The new driving force of artificial intelligence - silicon-based optoelectronics, thin-film lithium niobate and optoelectronic collaborative design, breaking the bottleneck of high-speed transmission of 200Gbps links - Xunshi Optical Communication Network (iccsz.com)

Abstract: Driven by artificial intelligence and ML technologies, data consumption continues to grow, creating an unprecedented demand for high-speed connectivity in modern data centers. To meet these needs, the industry is actively developing next-generation wired transceivers capable of supporting data rates in excess of 200 Gbps

Brief introduction

Driven by advances in artificial intelligence (AI) and machine learning (ML) technologies, data consumption has grown exponentially, creating an unprecedented demand for high-speed connectivity in modern data centers. As the complexity of AI models continues to increase, the number of parameters reaches astronomical (e.g., BaGuaLu of more than 37 million cores), the need for bandwidth and low-latency interconnects has become very important. This article will look at the ability to support more than 200 Gbps Industry trends, emerging technologies, and design considerations for next-generation wired transceivers for data transfer rates that are important for enabling seamless data flow in AI and data center applications, Tony of Alphawave Chan Carusone presented at the ISSCC2024 Forum titled, "The Impact of Industry Trends on 200+Gbps." Wireline R&D", this article will sort out the general content of it.

The megatrends driving the need for connectivity

1. AI connectivity and scalability

The rapid growth of AI and ML workloads has led to the deployment of large-scale computing clusters consisting of hundreds to thousands of accelerators (xPUs) interconnected via high-speed links. By 2027, it is expected to be about 50% of the market's revenue will be driven by AI-accelerated servers, with 20% of Ethernet data center switch ports connected to AI-powered servers. In addition, it is expected that 50% of these switch ports will be 400 Gbps or higher, 800 Gbps will grow faster than 400 Gbps by 2025 (Figure 1).

Figure 1: Projected Growth in AI Connectivity and Scaling (Source: Dell'Oro Group Data Center IT Capex Forecast, January 2023)

2. Disaggregate storage

Another important trend driving the need for high-speed connectivity is the rise of disaggregated storage architectures. By centralizing storage in shared pools, data centers can increase efficiency and enable larger shared pools, resulting in improved resource utilization. However, this approach relies on: Low-latency interconnects such as PCIe and CXL to ensure seamless communication between compute resources and disaggregated storage.

3. Wired transceiver trends

To meet the growing demand for bandwidth, wired transceiver data rates are doubling approximately every five years (Figure 2). This trend is expected to continue, in the near future, 200 Link transceivers will be widely adopted, with 400 Gbps and 800 Gbps link transceivers coming in the following years.

Figure 2: Released transceivers from 2010-2023, showing a trend of data rates doubling every five years (Source: ISSCC Forum)

4. Advantages of 200G links

A 200Gbps link has an advantage over a link with a lower data rate. For example, a 51.2 Tbps 1RU (rack unit) switch requires 32 modules, each with 16 x 100 Gbps optical links, doubling the number of lasers compared to an equivalent configuration of 8 x 200 Gbps links. By reducing the number of lasers, 200 Gbps Links can dramatically reduce power consumption and cost. In addition, higher per-channel data rates allow for flatter network topologies and higher arc switches, which reduces latency – a key requirement for AI workloads.

New technologies and considerations for 200G links

1. Inside the transceiver

To support data rates of 200 Gbps, wired transceivers must employ advanced digital signal processing (DSP) technology and robust forward error correction (FEC) Scheme. In order to mitigate the loss caused by severe channel loss (greater than 30 dB), a large number of equalizations such as decision feedback equalizers (DFEs) with a large number of taps must be used. In addition, DSPs such as Circuit Tap Limited Impulse Response (FIR) equalizers Technology can also help solve the problem of reflections in short cable channels.

FEC plays an important role in ensuring reliable data transmission over lossy channels. At 200 Gbps, a more robust FEC is required scheme, resulting in increased decoding complexity, power consumption, and latency. To balance the trade-offs between coded gain, power consumption, and latency, piecewise FEC (where each link segment is protected by its own optimized FEC) and series connections are currently being explored FEC (Double Protection for Optical Links) and other technologies.

An important architectural implication of adopting a soft decision FEC at 200 Gbps is the effective exclusion of analog serializer/deserializer (SerDes) architectures. Instead, it is necessary to use the FEC Tighter integration with analog front-end (AFE) facilitates analog-to-digital converter (ADC)-based DSP SerDes architectures.

2. 200G optoelectronics

It is currently being set at 200 Gbps per wavelength The application of research to various modulation techniques. Absorption modulated lasers (EMLs) are a promising option, offering modest swing requirements and the potential for differential drive configurations. However, there are still challenges in optimizing extinction ratio (ER) and chirp, especially at longer wavelengths.

Silicon-based optoelectronics (SiP) Mach-Zehnder modulators (MZMs) and microring resonator modulators (MRMs) are attractive due to their integration potential and low cost. However, for SiPs of 200 Gbps Achieving the required bandwidth, modulation efficiency (Vπ), and low optical loss at the same time remains a challenge for modulators.

Thin-film lithium niobate (TFLN) modulators are also being explored, offering high bandwidth and low drive voltage, but at a higher cost and with potential integration challenges.

3. Optical/electrical co-design

As data rates increase, co-design and co-optimization of optical and electronic components becomes increasingly important. For example, the encapsulated interconnect between a photodiode (PD) and a transimpedance amplifier (TIA) in a receiver has a significant impact on the wideband frequency response. Optimize trace impedance and use on-chip Technologies such as T-coils can increase bandwidth and reduce reflections.

In addition, the optimal design parameters may vary depending on the presence and function of DSP equalization. In the absence of DSP equalization, it is critical to minimize reflections, while in the presence of DSPs In a balanced situation, retaining some residual reflections can help to achieve better overall performance.

4. Optoelectronic Co-Packaged Devices (CPO)

To meet the challenges of chip-to-module interconnects and achieve higher total bandwidth, photoelectric co-packaged device (CPO) solutions are gaining traction. By integrating the optical engine with the ASIC in the same package, the CPO This eliminates the need for a reset timer, reduces power consumption, and reduces latency. However, CPOs also present some challenges, such as increased power density and thermal management within the package, as well as potential constraints to the innovation ecosystem.

Beyond 200 Gbps: Emerging Technology

1. Parallelism: WDM and PSM

In order to make the transmission rate of each wavelength more than 200 Gbps, technologies such as wavelength division multiplexing (WDM) and parallel single-mode (PSM) fiber architectures are currently being explored. WDM technology multiplexes multiple wavelengths on a single fiber, resulting in a higher combined data transmission rate. Compact modulation techniques, low-cost and low-loss wavelength multiplexers/demultiplexers, and multi-wavelength laser sources are key factors in enabling wavelength division multiplexing technology.

2. Higher-order modulation format

Increasing the baud rate and employing higher-order modulation formats such as 6-PAM and 8-PAM is to achieve data rates in excess of 200 Gbps per line potential avenues. However, these approaches require significant advances in analog bandwidth, DSP, and coding techniques.

3. Coherent optical communications

Coherent optical communication is a proven technology in long-haul networks that is being used for short-distance transmission within data centers. By utilizing coherent modulation formats such as dual-polarization quadrature amplitude modulation (DP-QAM), coherent links can transmit data at four times faster rates than intensity modulated and direct detection (IM/DD) links at the same baud rate.

Recent developments in lightweight coherent solutions tailored to transmission distances of up to 10 km have shown promising results. These solutions utilize the O-band (approximately 1310 nm) to reduce DSP power consumption while maintaining acceptable fiber loss for short-range applications. In addition, synchronous baud rate sampling DSP architectures are being explored to further reduce the power consumption and latency of coherent transceivers.

conclusion

Driven by artificial intelligence and ML technologies, data consumption continues to grow, creating an unprecedented demand for high-speed connectivity in modern data centers. To meet these needs, the industry is actively developing the ability to support more than 200 Next-generation wired transceivers with Gbps data rates.

Key technologies and considerations for 200 Gbps links include advanced DSP technology, robust FEC schemes, co-design of optical and electronic components, and the exploration of new optical modulation formats. In addition, optoelectronic co-packaging and coherent optical communications are emerging as promising solutions to meet the challenges of chip-to-module interconnects and enable higher data transfer rates within data centers.

When we set our sights on 200 Gbps At the same time, we are also actively researching wavelength division multiplexing technology, high-order modulation format, and short-range coherent optical communication. Collaboration across disciplines, including analog and digital design, coding theory, optics, and system architecture, is important to overcome challenges and enable seamless data flow in future AI and data center applications.

References

  [1]T. C. Carusone, “The Impact of Industry Trends on 200+Gbps Wireline R&D,” in IEEE International Solid-State Circuits Conference (ISSCC), 2024