Bench Talk

(Source: saicle / stock.adobe.com; generated with AI)

Artificial intelligence (AI) is commonly described as computer programming trained to sense or learn about its environment from data, act accordingly, and then adapt its behavior based on the results. This has led to another innovation called machine learning (ML), in which the system's performance improves over time as it recognizes patterns from data and learns better ways to analyze them. The more data the system is exposed to, the more accurate its predictions will be.

AI and ML are already working behind the scenes in many industries, analyzing vast amounts of data to recognize patterns and suggest courses of action. We see the benefits of these advances in a wide array of applications, from demand planning and understanding customer behavior to rapid image recognition for the latest safety systems.

Regardless of the application, AI systems have several features in common. The first is impressive computing power—the “brains” behind the intelligence. AI systems employ powerful processors that consume large amounts of energy to sift through the vast amount of data they are presented with.

Another feature of AI systems is the data. With such an emphasis on the computing power behind AI, the humble connector might be overlooked. Still, the infrastructure that supports AI relies upon the ability to transmit and receive a tremendous amount of information without error. With connectors playing an integral role in signal integrity (SI), power supply, and thermal management, let’s take a closer look at how these components factor into the AI revolution.

Signal Integrity

SI describes the quality of electrical signals transmitted through connectors and over cables, and as data speeds increase, it has grown in importance. Many factors affect SI, including the external environment's electromagnetic interference (EMI). Because these factors have the potential to cause interference from crosstalk, the cables, connectors, and printed circuit board (PCB) traces through which the signal travels must affect the signal as little as possible. The latest connectors feature reduced pin spacings, small sizes, and low profiles to make them suitable to handle SI for the latest handheld devices.

Manufacturers are developing new methods of connecting processors to minimize signal loss, especially in the data-hungry world of AI. In particular, the losses generated by high-speed signals passing through PCBs are very high, and in some cases, conventional PCB-mounted connectors are giving way to more advanced technologies.

Some designers are returning to the wires and cables that PCBs were introduced to replace decades ago. By mounting the cable as close as possible to the processor, the signals pass through cables in order to bypass the need for PCBs. Direct attach cables (DACs) transmit data with fewer losses over greater distances between servers in the latest data center.

Some manufacturers are introducing active solutions to boost the signal over longer distances with active electrical cables (AECs). These use active devices known as retimers, which recondition the data signal as it travels through the AEC on the way in and out. The retimer of an AEC creates a cleaner signal, removing noise and amplifying the signal to minimize losses. The drawback is that these active devices require power.

The Problem of Power

Power is one of the greatest concerns for AI system designers. The data center is home to many powerful processors, consuming vast amounts of energy, and connectors are essential for ensuring power integrity (PI), which aims to provide power to the system within acceptable limits. This means minimizing the voltage fluctuation across the consumer despite the fluctuating current demanded by the consumer. The cabling and connectors over which power is transmitted also significantly impact the PI of a power network.

Power connectors for data centers must be chosen with care. Should connectors feature fewer, larger contacts, or will a larger number of smaller pins be better? There is no single rule that is true for all applications. Some applications use unregulated input voltages where the system can tolerate greater AC fluctuations so that a single larger blade-type contact will be more suitable. A larger contact will be able to carry higher currents at greater voltages. However, once the power supply has been regulated, the problem of AC fluctuation becomes more important. In this case, the inductance of the connector will be lower if smaller contacts are used, connecting several in parallel to reduce the current transmitted through each.

Reducing Heat Through Smart Design

With such power demands on data centers in the growing age of AI, heat becomes another issue to address. Power connectors must play a role in thermal management. Every electrical circuit has resistance. The value may be small, but when passing power through that circuit, any resistance causes some of the energy to be converted into heat. With enough power, this can cause a rise in the temperature of the connector and, therefore, everything else around it.

When added to the heat generated by powerful processors, the challenge of removing excess heat becomes significant. Manufacturers are playing a key role in thermal management by incorporating smart design features into connectors. Some connectors, especially those intended for board-to-board applications within server cabinets, incorporate cavities to allow air to flow over and through them. Other connectors use high-conductivity alloys in the design of their power contacts, reducing contact resistance and allowing more energy delivery with a lower overall temperature rise.

Conclusion

Artificial intelligence and machine learning systems are changing how we handle data, but they cannot exist without robust and capable infrastructure to deliver data and power while removing heat. Although the connector may appear trivial initially, a poorly designed connector can make or break even the most advanced artificial intelligence systems.