An Introduction to 224G System Architecture

Artificial intelligence (AI), machine learning (ML), streaming and other high-performance applications are placing unprecedented demands on data centers. The need for faster speeds, higher bandwidths and lower latencies are pushing the limits of today’s data centers — a challenge that is only expected to grow as applications like generative AI evolve.  

Systems operating at 224G PAM4 (Pulse Amplitude Modulation 4-level) are now a necessity for data centers to support data-hungry applications and workloads, as they provide a range of benefits across speed, bandwidth and latency. But operating at 224G isn’t a simple upgrade — it will require a new era of facilities purpose-built for emerging technologies.  

224G PAM4 is a high-speed data transmission technology that utilizes a PAM4 modulation scheme to achieve a data rate of 224 Gigabits per second. 224 represents the next evolution of data transmission, doubling the rates of 112G systems, and is used to support the growing bandwidth and performance requirements of applications like generative AI, autonomous driving, 5G/6G and high-performance computing. 

PAM4 is a multi-level signaling scheme that uses four different voltage levels to represent two bits of data per symbol. Unlike traditional non-return-to-zero (NRZ) signaling which uses only two voltage levels to transmit data (0, 1), PAM4 can achieve twice the data rate (00, 01, 10, 11) by using four levels. By transmitting two bits per symbol, PAM4 operates at a lower symbol rate than NRZ modulation and therefore consumes less power than NRZ when producing the same data rate. Further, this bandwidth efficiency makes PAM4 a more scalable technology than NRZ for high-bandwidth applications. 

PAM4 has unique design considerations because its structure can make PAM4 more susceptible to intersymbol interference (ISI) — a form of distortion in which symbols interfere with one another. ISI can introduce issues such as an increased bit error rate (BER) and reduced data integrity, which can be serious problems for applications such as cloud computing, AI and ML, which require accurate and reliable transmission of data. ISI emphasizes the need to work with reputable vendors such as Molex that use advanced equalization techniques to manage signal integrity in PAM4 solutions. 

PAM4 is the preferred modulation scheme for transmitting data at 224 Gbps due to higher bandwidth efficiency, reduced power consumption and improved scalability. Since PAM4 utilizes two bits per symbol, 224G PAM4 operates at a symbol rate of only 112 Gbaud. A 224 Gbps-NRZ system would need to operate at a symbol rate of 224 Gbaud to achieve the same data rate, increasing the power required and limiting the scalability.  

Outside of the symbol rate and encoding scheme, the components within the system — such as the cables and connectors — along with hardware such as transceivers must be designed to support 224G PAM4.  

For emerging applications like generative AI and high-performance computing, 224G-capable data centers will not be optional. The need for faster speeds, lower latencies, greater scalability and even more mechanical robustness will stress today’s most advanced data centers — all of these requirements are met by 224G-PAM4 technologies but may require unique design requirements. 

224G networking offers a significant increase in both speed and bandwidth over the previous standards of 112 Gbps and 56 Gbps. These increases are necessary to support critical applications like facial recognition, 5G, 6G rollout and vehicle-to-everything (V2X) communication.  Faster speeds and higher bandwidths also help the training of highly sophisticated AI models, which consume and produce massive amounts of data. 

224G PAM4 technology can also help to reduce transmission latency. Transmission latency is the time it takes for data to travel from the source to the receiver. Lower latency is important for applications where delays could cause system failure or interrupt user experiences. For instance, video conferencing and AR/VR applications require low latency to provide a smooth and immersive experience. Remote surgery will rely on incredibly low latency to ensure procedures are performed safely and accurately. And in autonomous vehicles, latency can influence how quickly a vehicle reacts to road conditions and the surrounding environment. 

224 Gbps-PAM4 networking components are designed to be more mechanically robust than previous generations of solutions. This is important for ensuring reliable operation in data center environments where high levels of heat and vibration can cause system failure.

Scalability, flexibility and modularization are necessities for future-proofing to new and revolutionary technologies like 6G networking where high bandwidth, low latency performance will be required. 224G solutions, like Molex’s Mirror Mezz Enhanced Connectors,  allow architectures to quickly scale to both increasing and decreasing demands, ensuring efficient and reliable operation when and where it is needed most. 

Although many of the most demanding applications for 224G solutions are only at their early stages of emergence, their adoption is unprecedented. Generative AI is quickly becoming commonplace across business operations, consumer applications and for potential integration into systems and devices — emphasizing the need for supporting infrastructure. Further, established applications like streaming, the cloud and the internet of things (IoT) are continuing to grow but require more capable solutions to evolve. 

Generative AI applications are quickly gaining traction across industries and use cases. Most known for being able to produce text, images, video and other forms of media at a quality similar to that of a human, these applications require massive amounts of data to train to increasing levels of sophistication. Additionally, the simple operation of generative AI — in which some form of input prompts a response — produces large amounts of data. 

Machine learning is a subset of AI that enables a machine to learn from data, make predictions or decisions and improve processes. As such, the use of ML applications is growing steadily within businesses, such as for forecasting, fraud detection, product recommendations and energy allocation.

However, like generative AI, ML applications are also data intensive — both in requiring large data sets for training and in the continuous data ML models can produce.   

High-performance computing uses powerful computers and software to solve computational problems that are too complex for traditional computers. Although HPC is already used in a wide range of fields, including scientific research, engineering, finance and manufacturing, its capabilities will increase significantly when combined with the scalability of 224 Gbps-PAM4 data centers.  

Digital twins are virtual representations of physical objects or systems that are created from data sourced from sensors and other devices attached to the object or system. Digital twins can be used to monitor current conditions, predict future behaviors and simulate potential scenarios. For instance, digital twins can be used to simulate the effects that environmental conditions may have on the logistics of a trucking company — such as the potential of a hurricane or blizzard. Digital twins can also be created at the device and even component level to predict the effects of use, such as accidental drops.  

The Internet of Things (IoT) is a network of physical objects that are embedded with sensors and other devices that allow them to collect and exchange data. By the year 2030, it is estimated that there will be close to 30 billion IoT devices connected to the internet — a growth driven by the increasing affordability and availability of IoT devices and the growing demand for data-driven insights. 

224G PAM4 networking will be critical to ensure data from IoT devices can be transmitted, processed and delivered through the cloud. 

Made a household necessity by the likes of Netflix and YouTube and further fueled by an increase in video networking, audio and video streaming is expected to continue its growth trajectory, especially as it becomes an alternative to typical broadcast media. Streaming’s dependency on high bandwidths, high speeds and low latencies ensures content is consistently available at high resolutions. 

Although 5G networking is still being rolled out, the early plans for 6G are already in the works. Whereas the impact of 5G became most prominent in consumer devices, 6G is expected to be more focused on the workplace and will help realize operational goals such as Industry 4.0 — placing even greater emphasis on speed and reliability. 

Autonomous vehicles are expected to become widespread in the next decade, and 224G networking will be essential for their deployment. Autonomous vehicles and the supporting vehicle-to-everything (V2X) infrastructure require high-bandwidth, low-latency networking to effectively communicate and ensure safe conditions for drivers, passengers and pedestrians alike

Edge computing will become more powerful, capable and reliable due to integration with 224G data centers, enabling edge networks to better support applications such as smart homes, telemedicine, smart grids and the industrial internet of things (IIoT). Even technologies like augmented reality (AR) and virtual reality (VR) will benefit from improved edge computing performance by creating a more lag-free, immersive experience.

Quantum computing uses the principles of quantum mechanics to solve problems that are too complex for most computers. Quantum computing is expected to revolutionize many industries from healthcare to research and financial modeling but will require high-bandwidth networking to realize its potential.  

Although 224G adoption is a necessity for the continued emergence of new technologies, it is not as simple as replacing a few components and systems. Several barriers to adoption exist that, while not insurmountable, stress the need to work with a trusted provider specializing in 224 Gbps-PAM4 technology. Identifying and addressing challenges at the earliest stage of the design process will help minimize costly errors.  

Data center operators are always looking for ways to improve performance and efficiency, but they also need to be mindful of the cost and disruption involved in upgrading to new technologies. The transition from 112G to 224G PAM4 networking requires new devices, components and system architectures that are capable of operating at 224G. Existing networking installations, such as switches, routers and cables, will need to be replaced. 

Minimizing signal integrity issues is one of the greatest technical challenges to implementing 224G PAM4 systems. Careful consideration is needed during the design and deployment stages to ensure long-term reliability. More information on signal integrity is provided within the section titled “What Influences Signal Integrity in 224G Channels?”

Compatibility between 224G networking equipment from different vendors is critical, but there is currently no finalized standard for 224G PAM4 networking. This means that data center operators need to be careful when choosing 224G equipment to ensure compatibility with other equipment in their network, emphasizing the importance of working with a trusted, recognized solution provider.  

224 Gbps-PAM4 is more complex than previous generations of networking. Data center operators need to have the expertise to design, deploy and manage 224G networks, especially in how to address signal integrity challenges. This can be a challenge for smaller data centers that do not have the in-house expertise and even for hyperscale facilities where acquiring skilled operators has become increasingly competitive. Molex helps address this issue for our customers by providing a consultative experience at the earliest stages of design to ensure challenges are recognized and addressed to ease implementation, maintenance and operation. 

224G networking equipment consumes more power than previous generations. This can be a barrier for data center operators who are trying to reduce their energy consumption. That said, 224 Gbps PAM4 solutions are also more power efficient than comparable NRZ solutions. 

Despite the barriers, data centers must adopt 224G networking in order to support the rapidly growing data demands of generative AI, ML, digital twins and other emerging applications. The benefits of 224G PAM4 networking far outweigh the challenges and risks in not doing so. 

Connectors play a critical role in 224G system architectures, such as near or at ASICS, at cable interfaces, on the fabric channels and on the scale out channels. This places unique emphasis on the quality of the cables and connectors used. 

Some ASICs are now being placed on the PCB as modules, requiring a board-to-board or mezzanine connector system.  

The fabric channel is made up of a cable and connector system that enables the high-speed transfer of data between modules and can provide designers with flexibility in their architecture selection. 

Connectors exist in several locations within scale out channels such as standardized high-speed copper connectors that connect copper or optical cables to the PCB. Further, a flying lead cable is connected to a near-chip connector to transfer signals into or near the ASIC. 

Operating at 224G poses significant design challenges, most notably in balancing signal integrity with mechanical form factor requirements. Failing to address this balance early in the design stage can lead to devastating results ranging from selecting incompatible components to challenging thermal management and decreasing performance. When prioritizing signal integrity, taking a system-level modeling approach with S-parameter models for the candidate connector products is the best way to qualify a design. Here are several factors found in 224G systems that influence signal integrity. 

The interface, cable and land pattern of connector interface define the input impedance.   Further, the physical structure of the interface determines pin-to-pin crosstalk and emissions from the connector body.

The load capacitance and package inductance at the I/O pins negatively impact signal integrity. Additionally, the chip land patterns going into the PCB or module connector must be wideband to support high frequencies like 224G.

The circuit board can introduce capacitive loading via parasitic capacitance along transmission lines.

Managing losses due to copper roughness is pivotal at mid-range frequencies operating below or near Nyquist. For frequencies near or above Nyquist, dielectric losses become dominant. Crosstalk should be minimized and maintained below certain limits up to Nyquist, with -40 dB being typical.  

Molex is a leading provider of  for high-speed data centers, offering a comprehensive family of products that support both scale-up and scale-out architectures. Molex's 224G solutions are designed to meet the growing demands of high-performance computing workloads, enabling data centers to scale up their performance and scale out their capacity to meet the needs of their customers.

With modular hardware, architects can add or remove modules to tailor each chassis to their specific needs, incrementally scaling power as necessary. The new Mirror Mezz Enhanced connector enables this modularity by supporting 224G PAM4 with improved impedance tolerances and crosstalk.  Mirror Mezz Enhanced also maintains an industry-leading density and the hermaphroditic nature of the Mirror Mezz line to reduce SKUs. When used in conjunction with the Inception/CX2-DS fabric, a system can connect between separate chassis using backplane connectors and cables. 

Molex's next-generation, MSA-compliant OSFP, QSFP and QSFP-DD product families include SMT, BiPass, Direct Attach (DAC) and Active Electrical Cable (AEC) solutions that are built for 224G. These robust cables and connectors offer superior mechanical durability and excellent shielding to minimize crosstalk and deliver better signal integrity (SI) performance at a higher Nyquist frequency.

The demand for data is insatiable, and satisfying the need for higher speeds and bandwidths is a necessity to enable the emerging applications that will support the day-to-day functions of businesses and people. Molex is paving the way for the next generation of data centers by providing a first-to-market complete portfolio of 224G products and solutions.

Molex's co-development approach with our customers has ensured that we’ve designed and manufactured the highest quality interconnect solutions to support applications such as generative AI. And our diverse team of experts specialized in design considerations across hardware, architecture, connectivity, mechanical integrity and signal integrity provides our customers with a personalized, consultative experience starting at the earliest stage of design. 

To learn more, explore Molex’s 224G products and solutions.