In the coming years, scaling AI accelerator clusters in data centers will face increasingly complex challenges. System architects need to address three challenges simultaneously:

Deliver better performance to meet growing bandwidth demands. Control costs while expanding computing power and complexity. Continue to improve energy efficiency. These three challenges keep network operators awake at night.

While the emergence of new technologies creates opportunities for innovation, they also overwhelm data centers. New AI and machine learning workloads, such as generative AI and large language models (LLMs), are driving data bandwidth beyond traditional interconnects, with speeds rapidly doubling to 800G and soon to 1.6T.

Technical Limits of Terabit Speeds

To meet the growing demand, data centers rely on two solutions: 400 Gigabit and 800 Gigabit (400G/800G) network equipment, using copper cables for short distances and fiber cables for long distances. However, both technologies are reaching their technical limits in terms of terabit interconnect speeds.

Copper wire is the preferred interconnect solution for short-distance applications due to its low cost, simplicity and high reliability. The limitations of copper wire are that channel loss due to the skin effect can severely limit the transmission range of the cable, and as the transmission speed increases, the cable thickness also increases, as shown in below picture.

Copper cables are not sufficient to support network speeds of 1.6T and above. At terabit speeds, copper cables are too short and too thick to be scalable and are not suitable for deployment in high-density data centers.

The shift to optical interconnects

For many AI-related workloads, hyperscalers will turn to optical interconnects, such as active optical cables (AOCs). Optical interconnects can provide connectivity up to several kilometers, but are more complex, power-hungry, and costly because of the additional components required for electrical-to-optical conversion, such as optical DSPs, transimpedance amplifiers (TIAs), laser drivers, and lasers.

These optical cables integrate advanced digital signal processors (DSPs) and complex optical components to transmit and receive optical signals at high speeds. AOCs can support longer cable lengths than copper cables and are thinner and lighter. While this makes them easier to deploy, optical technology is inherently unreliable because optical performance varies with temperature and will eventually fail.

Optical DSP electronics can significantly increase latency, which can degrade network performance. Adding optical engines and DSPs can quickly become prohibitively expensive, costing up to five times more than copper cables. The same components also significantly increase the power consumption of the cables, increasing the energy requirements of data center operations.

A third option besides copper and fiber

All of this leaves hyperscale data center operators in need of a solution that overcomes the limitations of copper and fiber technologies while remaining cost-effective for large-scale deployments. This is where a third option comes in: e-Tube, a scalable multi-terabit interconnect platform that transmits RF data over plastic dielectric waveguides.

As shown in Figure 2, the active RF cable (ARC) uses e-Tube technology with an integrated mmWave RF transmitter to up-convert terabit data from the electrical domain to the RF domain. The antenna radiates the wireless signal, which propagates through the e-Tube core.

At the other end, a complementary mmWave RF receiver and antenna receive the wireless signal and convert it back to electrical. To the two systems connected by ARC, the interconnect appears as a single electrical system. ARC manages the conversion of electrical to RF and RF to electrical, making the conversion process transparent to the two connected systems.

Using plastic as the cable medium allows data to be transmitted efficiently and at low cost.

Made from common low-density polyethylene (LDPE) material, e-Tube cables are not subject to high-frequency losses like copper wires, making them a scalable interconnect solution for any data speed from 56G to 224G and beyond. The low-power RF transmitter and receiver ICs used for data transmission achieve the industry's best 3pj/bit energy efficiency with latency in the picosecond range.

Lighter, thinner, lower power consumption

The result is a cable that provides 10x greater reach than copper, is 5x lighter, 2x thinner, 3x lower power, 1,000x lower latency, and 3x lower cost. e-Tube addresses bandwidth needs that cannot be met with copper and fiber interconnect technologies. As data centers transition to 1.6T and 3.2T speeds, it is an ideal copper replacement for intra-rack and adjacent rack connections.

To accelerate deployment, this innovative interconnect technology, the e-Tube RF SoC, is manufactured using mature, standard semiconductor process technology and established IC packaging technology. For decades, the "connectivity" of connectors and cables has been mass-produced using copper twinax manufacturing technology. Cable designs meet industry-defined MSA package specifications such as OSFP and QSFP-DD, as shown in below picture.

This provides flexibility for different system designs as it helps ensure compatibility with existing network infrastructure equipment from different manufacturers.

As data center hardware rapidly evolves to support LLM and generative AI computing needs, a third interconnect option is needed to alleviate the limitations of copper cables, which are much less expensive and energy efficient than optical fiber. e-Tube RF based on plastic media is expected to revolutionize computing fabric interconnects, providing a unique combination of power efficiency, longer cable reach, lower latency and cost points to expand the scale of AI clusters in data centers in the next few years.

Source: This article is translated from allaboutcircuit

Reference link https://www.allaboutcircuits.com/industry-articles/beyond-copper-and-optical-a-new-interconnect-eyes-next-gen-data-centers/

View more at EASELINK