The New Era of 800G Optical Transceiver

The rapid development of AI large-scale models and related applications, represented by ChatGPT, has made computing power a crucial infrastructure for the AI industry. In addition to GPU performance, communication has become one of the bottlenecks in supercomputing. Therefore, AI servers have stringent requirements for underlying data transmission rates and latency, necessitating high-speed optical modules. This article explores the emergence of 800G optical transceivers and their potential in the AI era.

The Evolution of 800G Optical Transceiver

Increasing Bandwidth Demand

In recent years, with the emergence of new businesses such as VR, IoT, and cloud computing, the market has higher requirements for network bandwidth, concurrency, and real-time performance. With the continuous increase in bandwidth demand, although 100G, 200G, and 400G optical transceivers will still have the largest market share, 800 Gbit/s optical transceivers will gradually emerge and are expected to achieve large-scale deployment by 2025, occupying the market mainstream with its performance advantage.

Potential of LPO Technology

When it comes to 800G high-speed optical transceivers, LPO technology is the most promising solution in the 800G era. lPO (Linear-drive Pluggable Optics) is a linear-drive pluggable optical module that uses only linear analog components in the data link and requires no CDR or DSP design. Compared to DSP solutions, LPOs can significantly reduce system power consumption and latency (by nearly 50% compared to DSPs) while sacrificing system error rate and transmission distance. This is highly compatible with the short-range, high-bandwidth, low-power and low-latency data connectivity requirements of today’s AI computing centers.

As a result, LPOs are expected to be the first to capture the huge market for switching networks inside data centers in the future. As North American cloud service providers actively expand their computing resources, companies such as Microsoft, Meta, AWS, and Google may gradually accept LPO solutions in the future. 800G LPO is the most promising technology route in the AI era.

Electrical and Optical Interface Architectures of 800G Ethernet

Research shows that for electrical interfaces, when the single-channel speed of the electrical interface is the same as the single-channel speed of the optical interface, the architecture of the optical module will reach its optimal state, with advantages such as low power consumption and low cost. A single-channel 100 Gbit/s electrical interface will be the ideal electrical interface for an 8×100 Gbit/s optical transceiver, and a single-channel 200 Gbit/s electrical interface will be the ideal electrical interface for a 4×200 Gbit/s optical transceiver. In terms of packaging, 800 Gbit/s optical transceivers may exist in different forms such as double-density four-channel small form-factor pluggable (800G QSFP-DD) and eight-channel small form-factor pluggable (800G OSFP).

There are three main optical interface architectures for 800 Gbit/s optical transceivers:

8×100 Gbit/s PAM4 Optical Transceiver

The PAM4 (4-level pulse amplitude modulation) transceiver operates at 53 Gbd and uses eight pairs of digital-to-analog converters (DACs) and analog-to-digital converters (ADCs), eight lasers, eight pairs of optical transceivers, and one pair of 8-channel coarse wavelength division multiplexing (CWDM) or LAN-WDM multiplexer and demultiplexer.

4×200 Gbit/s PAM4 Optical Transceiver

The PAM4 transceiver operates at 106 Gbd and uses four pairs of DACs and ADCs, four pairs of optical transceivers (including four lasers), and one pair of 4-channel CWDM or LAN-WDM multiplexer and demultiplexer.

800 Gbit/s Coherent Optical Transceiver

It operates at 128 Gbd under 16QAM modulation. It uses four pairs of DACs and ADCs, one laser, and one pair of optical transceivers and can use a fixed-wavelength laser in data center coherent optical modules to reduce cost and power consumption.

Development Trends of 800G Optical Transceiver

800G optical transceiver is currently in the first-generation stage and will continue to develop in the next five years. ZTE’s R&D team believes that there are three major trends for the future development of 800G ethernet:

Single-Mode Migration

Due to the bandwidth limitations of multimode fibers, the transmission distance of 100 Gbit/s PAM4 VCSEL+multimode fibers is limited to 50 meters. If OM5 fiber is used, system costs will increase. In the future, the sinking of single-mode optical interface solutions is an inevitable trend, which will benefit SiPh technology.

Arrival of Single-Wavelength 200 Gbit/s

Although 112 Gbd EML technology has developed rapidly and prototypes have been produced, the 55 GHz bandwidth resources are slightly insufficient. The application prospects of SiPh modulators and silicon-based thin-film lithium niobate for the 200 Gbit/s PAM4 rate level are very broad.

Coherent Migration

With the improvement of transmission rates, coherent technology solutions will further expand their applications to shorter distances such as 40, 20, and 10 km based on the 80 km transmission distance. At the same time, non-coherent solutions are also striving to expand to longer distances. The fact that the coherent solution only requires one laser, modulator, and receiver will make it cost-competitive with PAM4. PAM4 requires four simple lasers, modulators, and receivers. Even though these devices are relatively complex at 800 Gbit/s, they are enough to quickly reduce the overall system cost, keeping PAM4 competitive. The competition between coherent and PAM4 transmission has already begun, and the future result requires continuous observation.

The Impact of the AI on 800G Optical Transceiver Deployment

Examples of AI Applications

At 4:00 a.m. on May 17th, Tesla held its 2023 shareholder meeting, and the Tesla robot (Optimus) once again appeared in front of the public.
In 2021, Musk displayed the approximate form of the Tesla Bot at Tesla AI Day. However, at that time, it was still just a concept. In October 2022, the humanoid robot made its first appearance, but at that time, Optimus needed assistance from a person and could not work normally.

Today, several months later, the humanoid robot can achieve basic upright walking and can perform assembly tasks on another robot. The finger joints can meet the requirements of grasping power tools, screws, and fabric covering the frame.

Musk has high expectations for Optimus and even said at the meeting that everyone will have a humanoid robot in the future, and some people will even have more than one. He believes that this market will surpass the demand for electric cars and will reach a scale of billions. Musk openly stated that Tesla’s long-term value may come from humanoid robots, and he is very confident in this prediction.

At the ITF World 2023 Semiconductor Conference, NVIDIA CEO Huang Renxun stated that the next wave of artificial intelligence will be embodied AI, which is an intelligent system that can understand, reason, and interact with the physical world. An embodied AI robot requires understanding human language, breaking down tasks, planning subtasks, recognizing objects while moving, interacting with the environment, and ultimately completing the corresponding tasks.

Why is 800G more significant than 400G for AI servers?

Firstly, AI servers require high data transmission rates and low latency, and the corresponding top-of-rack switches need to match the underlying data transmission bandwidth, and may also require latency redundancy, requiring high-speed optical modules. For example, the NVIDIA DGX H100 server is equipped with 8* H100 GPU modules, assuming each GPU requires 2* 200G optical modules, then each server needs at least 16* 200G modules, and the corresponding top-of-rack switch ports need at least 4* 800G.

Secondly, 800G optical chips are more economical and cost-effective. 800G uses 100G EML chips, while 200G/400G uses 50G optical chips. According to calculations, for the same rate, the cost of a 100G optical chip is 30% lower than the cost of two 50G optical chips.

However, it’s worth noting that 400G optical transceivers still have their place in the industry. While they may not be as fast as 800G optical transceivers, they still offer a significant increase in bandwidth over older technologies, making them a cost-effective solution for many organizations. Additionally, some applications may not require the full capabilities of an 800G ethernet, making a 400G ethernet a more practical choice.

400G 800G Transceivers

From March 5th to March 9th, US time, the 2023 Optical Fiber Communications Conference and Exhibition (OFC) was held locally in San Diego, USA. HPC and data center high-speed optical network connectivity product and solution provider NADDOD presented its new 800G and 400G products at the 2023 OFC.

NADDOD staff met with industry experts at OFC 2023, and demonstrated live technologies of new high-speed optical transceiver products.

During the exhibition, NADDOD thought leaders participate in technical sessions and panel discussions focused on accelerating the evolution to 800G and leading the conversation on next-generation technologies for the evolving network.

NADDOD provides high-speed optical network products and turnkey HPC networking solutions for customers in different industries such as Financial, Healthcare, Education, Government and Telecom. For the up-to-date optical connectivity solutions, some 400G and 800G optical transceiver products have been launched on NADDOD official website.

Conclusion

The explosive demand for AI servers has led to the deployment of 800G optical transceivers, which offer a significant increase in bandwidth and performance over previous generations. While 400G optical transceivers still have their place in the industry, the industry is continuing to push forward with the development of even faster and more advanced technologies. LPOs, or linear-drive pluggable optics, are a cost-effective and practical 800G solution for AI servers. Despite the slowdown in capital expenditures, the demand for optical modules from cloud computing providers is expected to continue growing due to the significant gap in AI computing power. The market for 800G optical transceivers presents new opportunities as 5G, edge computing, and data center construction drive the need for higher bandwidth and faster data transmission.


Related Products:
800G OSFP-800G-DR8+
800G OSFP-800G-DR8
800G OSFP-800G-2xFR4
400G QDD-400G-SR8


Related Resources:
400G Optical Transceiver Module Types
The Next Station of Data Center Networks—200G vs 400G
Four Types of Typical 400G Network Solution Plan Explained
FAQs for 400G Transceivers and Cables