By Sean Park, Founder and CEO, Point2 Technology
[CONTRIBUTED THOUGHT PIECE] Data centers are in the midst of a rapid transformation to support the increasing number, complexity and size of AI and HPC workloads. As part of that transformation, large language models and generative AI applications are driving hyperscalers to adopt a new data center-scale switch fabric, called the AI backend network. And there is a new interconnect technology for this switch fabric based on a common and inexpensive material that outperforms anything in the market.
The AI backend network needs to operate at multi-terabit speeds to connect thousands, tens of thousands, and even hundreds of thousands of GPUs, TPUs and CPUs in the same cluster to operate together as a giant accelerator. The demands on the new switch fabric have elevated the importance of the cables, or interconnects, that connect chips in the same rack and between racks. The interconnects are increasingly critical infrastructure for AI/ML data centers because they ensure the optimal compute efficiency of the network accelerators.
Conventional interconnects rely on copper and fiber-optics, but both have limitations that are increasingly exposed with escalating AI workloads. Copper-based interconnects are reaching their physical limits as network speeds rise to 1600G and above. And while optical interconnects can support multi-terabit speeds over longer distances than copper, they are also power hungry and expensive, and they pose significant heat and reliability challenges.
Taken together, these limitations have left hyperscalers and enterprises seeking alternatives, including e-Tube technology.
e-Tube technology is a new, scalable interconnect platform that uses radio wave transmission over a dielectric waveguide made of – drumroll – common plastic material such as low-density polyethylene (LDPE). While waveguide theory has been studied for many years, only a few organizations have applied the technology for mainstream data interconnect applications. Because copper and optical interconnects are historically entrenched technologies, most research has focused on extending copper life or improving energy and cost efficiency of optical solutions.
But now there is a shift toward exploring the e-Tube option that delivers a combination of benefits that copper and optical cannot, including energy-efficiency, low latency, cost-efficiency and scalability to multi-terabit network speeds required in next-gen data centers.
The key metrics for data center cabling are peak throughput, energy efficiency, low latency, long cable reach and cost that enables mass deployment. Across these metrics, e-Tube technology provides advantages compared to copper and optical technologies.
Traditionally, copper-based interconnects have been considered an inexpensive and reliable choice for short-reach data center applications, such as top-of-rack switch connections. But copper wires are physically constrained by a phenomenon called skin effect, which causes greater levels of signal loss as frequency increases. This means that at higher data rates the signal can only be transmitted and received at significantly shorter cable lengths. To compensate, thicker copper wires (AWG) can be used. However, thicker AWG results in thick, heavy, and rigid cable bundles that cannot be deployed and serviced in high-density data center rack designs.
Figure 1. e-Tube cables are more affordable than even copper interconnects. (credit: Point2 Technology)
In contrast, skin effect does not happen with e-Tube technology. In fact, its channel loss does not change as data rates increase. As a result, the same e-Tube cores can support current network speeds as well as whatever speeds the future brings, from 800G to 1600G, 3200G, and beyond.
Despite copper’s high level of signal loss, its low cost continues to make it attractive for select, short-reach applications. However, plastic-based e-Tube cables come in at similar price points as copper (Fig. 1) and they are up to 80 percent lighter in weight and about half as bulky. e-Tube’s thinness and flexibility (bend radius) helps to alleviate rack congestion and simplify field servicing. In combination with its pricing, this makes e-Tube interconnects an alternative to copper in multi-terabit in-rack and adjacent-rack use cases up to seven meters.
Compared to active optical interconnects, such as linear pluggable optical (LPO) or re-timed active optical cable (AOC), e-Tube technology offers impressive advantages. Optical cables are more expensive than both e-Tube and copper-based cables (Fig. 1) because they require expensive electrical and optical assembly components to make the conversion between the electrical and optical domains. Many manufacturers supply optical cables, but building the cables and modules with nanometer wavelengths remains challenging, requiring high-precision alignment and production processes to achieve acceptable yield.
Figure 2. As an electrical technology, e-Tube provides energy-efficiency advantages over interconnects that require power-hungry optical assemblies. (credit: Point2 Technology)
While newer technologies such as co-packaged optics (CPO) improves the energy efficiency from existing AOC, it remains an optical technology that is expensive and prohibitive for reliable deployment for high-volume in-rack, adjacent rack, and backplane applications.
Since e-Tube is an all electrical technology, it does not require expensive, power-hungry optical assemblies and complicated optical DSP (digital signal processors). The technology is roughly 50 percent more energy-efficient than the lowest-power optical technology shown in Figure 2, with e-Tube, CPO and re-timed optics compared at 100G and 200G (lane rates used to build 800G and 1600G cable bundles). e-Tube’s energy efficiency comes from its simple, elegant architecture, where only an RF (radio frequency) transmitter and receiver IC (integrated circuit) pair are required for electrical-to-RF conversion for terabit-level data transmission.
The greater energy efficiency of e-Tube technology alleviates data center load on the power grid in two ways. First, power saved per cable translates to lower operating cost, as shown in Figure 3. Second, with less power (heat) to dissipate, significant energy savings from the cooling system can be achieved in high-density data center installations.
As an all-electrical technology, e-Tube also performs reliably across temperature ranges, whereas in small-module form factors the high power consumption of an optical module creates thermal management challenges that lead to reliability issues. In optical cabling, the required laser, photodiode and other components are adversely impacted at high temperatures, leading to reduced signal quality, increased bit error rate, higher product wear and potential optical assembly failure.
Figure 3. The lower power consumption of e-Tube cables translates into annual savings. (credit: Point2 Technology)
e-Tube technology also delivers advantages in latency, which is critical for AI/ML servers to reach peak performance. Lower latency enables fast synchronization of accelerator servers for parallel processing and provides efficient architecture that can scale to meet the demand of future AI workloads.
Figure 4. e-Tube latency is three orders of magnitude better than traditional optical cables. (credit: Point2 Technology)
At about 80 picoseconds, e-Tube latency is three orders of magnitude better than traditional optical cables (Fig. 4). The higher latency for optical cabling comes from the DSP, which includes multiple processing blocks (ADC, DSP, memory and DAC). e-Tube’s low-latency performance stems from its simpler architecture, with only an RF transmitter and a receiver IC pair, resulting in minimal delay when transmitting data.
e-Tube has been in development for several years anticipating the need for new interconnect solutions to meet the needs of next-gen AI/ML data centers. RF transmission through waveguides is not a new concept. Companies have commercialized flexible waveguides with RF wireless transmission at lower speeds and short distances for test and instrumentation applications, as an example. e-Tube is the first use of plastic waveguides for terabit data transmission in mainstream applications, such as data center interconnect. Increasingly, today’s data centers will include new cabling – moving away from traditional copper and optical cables – to e-Tube. The current and future use cases demand an offering like e-Tube that delivers better power consumption, cost, module latency, cable reach and more.
In sum, e-Tube technology offers a unique combination of power efficiency, cable length, low latency, and attractive cost structure to make it an ideal interconnect solution to scale with the compute requirements of future AI/ML data centers.
Figure 5. e-Tube offers benefits over common copper and optical alternatives for AI/ML data centers. (credit: Point2 Technology)
Sean Park is founder and CEO of Point2 Technology.
