​

  (Shutterstock)

The use of silicon photon chips is expanding, with a compound annual growth rate of 36% in the overall market value. Five big tech giants are actively investing in development.

 

Since the introduction of silicon photon technology to the market with generative AI, its importance and attention have soared. In the data center boom of generative AI, there is not only a higher demand for computing power but also increased standards for data transmission speed, bandwidth, loss, and power consumption. In industry dynamics, silicon photon is identified as one of Taiwan's four potential markets! The article mentioned an estimation of compound growth rate may be as high as 36% through 2027. Silicon photon technology transforms "electrical signals" into "optical signals" for message transmission. Since light itself has no charge and no mass, signals do not interfere with each other, and there is low energy and signal loss.

 

In the future, silicon photon chips are expected to be applied in various fields, including lidars for autonomous vehicles, consumer healthcare, and high-speed transmission in AI data centers. The ultimate goal is an All-Optical Network (AON), where all communication between chips, including random storage, transmission, exchange, etc., is done using optical signals. Research firm Yole Intelligence estimates the total market value of silicon photon chips to reach $972 million by 2027, with a compound annual growth rate of 36%. Currently, the main participants are five major tech giants, including Broadcom (AVGO), Marvell (MRVL), NVIDIA (NVDA), Cisco (CSCO), and Intel (INTC).

 

Generative AI is driving a significant increase in the demand for high-speed transmission in data centers. Silicon photon network equipment is expected to become another essential tool for AI development.

 

Silicon photon has become a market focus due to the emergence of generative AI, leading to a substantial rise in the demand for high-speed, low-latency, and low-power transmission technology within data centers. The development of generative AI can be divided into two main stages: training and inference. In the training phase, the need for analyzing large amounts of data increases the demand for chip processing speed, causing a explosive growth in the demand for High-Performance Computing (HPC) GPUs. However, having high-power GPUs alone is not sufficient. A large amount of data and computation results need to be transmitted rapidly between GPUs, requiring an increase in the bandwidth of communication between GPUs and a further reduction in power consumption.

 

The internal data transmission pipeline within a data center is like a highway about to end after a long weekend. As traffic increases, if lanes cannot be added, traffic jams often occur, significantly reducing the speed of travel. Similarly, when the volume of messages within a data center increases (due to a large amount of data and analysis results), if the bandwidth of communication between GPUs (lanes) is not increased, bottlenecks in signal transmission occur, and the speed of message transmission (travel speed) slows down. Moreover, when data is transmitted using "electrical signals," the resistance of metal copper wires causes signal loss or damage, leading to errors in the results of Large Language Models (LLM) and generative AI. The application value is at risk of a significant reduction.

 

Furthermore, when large language models enter the inference stage, there is a need to rapidly and accurately transmit a large amount of data to quickly provide information to users. Therefore, data centers, especially cloud data centers, have a significantly increased demand for network equipment (such as optical transceivers and switches) with high bandwidth, low latency, and low loss. This is to eliminate data transmission bottlenecks within data centers, maximizing the practical value of generative AI. Thus, in the AI gold rush, silicon photon network equipment is expected to become another essential tool, alongside high-performance GPUs.

 

The Development Timeline of Silicon Photonics: From Integrated Circuits to Integrated "Light" Paths

 

In the current market for silicon photonics optical transceivers, the mainstream product is the pluggable transceiver, which looks like a USB interface and connects to two optical fibers for transmitting and receiving light. Electronic signals have to go through circuit boards and carriers before reaching the Switch ASIC, resulting in slow signal transmission and the possibility of signal loss. The next-generation Co-Packaged Optics (CPO) is gearing up, intending to assemble the optical transceiver and Switch ASIC on the same carrier board. This shortens the path for electronic signal transmission to just the carrier board, speeding up signal transmission and reducing signal loss.

 

The next generation of silicon photonics technology aims to address the issue of electrical signal transmission between computing chips using Optical I/O. It replaces electrical connections between computing chips with optical links, utilizing optical waveguides for communication between computing chips. This accelerates the computing speed of AI data centers, resolves current computational bottlenecks, and gradually shifts all inter-chip communication to optical signals, moving towards a transition from electronic networks to fully optical networks.

 

Pluggable Transceivers: Main Application Product in Silicon Photonics Network Equipment

 

Silicon photonics network equipment includes optical transceivers (Transceivers), cables, optical switches, sensors, optical attenuators, and modulators. Among them, optical transceivers have been widely adopted in data centers and are the primary application product of silicon photonics chips. Optical transceivers are crucial components for signal transmission in optical fiber communication and play a vital role in achieving a fully optical network. They convert electronic signals used by regular communication devices into photon signals for optical fiber communication and vice versa. According to a report from research firm Yole Intelligence, optical transceivers in data centers will continue to dominate the silicon photonics market, with a Compound Annual Growth Rate (CAGR) reaching up to 22%.

 

Pluggable Transceivers' Specifications Led by Marvell and Broadcom, Intel and Cisco Follow Suit

 

The mainstream product in silicon photonics optical transceivers is the pluggable transceiver, which means inserting the transceiver onto the printed circuit board (PCB). The path for electronic signal transmission requires passing through the circuit board and carrier before reaching the Switch ASIC. The maximum transmission speed of pluggable transceivers can reach up to 800 gigabits per second (Gb).

 

Marvell's COLORZ 800, launched in August 2023, is the industry's first pluggable transceiver with a transmission speed of 800Gb per second and a transmission distance of up to 1,200 kilometers. It can be used to connect data centers (DCI) in different cities and is expected to begin sample testing in the fourth quarter of 2023, with shipments expected within 9 to 12 months after testing. Broadcom offers a diverse range of 800Gb products, including pluggable transceivers with transmission distances ranging from 100 meters to 10 kilometers.

 

Intel showcased an 800Gb pluggable transceiver at the 2022 European Conference on Optical Communication (ECOC) but has not yet announced an official shipping schedule. Cisco's latest product, Cisco QSFP-DD800, does not provide a single port with a bandwidth of 800Gb per second; it only offers a port with a transmission speed of 400Gb per second, indicating that there is still some distance in product specifications compared to the two major leaders, Broadcom and Marvell.

 

Co-Packaged Optics (CPO) Transceiver, Shortening Signal Transmission Distance, Speeding Up Data Transfer

 

With approximately 80% of data transmission currently occurring within data centers and considering the spine-leaf architecture of data centers, the switch plays a crucial role, with each data transfer passing through at least three layers of switches. Data travels from the server storing information to the top-of-rack switch (illustration 1), then from the top-of-rack switch to the leaf switch (illustration 2), to the spine switch (illustration 3), then to the leaf switch of the destination server (illustration 4), and finally reaching the destination server (illustration 5).

 

A switch consists of an optical transceiver and a Switch ASIC (Application-Specific Integrated Circuit customized for switch design). The next generation of Co-Packaged Optics (CPO Transceiver) assembles the optical transceiver and Switch ASIC on the same carrier board, making the transmission path for electronic signals only pass through the carrier board. This shortens the signal transmission path, reducing losses and delays during signal transmission, and speeding up the internal data transfer within the data center.

 

CPO Co-Packaged Switch Specifications Led by Broadcom, Followed Closely by NVIDIA and Marvell

 

Broadcom's latest generation switch, Tomahawk 5, achieves a data transmission bandwidth of 51.2 terabits per second (1 terabit = 1,000 gigabits, 51.2 Tb = 51,200 Gb). Broadcom announced in a press release on March 15 that Tomahawk 5 has started shipping, making it the market's first and currently the only commercially available switch with a bandwidth of 51.2 Tb per second. The highest transmission speed of a single port is up to 800 gigabits per second.

 

NVIDIA unveiled Spectrum 4 designed for large data centers at the GTC conference in March 2023, with a bandwidth of 51.2 terabits per second. The bandwidth of each port is doubled, and the highest transmission speed per port is comparable to Broadcom's Tomahawk 5 at 800 gigabits per second. NVIDIA stated that Spectrum 4 is expected to start shipping by the end of 2023, aiming to become a strong competitor in the market, leveraging NVIDIA's leading GPUs and its highly adhesive ecosystem.

 

Marvell also introduced the Teralynx 10 switch with a data transmission bandwidth of 51.2 terabits per second. Marvell mentioned in its fiscal year 2024 Q2 earnings call on August 25 that Teralynx 10 has entered the trial phase and is expected to contribute revenue starting from fiscal year 2025 (after February 2024). However, the penetration rate is expected to increase in the more distant future, indicating that the shipping time for Teralynx 10 might be 1-2 quarters behind NVIDIA's Spectrum 4. Nevertheless, the highest transmission speed per port of Marvell's Teralynx 10 surpasses both Tomahawk 5 and Spectrum 4, reaching 1.6 terabits per second.

 

Cisco announced the Cisco Silicon One G200 switch on June 20, adopting CPO technology similar to Broadcom's Tomahawk 5, NVIDIA's Spectrum 4, and Marvell's Teralynx 10. It also achieves a bandwidth of 51.2 terabits per second. Major cloud service providers like Amazon (AMZN), Google (GOOG), and Microsoft (MSFT) are currently testing the G200, but an official shipping date has not been disclosed.

 

Looking at the shipping order of the latest generation switches with a bandwidth of 51.2 Tb per second, Broadcom leads the CPO switch market, with NVIDIA and Marvell closely following. Cisco, with no shipping schedule yet, still has work ahead to catch up.

 

Next-generation Optical I/O technology has lower latency and greater growth potential than CPO, with Ayar Labs currently leading the way.

 

The mentioned CPO technology focuses on interconnecting servers within data centers, co-packaging switch ASICs and optical transceivers. For chip-to-chip interconnects among computing chips like CPUs, GPUs, and DPUs (collectively referred to as XPUs), Optical I/O technology can be used by replacing electrical signals with optical signals for signal transmission. Optical I/O technology co-packages computing chips with optical transceivers to accelerate the transmission speed between computing chips. According to Yole Intelligence's expectations, the Optical I/O market is poised for a Compound Annual Growth Rate (CAGR) of 68% from 2022 to 2028, driven by the growing demand for high-speed computation in AI and machine learning. The CAGR from 2028 to 2033 is even expected to further increase to 81%. Both periods of CAGR surpass CPO's 41% and 69%, indicating that Optical I/O has greater growth potential than CPO.

 

Optical I/O also co-packages optical transceivers and processor chips on a carrier board, similar to CPO but replacing the Switch ASIC with XPU (CPU/GPU/DPU). The strength of Optical I/O lies in its shorter electrical transmission path compared to CPO, resulting in lower latency. CPO has a latency of 100-150 nanoseconds (ns), while Optical I/O can further reduce latency to around 5 nanoseconds. Latencies above 100 ns may be acceptable in Ethernet scenarios, but they are intolerable in situations where a series of XPUs share machine learning workloads.

 

Currently, Ayar Labs leads in Optical I/O technology, being the first company to successfully implement optical signal transmission between chips. Although Ayar Labs has not gone public, it has collaborations with Intel and NVIDIA. Its key partners in the supply chain include the foundry Global Foundries (GFS) and the laser component supplier Lumentum (LITE).

 

Silicon photonics concept stocks to watch include CPO switch leaders - Broadcom, NVIDIA, and Marvell, while for the medium to long term, Optical I/O technology merits attention, with Global Foundries and Lumentum being noteworthy.

 

In summary, with significant improvements in data transmission speed, bandwidth, losses, and power consumption standards, silicon photonics technology is poised to become a key development direction in future data transmission technologies. Looking at specifications alone, the current leaders in plug-in optical transceivers and CPO switches are Broadcom, NVIDIA, and Marvell. For Optical I/O technology, which connects chips, Ayar Labs leads the way, and particular attention can be given to Global Foundries in the supply chain, serving as a foundry, and laser component supplier Lumentum.

 

Copyright Statement

The copyright for this article belongs the author and CMoney Financial News. Reproduction without explicit permission is strictly prohibited and may result in legal recourse.

Disclaimer

The information presented on this website is intended solely for reference and does not imply any endorsement for buying or selling. Investors are advised to assume full responsibility for the risks associated with trading.

​

​