800G Ethernet: Advancing AI Large Models

Ethernet, a well-known local area network technology, has permeated every aspect of our daily lives. Since its inception in 1973, this technology was standardized in 1983. Over the years, Ethernet has steadily evolved, supporting higher bit rates, more nodes, and longer link distances, while maintaining backward compatibility.

 

From its early days of achieving only 10 Mbps transmission rates to now being able to achieve 800 Gbps, the remarkable development speed of Ethernet has continuously driven advancements in various application fields. Currently, 800G Ethernet is making a strong impact and is the latest hot topic following 400G Ethernet.

 

Ethernet: From 10 Mbps to 800 Gbps

Every standard upgrade of Ethernet technology represents a significant improvement in data transmission capacity. In 1985, the throughput was only 10 Mbps, which was upgraded to 100 Mbps in 1995 and further to 1 Gbps in 2000.

 

Today, Gigabit Ethernet is widely deployed, and 400G Ethernet is slowly emerging in large-scale enterprises, telecommunications companies, and other organizations that require ultra-high-speed backbone networks. On the Ethernet roadmap, there are even glimpses of 800 Gbps Ethernet and 1.6 Tbps Ethernet.

 

The initially widely used standard Ethernet with a throughput of 10 Mbps operated on thick coaxial cables using carrier-sense multiple access with collision detection. With the release of the IEEE 802.3u 100Base-T Fast Ethernet standard, the era of Fast Ethernet began. People could transition from 10 Mbit/s to 100 Mbit/s using twisted-pair Ethernet interfaces.

 

Later, Gigabit Ethernet was introduced, which uses fiber optics and twisted-pair cables as transmission media, with speeds of up to 1 billion bits per second. Since then, the development of Ethernet has continuously broken through design limitations. The 40 Gbps and 100 Gbps Ethernet standards were released in 2010, and the media systems can carry Ethernet signals at 40 Gbps and 100 Gbps on fiber optic cables and short-range coaxial cables.

 

With the introduction of 200/400G Ethernet, ultra-high bandwidth can fully meet the needs of various bandwidth-intensive applications and significantly reduce port costs.

 

Now, 800G Ethernet is accelerating its arrival. The Ethernet Technology Consortium (ETC) has developed the 800GBASE-R specification for 800G Ethernet in April 2020. The 800G Ethernet specification introduces new Media Access Control (MAC) layer and Physical Coding Sublayer (PCS), reuses two groups of existing 400G Ethernet logic from the IEEE 802.3bs standard, and makes modifications to distribute data across eight 106 Gbps physical channels.

 

Although 800G Ethernet is currently in the testing and refinement stage, there is still a long way to go before its large-scale deployment. However, many bandwidth-intensive applications have already started to approach 800G Ethernet. Especially with the exponential growth of data in end-to-end and user-facing traffic driven by new data center business flows such as AI models and machine learning, 800G Ethernet is poised to make a significant impact.

 

AI Large Models Drive the Development of 800G Ethernet

The demand for various bandwidth-intensive applications has been driving the development of higher-level Ethernet, and the recent surge in AI large models has undoubtedly sparked an early demand for 800G Ethernet in data centers. Since the resurgence of artificial intelligence with ChatGPT, leading technology companies such as Google and Baidu, both domestically and internationally, have been accelerating their deployment of AI large models. The arms race among major technology giants in the field of AI and GC has significantly driven the accelerated upgrading of computational devices in cloud data centers.

 

To cope with the explosion of data, the three core elements of advanced computational power in AI large models - computing power, storage, and network transmission - are all indispensable. Any shortcomings in any of these elements would affect the computational power of large models. As an essential optoelectronic conversion device in data center Ethernet, optical transceivers are directly matched with the skyrocketing computing power and data interaction of AI servers.

 

NADDOD predicted in 2021 that 800G Ethernet optical transceivers would dominate the market starting from the end of 2025. Now, driven by AI large models, the deployment of 800G Ethernet applications is undoubtedly advancing much earlier.

 

Companies along the industry chain have also confirmed the high market potential of 800G Ethernet. As a leading provider of comprehensive optical network solutions, NADDOD expects significant growth in Ethernet device sales for AI applications, driven by the surge in AI large models, and this growth trend will continue to accelerate.

 

800G Ethernet can expand the leaf-spine architecture to support large-scale network deployments at a lower cost. It is widely applied in AI large models, and the demand for optical transceivers indicates that the number of optical transceivers under the leaf-spine architecture will experience explosive growth.

 

Overall, 800G Ethernet brings outstanding value to AI large models in two aspects. Firstly, with the same computational power, it offers a multiplied computing efficiency. The larger bandwidth of 800G Ethernet allows for a multiplied increase in the transmission capacity of each port. Secondly, it enables greater computational power scalability, thanks to the highly interconnected leaf-spine architecture that maximizes interconnect density between servers, facilitating larger-scale computational clusters.

 

The rapid growth of the AI large model industry has greatly accelerated the implementation of 800G Ethernet. In the next two years, it is expected to achieve large-scale deployment.

 

Conclusion

The continuous growth of traffic and bandwidth in various bandwidth-intensive applications has not only generated significant demand for 800G optical transceivers but also raised higher requirements for Ethernet chips in switches and routers. The explosive popularity of AI large models further accelerates the implementation of 800G Ethernet. It is worth looking forward to the arrival of the era of 800G Ethernet and higher transmission-level Ethernet.