Silicon Photonics Outlook

Latest update time：2022-03-14

Reads：

With the emergence of new data center applications such as artificial intelligence, the data interconnect market based on silicon photonics technology for data centers is also gradually heating up. March is a very busy month for silicon photonics technology. AMD/Xilinx has just announced a collaboration with Ranovus to release a system that integrates Versal ACAP and Ranovus' Odin optical communication module. Marvell also released its 400G DR4 silicon photonics platform, and GlobalFoundries also released its silicon photonics process platform Fotonix. This article will analyze the prospects of silicon photonics technology. We believe that silicon photonics technology will play an increasingly important role in data centers in the future, and its technological development will evolve towards lower cost, lower power consumption, and smaller size.

Application of Silicon Photonics Technology in Data Centers

Traditionally, optical communications are mainly used in long-distance high-speed communications. For example, the fiber-optic broadband access that everyone is familiar with is an important example of such long-distance high-speed wired communications. In data centers, the commonly used data interconnection is based on copper cable communications.

However, in recent years, with the rise of applications such as artificial intelligence, data centers are also moving towards data interconnection based on optical communications. The main reason for this is that applications such as artificial intelligence will greatly increase the demand for data interconnection bandwidth, and conventional copper cable interconnection is increasingly difficult to meet the demand. For example, in artificial intelligence applications, data centers often need to perform a large amount of distributed computing, and in distributed computing, different servers need to frequently exchange a large amount of data, and the bandwidth of data interconnection often limits the performance of the overall task. This is also a feature of the big data era, that is, in order to process massive amounts of data, the performance of data access is almost as important as the computing performance, and sometimes even more important; and as the scale of algorithms becomes larger and larger, there are more and more occasions where multiple servers are needed for distributed computing, so high-speed data interconnection between servers has become an extremely important core technology, and this has become the main reason for data centers to introduce ultra-high bandwidth data interconnection based on silicon photonics. A few years ago, when artificial intelligence was just emerging, Nvidia introduced InfiniBand based on optical interconnection in machine learning training to help high-performance training through cooperation with Mellanox, and today similar ultra-high-speed optical interconnection has almost become a standard feature of data centers.

In the future, we believe that big data and artificial intelligence will continue to be the main driving force of the data center market. Because of this, the demand for optical data interconnection in data centers is expected to increase further, which will further increase the market size of silicon photonics. As a result, we have seen that several major semiconductor industry giants are actively deploying related silicon photonics technology.

Silicon Photonics Technology and CMOS Chip Design

We believe that silicon photonics technology is currently undergoing an important technological innovation, the core technology of which is co-packaged optics (CPO) technology. Using this technology, silicon photonics modules and ultra-large-scale CMOS chips can be packaged together in a more compact form, thereby further improving optical interconnect technology in data center applications in terms of cost, power consumption and size.

CPO technology mainly refers to the integration of silicon photonic modules and CMOS chips in the form of advanced packaging (such as 2.5D or 3D packaging). Before the rise of CPO technology, the current traditional technology is to separate silicon photonic modules and CMOS chips into two separate modules, and then connect them together on the PCB board. The advantage of doing this is that the design is more modular, and the CMOS chip or silicon photonic module can be replaced separately if there is a problem, but it is disadvantageous in terms of power consumption, size and cost. For example, since the output of the silicon photonic module is ultra-high-speed data, these data will encounter large signal loss when connected to the CMOS chip using a PCB board, so a high power consumption is required to support high data rates. In addition, the cost of designing a PCB that supports ultra-high-speed signals also requires a high overhead, and in terms of size, discrete silicon photonic modules and CMOS chips are usually less integrated, which also limits the further improvement of server density in data centers.

CPO is a technology designed to solve these problems. When advanced packaging technology is used to integrate silicon photonic modules and CMOS chips into the same package, first of all, the data connection quality (signal loss) between the silicon photonic module and the CMOS chip is much better than that of the PCB board, so power consumption can be reduced; on the other hand, after large-scale mass production, advanced packaging can also bring cost improvements. Finally, after using CPO, since everything is integrated in the same package, the overall system integration is greatly improved and the size is reduced, which can also increase the popularity of silicon photonic technology in data center application scenarios.

At present, many giants with investments in the field of silicon photonics are vigorously developing CPO technology. As mentioned earlier, ultra-high-speed interconnection brought by silicon photonics technology is the core technology of distributed high-performance computing such as artificial intelligence. Therefore, companies with data centers as the main market, such as Nvidia and AMD, have a lot of layout and investment in CPO. Since Nvidia acquired Mellanox, it has further strengthened its capabilities in high-performance optical interconnection. In terms of CPO, Nvidia has also announced its current research technology, which is to use CPO technology to integrate GPU and silicon photonic chips in the same package to support 24 NVLINKs at the same time, thereby achieving 4.8Tbps ultra-high-speed interconnection.

AMD, which competes directly with Nvidia in the data center, is also making great efforts to develop CPO technology. We believe that the cooperation with Ranovus announced recently is an important investment of AMD in CPO. Ranovus is an important technology innovation company in the field of CPO. The CPO technology jointly developed by it, TE Connectivity, IBM and Senko has realized the integration of silicon photonic modules and transceiver modules in the same chip with a high degree of integration (unlike conventional CPO technology, where silicon photonic modules and transceiver modules are still two separate chips), thus achieving better integration and performance. Ranovus calls its unique technology CPO 2.0, and we also expect that AMD can further enhance its capabilities in the field of optical interconnection in the data center through cooperation with Ranovus.

In addition to Nvidia and AMD, which are mainly used in their own high-performance computing systems, important network communication providers such as Broadcom and Marvell have also invested a lot in the field of silicon photonics and CPO. Marvell has just released its 400G DR4 silicon photonic chip at the OFC conference. It has a high degree of integration, including transceiver modules and laser driver modules. The chip is integrated with Marvell's Teralynx network switch through CPO technology to achieve ultra-high data bandwidth.

Silicon Photonics Technology Platform

In terms of technology, GlobalFoundries is probably the most active among the mainstream foundries in terms of investment in silicon photonics technology. In terms of silicon photonics technology, GlobalFoundries has been actively deploying it for several years and can currently provide advanced silicon photonics process platforms, including various optical waveguides, phase shifters, polarizers, photodiodes, etc. In addition to silicon photonics technology, GlobalFoundries also provides advanced packaging options to help customers implement CPO technology. In the Fotonix platform just released by GlobalFoundries, silicon photonics and CMOS are further integrated on the same chip, thereby achieving higher performance and integration. According to information released by GlobalFoundries, its current Fotonix platform customers include Broadcom, Marvell, Nvidia, Cisco, Ranovus and other important manufacturers in the field of silicon photonics, and the future prospects are promising.

In addition to GlobalFoundries, Intel is also actively developing silicon photonics process technology. In the high-speed network switching chip market, Intel is pushing its Tofino solution, which also includes self-developed silicon photonics technology and advanced packaging technology. Intel actually has deep technical accumulation in these two fields. Six years ago, the current Intel CEO Pat Gelsinger (then vice president) was pushing silicon photonics technology and believed that silicon photonics would become a mainstream technology in the future. With Intel's accumulation in silicon photonics process and packaging technology, we believe that Intel will become a strong competitor in this field in the future.

In summary, we believe that in the future, the main technological evolution of silicon photonics will focus on higher integration and packaging with CMOS chips, and advanced silicon photonics manufacturing processes and packaging technologies will become the core technical support for the evolution of silicon photonics technology.

*Disclaimer: This article is originally written by the author. The content of the article is the author's personal opinion. Semiconductor Industry Observer reprints it only to convey a different point of view. It does not mean that Semiconductor Industry Observer agrees or supports this point of view. If you have any objections, please contact Semiconductor Industry Observer.

Today is the 2979th content shared by "Semiconductor Industry Observer" for you, welcome to follow.

Latest articles about

■TSMC's 2nm is too powerful, UMC is too miserable

■The United States has heavily funded this semiconductor technology

■Tesla is also snapping up HBM 4

■Intel's next-generation AI chip is exposed for the first time

■Nvidia releases its largest chip yet

■The Danish robot giant invites you to "do something" together

■Self-developed DPU released: Microsoft chip, full of firepower

■In the post-Moore era, optical computing chips have become the key to breakthrough, and domestic manufacturers have great potential!

■Open source software in crisis

■The global semiconductor equipment giants are all in trouble