Hacker News Reader

Comment by amelius | Hacker News Reader

> HITOP performs about 1 trillion operations per second, whereas sophisticated Nvidia chips can chug through 300 times as much data

Wow, it sounds like optical computing is already closer than I expected.

View on HN

> But the optical chip’s efficiency is compelling. “The game here is that we lowered the energy cost 1,000 times,” Chen said.

The second part of that paragraph is even more important. The optical equivalent is actually ~3x more efficient than the electronical version. With 300x HITOPs you would only consume 300 energy instead of the 1000 energy of the Nvidia version.

Simplified, that means 3x longer battery life, or 3x less heat, or 3x lower energy bill.

> 3x longer battery life, or 3x less heat, or 3x lower energy bill.

Can we not have the same battery life, same heat, and same power bill and instead do 3 times as many things?

I see you work in the mobile phone industry

If you have a laser, you can print an optical neural network using a consumer 3D printer. The present barrier to adoption, as the article describes, is the cost of transitioning from the optical space to the electronic space.

Do you know where does this cost lie? Say for an organization with no prior spending on "legacy" electronics based hardware.

Is it the sunken cost in existing hardware, or developing the rest of the "value chain" to make photonics work, or something else altogether?

You have to miniaturize the components and be able to cram billions on a die of a few square centimeters at most. The industry has lots of experience doing that with electronics. Almost none with photonics.

The silver lining is that one might need massively less photonic components to get the job done than electronic components due to the former's theoretically magnitudes higher switching frequency. And we might reserve them for graphics and AI at first.

> If you have a laser, you can print an optical neural network using a consumer 3D printer.

I would love to look into this more. Do you have a reference that would give some instructions on how to give this a go?

You could think of a lens as a "Fourier transformer," not in the LLM sense directly but in the sense of something that executes a Fourier transform. See the excellent videos by Huygens Optics including this one: https://www.youtube.com/watch?v=Y9FZ4igNxNA entitled "Fourier optics used for optical pattern recognition."

Having established that, well, it's not hard to imagine that something that's good at running Fourier transforms is probably good for other situations where you need to run a lot of dot products in parallel. That, in turn, should sound like an awfully familiar problem...

Ok, but dot products and Fourier transforms are both linear operations. Maxwell's equations are linear (light also obeys them).

Neural nets, on the other hand, generally require nonlinear operations. As do most computer programs.

So the explanation here is either not complete, or perhaps applies to a very restricted type of neural network only.

Good point. I know there are people working on optical NNs, but I have no idea how they handle nonlinear activation functions.

There’re a handful of ways to introduce electro-optical and all-optical non-linearities into optical neural networks. Here’s a fair survey of the field: https://dl.acm.org/doi/full/10.1145/3607533

Reply on news.ycombinator.com