Each week we find a new topic for our readers to learn about in our AI Education column.
Speed kills, but it can also sustain an ecosystem.
When we say speed kills, we’re often referring to driving or crucial decision-making, when going to fast can actually put ourselves and others in mortal peril. If we’re going too fast when we’re suddenly brought to a stop by the ground, or by unforeseen circumstances, we’re done for.
On the other hand, there are entire industries that have built and sustained themselves in part on speed, and two of them happen to intersect right here in our AI & Finance newsletter: technology and finance. That’s why we’ve decided to talk about speed in this weeks AI Education column, which is about optical connectivity.
You may recall that we’ve written earlier this year about optical computing, where instead of wires carrying electricity to move information within integrated circuits, light is used, being faster and in many ways more efficient. Optical computing is technically a form of optical connectivity taking place on a microscopic scale, within microprocessors. Optical connectivity, however, is more concerned with moving information with light on a grander scale.
And it really isn’t a new thing. Not at all.
Okay, What Is Optical Connectivity?
Optical connectivity is just using light to move information. It’s the information age-version of semaphore or an aldis lamp or smoke signals, older methods of communicating using light, and it carries some of the same benefits—the chief one being that in a vacuum, light is capable of moving near the speed limit for everything in the universe, and we have no known way to make information move faster. The Minutemen did not have to wait long at all to see if Paul Revere had hung one lantern or two in the Old North Church.
Proposed in the 1960s, fiber optic networks, where information is converted to light and transmitted over distance via a reflective glass tube, started coming into service in telecommunications in the 1970s. We were moving ideas by cable at near the speed of light 50 years ago. By the mid-to-late 1980s, with the rise in computing, the backbones of national and global fiber networks were coming online—this was also when the ethernet local area network (LAN) became a thing. In the 1990s, fiber was used to support the growth of the internet and cable television and the first data centers, with internal optical connectivity, were built. In the 2000s, fiber came to the home.
Of course, in the 1990s and 2000s optical connectivity also came to financial markets in a big way, eventually sparking a race to link exchanges and hubs with fiber to facilitate high-frequency trading—as is detailed in Michael Lewis’s book “Flash Boys.” It’s hard to believe that book came out 14 years ago—it’s opening frame, which takes the reader to the laying of a fiber optic cable between Chicago and New York, was set way back in 2009, long before generative AI came about.
That was the first peak in optical connectivity, in the late 2000s, as in-home high-speed internet and networking became for many people a norm, even if they didn’t know that their faster internet speeds were because the information coming into their computers (and sometimes their TVs) was, in pure layman’s terms, being converted to light for transport. Of course, eventually the LAN moved from ethernet to wireless, and much of our personal computing moved onto devices, and there also happens to be a movement afoot to take personal connectivity off the fiber and into air and space… but that doesn’t mean fiber is dead.
Optical Connectivity Is More than Fiber
So we’ve glossed over a couple of problems here with optical connectivity—the big one being how do we get from electrical signals to light, and back, while retaining the integrity of the data that we’re transmitting? Fiber cables only handle the transmission of a light signal—not the conversion of electricity to light. Keep in mind that traditional electrical transmission involves converting our information—our voice, or our data, or television images—into electricity that is broadcast over copper cables. These issues are handled by some of the other equipment used for optical connectivity, which includes:
- Tranceivers, or transmitters and receivers, which encode and decode information in electrical signals into light;
- Multiplexers and demultiplexers that allow multiple optical signals to move through one fiber; and
- Amplifiers specialized to boost an optical signal without converting it back to electrical form.
What Optical Connectivity Has to Do With AI
Optical connectivity has some advantages aside from its speed— it can’t be disrupted or harmed by powerlines or natural electromagnetical interference. Optical signals move so fast that they are difficult to intercept. It’s also become cheaper to lay fiber than it is to hang more copper cable—yet fiber is more expensive because of the switching equipment involved. It can support more bandwidth than traditional cable. It can also extend for longer distances without needing signal boosters.
So, if we’re building a distributed computing system to support AI models, and we’re using several data centers, it’s a good thing if our data centers are connected by fiber—our data is more secure and less likely to be disrupted in transit, it can move quickly and efficiently and with lower likelihood of degradation than it would if it were transmitted over copper. Within and between data centers, optical cable interconnects the clusters of AI chips and GPUs used to support the machine learning modules that train our AI models.
As the fiber infrastructure originally built for telecommunications has been adopted for AI, there’s just not enough fiber out there for the AI buildout that is being envisioned, not only on a local level to specific geographies or data centers, but also on the levels of regional, national and global backbones. AI is also placing a greater demand on optical networks to improve to even higher levels of data fidelity and resiliency. The acceleration towards more distributed computing is also putting pressure on local and global optical resources.
