
AI has a huge problem: Training them uses far too much energy. Even Sam Altman has warned that the next wave of AI systems will consume vastly more power than expected, and our current energy systems will struggle to cope. As such, AI development may hit a ceiling and stagnate. This wouldn’t be a problem if your current AIs were advanced enough, but they still can’t safely replicate tasks we humans find relatively easy, like driving. So, to unlock our AI future, we either need a clean energy revolution to provide the copious amount of energy required, which is highly unlikely, or we need a revolution in computing to drastically reduce AI power consumption. Well, researchers have done just that and developed an AI-optimised photonic chip that could solve this entire conundrum.
Let’s start at the beginning. What is a photonic chip?
Well, your computer and phone use electronic chips. These are comprised of millions of minuscule switches that can turn each other on or off. They work together to perform calculations and run code. Photonic chips do the same thing, except they use light rather than electricity. In a photonic chip, multiple light sources are injected into a transparent material and interact with each other through waveguides, lasers, polarisers, and phase shifters, then exit the chip. If the light destructively interacts, in which the light waves are out of sync and cancel each other out, it can be read as if a switch is off. If the light constructively interacts, where the light waves are in sync, and the brightness increases, it can be read as a switch is on. This way, photonic chips can run the same code as a regular electronic chip.
However, photonic chips can also work differently than standard chips. You can design them to run multiple calculations inside one “unit.” This way, a single photonic unit might have many inputs and many outputs, and when light passes through, it runs multiple calculations in a single go. Not only that, but it runs them at light speed! As such, they can compute at a far faster speed than electronic chips in certain circumstances. Another advantage is that photonics require way less energy to run than electronic chips and are far more efficient, as they don’t incur massive amounts of thermal loss. What’s more, as the complex computation is handled by light, not hard-to-manufacture microscopic switches, they have the potential to be cheaper, too.
In other words, they are one of the few ways we can break past the current constraints of computing technology.
So, why don’t we all use photonic chips then? Well, they are currently insanely difficult to manufacture. Minor tolerance issues can easily render a photonic chip useless. As such, we need an easy, efficient and reliable way to manufacture complex photonic chips if they are to become a widely used technology.
This is where these researchers come in. Rather than using waveguides, lasers, polarisers, and phase shifters to manipulate the light, they use wavey silicon. Silicon is a transparent material, making it an ideal medium for this application. The researchers used a slice that is only around 150 nanometres thick, making it narrower than the wavelength of the light passing through it. This means that they can control the propagation of the light through the silicon by varying its thickness. As such, they can design a topographical shape for the silicon that will create the light interactions that power a vast array of different computations. Moreover, we are already extremely good at manufacturing silicon to this level of accuracy, as regular electronic chips use silicon too. This means that, in theory, these photonic chips could be cheap to manufacture, easy to produce at scale, and, therefore, could be widely used.
But these researchers weren’t going to stop there! They used this new-found technology to design and test a photonic chip that can perform vector-matrix multiplication in a single light pulse. This mathematical operation is a huge part of AI training, as it powers the neural networks behind AIs.
In other words, these researchers may have just invented the perfect AI chip that could be mass-produced and train AIs using far less energy and in a fraction of the time. As such, this technology could be the breakthrough we need to enable AI models that are orders of magnitude larger than we have today and exponentially more advanced and reliable without breaking our energy grids. For now, this technology is still just a lab experiment and far from ready to be used in the digital wild, so don’t expect this revolution to come any time soon. But it is coming, and we have to brace ourselves for when it does.
Thanks for reading! Content like this doesn’t happen without your support. So, if you want to see more like this, don’t forget to Subscribe and follow me on BlueSky or X and help get the word out by hitting the share button below.
Sources: Nature, The Brighter Side, Synopsys, Planet Earth & Beyond