AI At The Speed Of Light
Lightelligence is a small team of 24 veteran researchers that aims to introduce light-based technology into the world of AI.
Though being a fairly small team, their capabilities are not to be underestimated, they all have degrees from the top universities, 12 PhDs collectively and over 210 hours of industry experience all-together.
According to Venturebeats, Lightelligence claims they have built a prototype chip that utilises the higher speed, lower latency and increased power efficiency to accelerate the computations required for artificial intelligence algorithms. They claim the chip improves latency by "10,000 times". This may seem like a huge deal at first but latency isn't the same as speed, this just means the operations sent to the accelerator have virtually no delay. The team also claims the chip has power consumption "orders of magnitude" lower than other silicon-based chips. Now here is the one truly important thing, compared to other similar-sized silicon-based AI accelerators, matrix-multiplication and other linear operations which form the backbone of AI operated roughly 100x times faster while crunching the MNIST database, a database filled with 70,000 images of hand-written digits.
But what prevented this innovation from happening before, why not earlier when light-based computation was a hot new thing? Well, the key difference is that the team here tried to ignore previous conceptions of computer architecture, which traditionally don't work well with light due to quantum effects, but instead tried to work with the laws of physics that govern light to make a system that's more efficient to light and then connect it separately to electronics. Due to this rethinking of how we should go about building light-based computers that led to this prototype, the prototype performs 100,000x faster in matrix multiplication compared to previous light-based models.
In the future, the team aims to make the system more stable and ensure its compatibility with popular AI frameworks such as Tensorflow, PyTorch and others. If this system were to be scaled up, then this computer might have as much importance as the first electrical computers that were created, which have all gone down in history as the ushering of a new era.