OFC 2022 : Evolution from Optical Pluggable to CPO

OFC2022: Evolution from Optical Pluggable to CPO

This year OFC has a workshop. The topic is "Is Paradigm Shift from Pluggable Optics to Co-packaged Optics Inevitable in the Next Generation of Datacenters?" The title was a question, to which the answer was "Yes", a consensus, but "might not be as smooth as shown". Here we collate the views of the huge players and some relevant progress for your reference.

1. What are the driving forces for CPO?

The driving force of CPO, in fact, is to answer why CPO is this technical route. The driving force of CPO mainly comes from the following three aspects:

a) Reduce power consumption

When it comes to CPO, everyone talks about its advantages in reducing power consumption. Different reporters give different power consumption values. Broadcom gives a 30% power consumption reduction, while Intel gives a 15% power consumption reduction. The typical metric is 5pJ/bit.

b) Increase bandwidth density

With the explosive increase of data volume, copper interconnection cannot solve the problem of high-speed data transmission between board-to-board and rack-to-rack, while optical interconnection can better solve the problem.According to Broadcom's estimates, bandwidth density can be increased by 50%. In addition, CPO can also solve the density problem of the high-speed IO interface of computing chips, known as Optical IO.The following figure shows the schematic diagram of CPO in two application directions: Switch chip and XPU computing chip.

c) Reduce costs

This point is closely related to silicon optical. By taking advantage of the high integration of silicon optical devices, CPO modules can support higher bandwidth rate transmission, thus reducing the cost per unit rate. The value given by Microsoft is <<1$/Gbps. Broadcom estimates a 40 percent cost reduction from CPO technology.

2. Difficulties and challenges of CPO

Although the advantages of CPO technology are obvious, there are also various technical difficulties and challenges. People are also concerned about the following points:

a) Yield

This point is the Achiles' heel of CPO technology. Since the optical chip is directly packaged with the electrical chip, if one chip is damaged, the whole module cannot work regularly. If Pluggable encounters this problem, it can directly replace an optical module to make use of its flexible advantages. To solve this problem, it is necessary to further optimize the process of silicon optical fab to improve the reliability and yield of optical devices and systems.

b) Encapsulation

As it involves high-speed signal transmission with high bandwidth, advanced packaging technologies such as 2.5D/3D are needed, including high-density bump and TSV technologies, so that the chip per unit area can support higher rate signal transmission. The introduction of advanced packaging technology will also bring new challenges to the yield. The following diagram shows Broadcom's chip 3d packaging structure. Similar to Intel's packaging solution, the PIC is placed above the EIC.

c) Heat dissipation control

The power consumption of both switch and XPU chips is very high. It is also a difficulty to ensure that the whole system can still work normally at high temperatures. In addition, the silicon nitride waveguide can be introduced into the silicon optical chip to reduce the influence of temperature on the optical device.

3. Comparison between external laser and on-chip integrated laser

On this point, most manufacturers choose external laser, only Intel is taking the line of on-chip integration, which is directly related to Intel's silicon optical process. The following table is a comparison of the two schemes proposed by Lumetum.

The external laser solution can independently control the heat dissipation of the laser, so as to avoid the influence of the high-power electrical chip on its heat dissipation and performance. The heat dissipation of the laser itself will not affect the PIC and EIC.The optical engine's bandwidth density is not affected because the external laser does not take up additional area of the PIC.If the external laser is damaged, it can be easily replaced. But the external laser requires additional optical components and fiber coupling package, the cost will be increased. In terms of link loss, the coupling loss from the laser to the optical chip also needs to be accounted for. In addition, because of the need to support multiple subsystems at the same time, the CPO system puts forward new requirements on laser optical power.

4. Progress of various manufacturers in CPO direction

a) Marvell

Marvell showed off its first CPO prototype, which has a bandwidth of 1.6Tbps and will support its 51.2-terabyte switch chip in the future. Marvell didn't disclose any additional details, but we'll post a picture of them here.

b) Broadcom

Broadcom released its first CPO product last year, and this year's report shows some details. A single CPO module supports 3.2Tbps, and the entire system consists of four CPO modules with a total bandwidth of 12.8Tbps, as shown in the figure below.

c) Ranovus

Ranovus has released its second generation product with a bandwidth of 800Gbps and announced a partnership with Xilinx.Its CPO module is shown in the following figure. It uses GlobalFoundries' silicon optical process platform.

After a brief summary, we all agree on the prospect of CPO, which can solve the two pain points of power consumption and bandwidth density, inevitable (inevitable) development direction. However, CPO is still in a relatively preliminary stage of development, and there are numerous technical difficulties to be solved, as well as deep cooperation with electrical chip, packaging and other fields. There are relatively high technical bottlenecks, which requires the efforts of the whole industry to move forward.In the short term, CPOs cannot be deployed in a wide range of data centers. In the future, CPOs will coexist with pluggable optical modules for a while. The role of silicon optical technology in CPO is irreplaceable. Most manufacturers adopt the technology of microring modulator, which requires the control circuit to adjust the working point of the microring. It is worth mentioning that a large part of the voices in the OFC conference come from traditional IC manufacturers, including Intel, Nvidia, Broadcom, Marvell, Cisco, etc., or from the cooperation with major manufacturers, such as Ranovus and AMD Xilinux. We don't see any Chinese IC in this field. We don't know whether there are Chinese manufacturers involved in this field.