Back to

Market Insights: 800G & 1.6T Silicon Photonics Optical Modules

Jason
Data Center Architect · Aug 7, 202415880Optical Transceivers

As the demand for high-speed data transmission continues to grow, silicon photonics technology has emerged as a pivotal solution for achieving higher bandwidths and lower latency. Silicon photonics integrates optical components with electronic circuits on a single silicon chip, leveraging the scalability of semiconductor manufacturing processes. This technology has gained significant traction, especially with the advent of 800G and 1.6T optical modules, which are crucial for modern AI data centers and high-performance computing environments. In this article, we address some common questions about 800G and 1.6T silicon photonics optical modules.

 

  1. What chips are included in 800G silicon photonics modules? What is the difference between 1.6T and 800G silicon photonics optical modules?

 

The types of chips are not significantly different. Basic electronic chips in a module, such as DSPs and drivers for the transmitter, and TIAs for the receiver, are essential for 400G, 800G, or silicon/non-silicon modules. Typically, 800G silicon photonics optical modules have two silicon photonics chips on the transmitter side, each with four channels handling 400G, totaling 800G. Each silicon photonics chip includes two laser sources, while the receiver side, in addition to the TIA, has a detector chip array for each channel, along with small components like capacitors and resistors.

 

  1. Will 1.6T silicon photonics modules require 4 silicon photonics chips?

 

No. The number of optical ports is fixed. Typically, 1.6T uses two 800G chips. Some companies may use an eight-channel chip for 800G and a sixteen-channel chip for 1.6T, but for mass production, two chips are more common.

 

  1. Does a single silicon photonics chip have 2CW light sources, and two chips have 4 CW light sources? Is this the main difference between silicon photonics modules and traditional optical modules on the transmitter side?

 

For traditional 800G optical modules, typically eight EML chips are needed. Silicon photonics require fewer chips, using CW light sources instead of modulated EML sources.

 

  1. Traditional optical modules use 8 EML chips. Do they need additional modulated light sources?

 

No, EML (Electro-Absorption Modulated Laser) is a type of external cavity electro-absorption modulated laser. Essentially, the laser is divided into two parts: one part is the DFB light source, which emits constant laser light, and the other part is the electro-absorption modulator made from an electro-absorptive material that is connected and grown in front of it. This material blocks light when voltage is applied and allows light to pass through when no voltage is applied.

 

 

  1. From a hardware perspective, the main difference on the transmitter side between silicon photonics and traditional optical modules is replacing 8EML chips with 2 silicon photonics chips and four CW light sources?

 

Correct. Traditional modules use EML chips, while silicon photonics separate the electro-absorption modulator into an independent optoelectronic modulator chip, with CW light sources as an addition.

 

  1. Is there no quantitative difference between 800G and 1.6T silicon photonicsmodules, only a difference in CW light source power?

 

Correct. Additionally, silicon photonics generally use non-hermetic packaging, entirely enclosed on the PCBA. Traditional modules, like those using EML chips, often use semi-hermetic packaging, where the laser chip is enclosed in a kovar tube shell, differing significantly from silicon photonics. In traditional methods, ensuring all wavelengths are the same is relatively simple. However, for longer distances and WDM applications, traditional modules may include numerous optical lenses to combine multiple laser sources into one path, often housed in kovar shells. Thus, the packaging form varies significantly.

 

  1. Is the packaging different between 1.6T and 800G silicon photonics modules, or between silicon photonics modules and traditional modules?

 

The packaging difference is primarily between silicon photonics modules and traditional modules. Traditional modules, especially those using WDM for distances over 1-2 kilometers, combine eight laser beams into one optical fiber using space optical coupling. In contrast, silicon photonics integrates this process on the chip, eliminating the need for space optical components. This fundamental difference means that the entire packaging form of silicon photonics modules is different.

 

  1. Does the packaging difference affect the device end?

 

Traditional laser modules have eight light sources. For short-distance transmission between switches, where each laser has the same wavelength, one module may output eight fibers. In these cases, the packaging difference between silicon photonics and traditional modules is minimal. However, for certain transmission applications over 1-2 kilometers, the eight lasers have different wavelengths, spaced at 20-nanometer intervals. Ultimately, the eight channels need to be coupled or transmitted into a single fiber. Traditional modules require additional lenses and mirrors to combine the eight laser beams into one before entering the fiber. These optical components, necessary for traditional modules, are integrated on the chip in silicon photonics.

 

  1. What is the estimated market penetration and shipment volume for 800G and 1.6T sipmodules in 2024?

 

Based on Nvidia's GPU orders, the estimated demand for 800G is 7-8 million units, while 1.6T could reach over 1 million units, including AOC and DR8. The specific distribution depends on each data center's architecture.

 

  1. What is the market share of silicon photonics versus EML in these volumes?

 

InnoLight is expected to capture 50-60% of the market, with silicon photonics and EML each taking half. 1.6T silicon photonics solutions are not yet in mass production, with EML still dominating.

 

  1. Why isn't 1.6T silicon photonics in mass production yet?

 

There are two approaches to 1.6T silicon photonics: using 16 channels of 100G per channel or 8 channels of 200G per channel. The latter is challenging due to silicon's limited high-frequency response bandwidth. Some companies, like InnoLight, are exploring bandwidth redundancy techniques to achieve 200G per channel on silicon, but mass production feasibility remains uncertain. Currently, there are samples, but the feasibility of mass production is still unknown. EML at 200G per channel is considered much more achievable. The final architecture of 1.6T could use either LPO or traditional hot-pluggable modules with DSPs, but this has not been fully determined. Using LPO offers significant advantages for silicon photonics over traditional EML.    

 

  1. Is there a short-distance solution?

 

For short distances, Broadcom’s 200G per channel solutions are mature, but 1.6T AOC solutions are still in development. Nvidia is exploring alternatives beyond silicon photonics and EML.

 

 

  1. Is 200G single-channel VCSEL mature?

 

Broadcom has produced samples, but the commercialization of single-channel 200G VCSELs is still uncertain. It remains unclear whether these can be scaled to the required industrial quantities. In contrast, EML technology has already proven its capability at this level and is currently in mass production. For 1.6T AOC solutions, Nvidia is also considering thin-film lithium niobate as a potential alternative to fill this gap, as it could address the current limitations of VCSEL technology.

 

  1. How do the four 1.6T AOC solutions compare? Which will mature first?

 

LPO is architecture-dependent and not directly related to AOC or DR8. Thin-film lithium niobate faces challenges due to the industry's maturity. Because data center orders are large, an incomplete and untested supply chain makes short-term application difficult, likely taking one to two years to mature. However, silicon photonics is mainly limited by bandwidth. Currently, the primary 1.6T solution being promoted is DR8, though the cost remains high. There is hope that a 1.6T AOC solution will be introduced in the future.

 

  1. Does Nvidia prefer CPO or optical engine solutions?

 

CPO is not widely adopted due to considerations of maintainability, power consumption, and economic value. Hot-pluggable modules remain the mainstream and most economical solution. While several major switch manufacturers have promoted CPO in recent years, it has not yet reached the economic tipping point to replace hot-pluggable modules. Last year, the industry began discussing LPO as a compromise solution.

 

  1. Will LPO be widely adopted?

 

LPO is very cost-effective because it removes the DSP from the optical modules and transfers this function to the switch's DSP. This requires a deeper integration between the optical module manufacturers and the switch manufacturers, unlike the previous hot-pluggable modules that could be compatible with different switches. LPO's removal of the DSP from the module and relying on the switch's DSP for processing means that coordination between the two is crucial. This approach is attractive as it can significantly reduce the cost for switch manufacturers, but the industry chain for LPO is not yet fully mature.

 

  1. Are there new components introduced by silicon photonics modules?

 

No, the basic structure remains simple, involving PCBA, SIMT, optoelectronic chips, lenses, fiber arrays, and shells.

 

  1. Are there changes in silicon photonics module connectors?

 

No significant changes. Silicon photonics modules use similar ports and fibers, with differences mainly in packaging and manufacturing processes.

 

  1. What power levels are used for laser sources, and what is the cost increase with higher power?

 

Currently, 30-50mW is sufficient. Scaling from 800G to 1.6T only doubles the bandwidth, not the number of channels, so power requirements scale proportionally. This depends on the number of channels in the chip. For example, a four-channel laser chip needs to supply light to two channels simultaneously. If a chip has eight channels, it will require four laser sources, with the power requirements scaling proportionally to the number of channels.

 

naddod

 

NADDOD is a leading vendor in the optical connectivity industry, specializing in next-generation high-speed network solutions for AI data centers, high-performance computing, and edge computing. We offer a comprehensive range of products, including optical modules, DAC, AOC cables, 1.6T InfiniBand XDR silicon photonics transceivers and 800G/400G Ethernet silicon photonics transceivers that deliver exceptional performance and reliability. With large inventories and a commitment to high-quality standards, NADDOD ensures quick availability and high reliability for all your high-speed networking needs. Contact our experts for detailed solutions tailored to your specific requirements and experience unparalleled support for your advanced network infrastructure.

We use cookies to ensure you get the best experience on our website. Continued use of this website indicates your acceptance of our cookie policy.