFP 2xFR4, etc. And FS provides modules with two different protocols, InfiniBand and Ethernet. No matter what application scenario you face, our products can meet your needs. From high-speed data transmission to network expansion, FS's 800G modules will provide you with excellent performance and reliability to help you achieve the success of digital transformation.

Summary

With the certainty of GPU orders driving demand, the mass shipment phase for 800G optical transceiver is set to commence from the latter half of this year. Serving as a vital link for AI computing power, the 800G optical transceiver are poised to witness accelerated market growth and deployment rates fueled by the expanding scale of data centers and the continuous rise in demand for AI training.