Chinese researchers have reportedly developed a new “three-lane highway” optical fiber system. The new system, the researchers explain, could dramatically expand the data-transfer capacity of future networks, notably those for artificial intelligence (AI).

Put in simpler terms, the innovation effectively “squeezes” more traffic from fiber-optic cables without having to replace or lay new cabling. According to the team, they have been able to increase traffic by as much as five times while also increasing transmission capacity per core by nearly 50%.

Today, most internet traffic travels through glass fibers using laser light. Importantly, each fiber doesn’t just carry one beam of light. It carries many wavelengths (colors) of light simultaneously, like multiple radio stations sharing the airwaves.

Presently, most commercial systems use a mixture of C-band (the workhorse) and L-band wavelengths. The latter is used to extend capacity when needed.

The team, they report, has managed to add a third, called S-band; hence the “three-lane highway” analogy. The fiber itself isn’t necessarily wider; they’re simply using more of the available light spectrum.

So, you might ask, why hasn’t S-band been used before? Well, in short, over long distances, optical signals tend to weaken. To overcome this, telecom companies use optical amplifiers placed along the route.

The industry’s amplifiers were optimized for C-band (and later L-band). S-band signals were more difficult and expensive to amplify cleanly. To this end, the Chinese team claims to have solved enough of these engineering problems to make commercial deployment practical.

But there is another trick up the team’s sleeve: multiple cores to the fibers. Normal optical fibers consist of a single core, but the team’s new ones reportedly actually contain four.

To this end, each of these core acts like an independent fiber. So, unlike traditional fibers with a single core and two optical bands, the new fibers actually have four cores, with three optical bands working simultaneously.

That’s a significant leap in capability. For applications like AI, this could prove to be very useful indeed, especially when it comes to removing data-transfer bottlenecks.

At present, most large AI models tend to use thousands of graphics processing units (GPUs), constantly exchanging data at the speed of light. A modern AI cluster can move terabits per second between machines.

5x the data transfer

Often, the bottleneck is not the GPU but getting data from one GPU cluster to another. So, if you can move five times more data through the same fiber infrastructure, you necessarily get faster training, larger distributed models, and lower network congestion.

Looking at the bigger picture, this isn’t a breakthrough in terms of fiber optics; similar achievements have happened in Japan, Europe, and the US. What is different is that this appears to have been commercially deployed, not just developed in a lab.

According to reports, the new optic fibers have run over about 21.7 miles (35 km) of an existing telecoms network, combining the innovations mentioned. Looking ahead, the team suggests the new cables could be deployed internationally, perhaps even in submarine cables.

It could also play a major role in China’s “Eastern Data, Western Computing strategy,” which seeks to shift data-intensive workloads from densely populated eastern regions to resource-rich western provinces, according to the developers.

Recommended Articles