I stopped reading after "Soon, you will not be able to afford your computer. Consumer GPUs are already prohibitively expensive."

> Critically, this processor achieves accuracies approaching those of conventional 32-bit floating-point digital systems “out-of-the-box,” without relying on advanced methods such as fine-tuning or quantization-aware training.

Hmm... what? So it is not accurate?

This paradigm for computing was already covered three years ago by Veratasium in https://www.youtube.com/watch?v=GVsUOuSjvcg.

Maybe not the specific photonic system that they are describing. Which I'm sure has some significant improvements over what existed then. But the idea of using analog approximations of existing neural net AI models, to allow us to run AI models far more cheaply, with far less energy.

Whether or not this system is the one that wins out, I'm very sure that AI run on an analog system will have a very important role to play in the future. It will allow technologies like guiding autonomous robots with AI models running on hardware inside of the robot.

Weirdly complex to read yet light on key technical details. My TLDR (as an old clueless electronics engineer) was the compute part is photonic/analog, lasers and waveguides, yet we still require 50 billion transistors performing the (I guess non-compute) parts such as ADC, I/O, memory etc. The bottom line is 65 TOPS for <80W - The processing (optical) part consuming 1.65W and the 'helper electronics' consuming the rest so scaling the (optical) processing should not have the thermal bottlenecks of a solely transistor based processor. Also parallelism of the optical part though using different wavelengths of light as threads may be possible. Nothing about problems, costs, or can the helper electronics eventually use photonics.

I remember a TV Program in the UK from the 70's (tomorrows world I think) that talked about this so I am guessing silicon was just more cost effective until now. Still taking it at face value I would say it is quite an exciting technology.

In 25 years we'll have #GlassModels. A "chip", which is a passive device (just a complex lens) made only of glass or graphene, which can do an "AI Inference" simply by shining the "input tokens" thru it. (i.e. arrays of photons). In other words, the "numeric value" at one MLP "neuron input" will be the amplitude of the light (number of simultaneous photons).

All addition, multiplication, and tanh functions will be done by photon superposition/interference effects, and it will consume zero power (since it's only a complex "lens").

It will probably do parallel computations where each photon frequency range will not interfere with other ranges, allowing multiple "inferences" to be "Shining Thru" simultaneously.

This design will completely solve the energy crisis and each inference will take the same time as it takes light to travel a centimeter. i.e. essentially instantaneous.

Interesting. Questions, the Nature paper being expensively paywalled:

- Is the analog computation actually done with light? What's the actual compute element like? Do they have an analog photonic multiplier? Those exist, and have been scaling up for a while.[1] The announcement isn't clear on how much compute is photonic. There are still a lot of digital components involved. Is it worth it to go D/A, generate light, do some photonic operations, go A/D, and put the bits back into memory? That's been the classic problem with photonic computing. Memory is really hard, and without memory, pretty soon you have to go back to a domain where you can store results. Pure photonic systems do exist, such as fiber optic cable amplifiers, but they are memoryless.

- If all this works, is loss of repeatability going to be a problem?

[1] https://ieeexplore.ieee.org/document/10484797