GPUs used to run AI require enough electricity to power a small town and generate enough heat to warm several more. However, as avid Toast readers you'll already know the most exciting potential solution: photonics.
And there’s a new breakthrough here, specifically in tensor operations (the mathematical heavy lifting behind AI). Think of tensors as a multidimensional container of information, like a box full of maths problems. Computers traditionally look at each calculation sequentially. Light can solve all of them at once.
A Finnish team encoded data into light waves then let physics do the work as the light propagates. No chips heating up, no power-hungry circuits, just light interfering with itself in mathematically useful ways. And they've now demonstrated it works for the operations AI actually needs like convolutions, with different wavelengths acting as separate computational channels.
Professor Zhipei Sun thinks this can be integrated onto photonic chips within three to five years and for an industry strangled by power bills and thermal limits, this is the sort of breakthrough that makes venture capitalists scramble for their chequebooks.
🧐 What's in it for me? AI could become dramatically faster and cheaper which could mean less money gets burnt using it, but more likely means there’ll just be a lot more AI. This could also unlock AI applications that are currently impossible due to power constraints.
💵 Out of the Lab: Data centres are haemorrhaging cash on power bills, and photonics offers an escape.
UltraCell Networks, a Leeds-based startup raising £5M, is already positioning optical architectures as the solution to AI infrastructure bottlenecks, claiming 80% power reduction and 85% latency improvements. If this single-shot tensor computing integrates with infrastructure like this, we could be in for a paradigm shift.
Lightmatter and Luminous Computing are building photonic AI chips, but this passive approach might leapfrog them both by eliminating active switching entirely.
Until next time, stay curious.
Like what you're reading? Share toast with a friend.