Let there be light! DARPA seeking physics-defying photonic computers to supercharge AI It's no lightweight matter. DARPA is putting about $35 million in total funding on the table in the hope that it will spur researchers to work around fundamental physical constraints and build much larger-scale photonic circuits that do more of the computing with light, not electronics. A recent solicitation from the …
There has long been a fair amount of dot-com moolah hosed onto photonics. It is proving a very expensive and time-consuming nut to crack. Sure the promise is every nark's dream, but what is a governmental watering-can going to achieve? A line on the CV and promotion to Departmental Overlord, perhaps?
I keep seeing this “we need more power” , we will be out of power by X” - build more power plants.
Has anyone done the math that includes Mores law?
It is starting to look like a money grab by all the FUD out there.
I just know 5 years ago, a rack that was 5kw, Today I can do 30 times the work on 175w.
Well I expect the VAX was bipolar, maybe ECL based logic. Very hungry power wise, but all there was back then. I actually did a bipolar chip in the 80's. 5W, and all it did was multiply. CMOS really started on a tear in the 80's and as the gate sizes dropped performance got better. I'm not sure there is much left to eek out of CMOS though. Gates are atoms thick now, so once you get to maybe 2 thick, how much lower can you go. I have a suspicion Moore's law is at an end, and unfortunately, the end of moore for less in the electronics world.
My initial reaction to reading that bit was similar to yours, but then I thought maybe they mentioned AI to get more hedge-funded money pit startups involved.
After all, if all these people are sinking their cash into the AI bubble, at least try and redirect some of it to good use!
Of course, cynically, it could just be someone at DARPA who has also drunk the cool-aid.
Everything they say about the problem is wrong.
… The millisecond latency of electronic circuits …
Today's optical modulators run at over 40GHz, so even with EQ, deserialisation and buffering, well under a nanosecond of latency.
... Unlike metal-oxide-semiconductor circuits, which regenerate signals and filter out noise, photonic circuits have a fundamental issue of optical attenuation and noise that can't be amplified out...
Exactly the opposite is true:
electronics adds noise, and in-band noise cannot be “amplified out” by any means. SNR is a ratio, and only gets worse with amplification.
Out-of-band noise can be filtered, yes, and simple optical-coating methods can achieve “impossible” 20th-order filtering – way beyond what silicon can do.
... as any attempt to amplify optical signals also amplifies the noise...
Optical noise-free amplification is entirely possible - and routinely used in Erbium-doped fibre amplifiers.
I only dabble in this subject, but my understanding is that the primary limitation of optical computers is fan-out. Optical switches require relatively significant power to function. After about three levels, the power levels are too low to activate the switches reliably. None of this was mentioned in the OP.
Is the 'photonic neural network' where as components switch they emit flashes of light, related to this?
From the article. it appears that light not only affects the components, but makes oscillation patterns appear that look
a lot like spiking oscillations in biological neurons.
https://www.nature.com/articles/s42005-025-02000-9
looked into this a while back, VCSELs can indeed be integrated though at the moment the efficiency isn't great.
An early use of these was in wireless optical mice as the light source is highly efficient (0.9mA) compared to the 30-40mA
of a regular red LED yet has comparable light output.
