Less than a single clock cycle (0.1 ns) of a fast conventional processor? This feels like the difference between electromechanical computing and transistors (skipping vacuum tubes entirely). These moments in life I hoped for as a kid, when you realise there's a step change between technologies you are familiar with and what's coming next are exhilarating and worrying. The downside is that everything we've created to this point, including all the bad applications, will be magnified by the same factor as the good things. Time to get our house in order.

There are several companies doing photonic computing (some with commercially available or soon-to-be-available chips). Unfortunately, it's currently all for naught, as existing available silicon hardware accelerators can currently do the computation faster than anyone can stream the data into the chip. The more interesting problem is data movement, as usual.

I do not understand your remark.

> do the computation faster than anyone can stream the data into the chip. The more interesting problem is data movement, as usual.

Is that not the selling point of integrated photonic circuits? You got photonic chips/components that are connected by photonic waveguides. So essentially photons replace electrons for all intents and purposes and the data moves at close to the speed of light.

My point is the following:

1. Photonics is a new technology. It is more expensive than silicon 2. Silicon can do the compute as fast as photonics, and we understand silicon. Building a silicon chip is 'easy'. 3. The problem is data transport, which silicon cannot do fast enough. We can compute inferences faster than we can get the data

Photonic compute doesn't help with (3). Photonic data transport does, and there are several companies who have entered this market, including several who have already exited with large deals. However, photonic data transport can terminate at photonic compute or silicon compute, and silicon compute is fast enough and cheaper / better understood.

I guess this bodes the end of the copper PCB, except for power supplies.

I'm going to guess most electronics is not going to be super high-speed. Sort of like how fiber optic was once the realm of only large businesses, then small businesses, and now is commonly available at home (although nowhere near 100% adoption). I think the copper PCB will eventually go away, but not for many decades.

570 ps is 0.57 ns.

Also, 0.1ns would be a single clock cycle at 10GHz

"5-by-6 pixel array"

These are not exactly high res images.

Equally important: "nine neurons". This is not a deep circuit.

When we say that a CPU is running at some frequency that's essentially the speed at which a signal propagates through all the gates of its slowest pipeline stage, and the individual gates have a transition frequency an order of magnitude higher.

So if a CPU is running at 4GHz, it completes a clock cycle in 0.25ns (250ps), but that means that individual transistors must have latencies <=0.025ns (25ps). If you built a tiny little 9-neuron single-layer network out of those transistors then the network's latency would be around the single-gate latency and it would appear to be 20x better than this photonic setup, and even if those 9 neurons were laid out sequentially it would outperform this photonic chip by 2x.

Now granted, that's comparing the performance of a mature IC process using equipment that cost billions of dollars to set up, to something that was built in a university lab on a substantially smaller budget, so direct performance comparisons aren't exactly fair in that sense. But it's also misleading to take a direct hardware implementation of an itty bitty neural network and handwave away the very real issues with scaling such an approach.

Yes, this is seriously blurring the line between image recognition and a bandpass filter. There doesn't seem to be anything in the way of actual computation going on until the step where the results are read.

Once a proof-of-concept exists, scaling it up is conceivable. Computers used to have kilobytes of RAM, now some have terabytes. Digital cameras used to have sub-megapixel resolution, now some are measured in gigapixels.

“1 pixel should be enough for everyone”

Most zoom calls are fine with 0 thank you very much

It could! https://opg.optica.org/oe/fulltext.cfm?uri=oe-28-19-28190&id...

Mmm...but scaling up is not the same in photonics as it is in digital. Everyone thinks of scaling in terms of digital, like Alan Turing, von Neumann, and later Moore talked about, if you have an amplifier you can amplify it with another amplifier, so as long as it's totally digital the signals can be manipulated at arbitrarily small scales. Not true for analog or photons or quantum, there's weird caveats and they're not well researched, this scaling works only for digital.

Sure, our existing understanding and models for scale may not directly apply.

That doesn’t limit the creation of new models and new understanding. Just because scaling might be different doesn’t mean it’s impossible.

Our existing models did not exist at some point in the past.

You can make digital optical logic gates - https://www.youtube.com/watch?v=pS1zAAD1nXI

Looks like a "modern version" of the numa-rete, developed by Paul Weston at Heinz von Foersters Biological Computer Laboratory (BCL): https://distributedmuseum.illinois.edu/exhibit/biological_co...

I'm really excited by the idea of using analog computing for AI, and the advent of analog chips. Anyone know if there's a cloud provider offering hourly access to analog compute?

As an aside, I recently learned of https://lightmatter.co/, a photonic chip manufacturer

The problem with photonic chips is that the devices are enormous in comparison. Something simple like routing the light around a 90 degree corner already takes in the order of a millimeter. Simply scaling down the waveguide causes the light to fly out of the corner, causing reduced signal strength and random scattering through the chip, heating it up. A chip with more than a few hundred devices will probably quite large already and will cause low yield, high cost and high power consumption (it likely requires cooling). I don't see optical chips replacing transistor based chips in my lifetime, except for a few specialized applications (high frequency radio maybe).

Are the network weights hardcoded (like a ROM), or is the network programmable?

The article implies that the weights are programmable, but the overall shape is fixed.

Isnt this a year old? https://arxiv.org/abs/2106.11747

Now make it able to work directly on visible light or at least the output of an image intensifier and you can make some pretty fast killer drones.

One could do preprocessing like edge detection with this, like happens in biological retinas. Then a lot less data needs to be processed down the line.

We know Quantum Computing is more computationally powerful than classical digital computing. But I've always been curious if there's a middleground for non-quantum analog computing - does anyone know of theoretical work here? Could we potentially build a non-quantum analog computer that's more computationally powerful than a digital computer?

It's a big area of research in AI.

Veritasium has two great videos on it:

https://www.youtube.com/watch?v=GVsUOuSjvcg

https://www.youtube.com/watch?v=IgF3OX8nT0w

I found a hackernews discussion for the second one: https://news.ycombinator.com/item?id=29645610

For a somewhat related topic (analog image recognition), for those interested, I can highly recommend this video: https://www.youtube.com/watch?v=Y9FZ4igNxNA

Not sure how image recognition rates (frames/sec?) translate into meters/sec.