A one hundred thousand-fold enhancement in the nonlinearity of silicon | Hacker News Reader

A one hundred thousand-fold enhancement in the nonlinearity of silicon

phys.org

80 pointsbookofjoe29 comments

> The block can cool with relaxation times measured in nanoseconds.

At what density?

I mean sure if you have a single device in isolation it might be able to cool in a few nanoseconds. What matters is the density at which you can have a checkerboard of these things toggling in opposite directions without neighbors inducing failures. Otherwise we're simply trading the propagation time of electrons for the propagation time of heat rather than the propagation time of light.

Light may move fast, but heat doesn't.

Edit: also, I'm skeptical about power efficiency if the principle of operation for this thing is fundamentally based on turning free energy into heat. Generally heat is a waste product of the switching event, leaving hope that future generations can continue to reduce that waste. Here, the waste product is what makes it work, meaning that it's probably very inefficient (joules per switching event) and unlikely to improve much. Current mode logic is wicked fast, but never caught on (except for I/O drivers) because fundamentally it works by burning up energy into heat -- it's crazy inefficient and hasn't improved after 20ish years.

https://www.nature.com/articles/s41467-020-17846-6 is the actual paper that has much better info and appears to be public access for now.

Silicon photonics has a lot of potential, but not for replacing transistors. I wish researchers would keep their claims of potential impact in the realm of possibility.

Why not? Could you develop why it could not replace transistors.

Read the article: photons don't interact with each other. Electrons do.

Photons are great for communicating. We still don't know how to use them to do computing without first transferring back into the electrical domain.

Edit: in a certain sense, this work hasn't solved that problem, it' just transfers to the thermal domain instead of to the electrical domain.

> We still don't know how to use them to do computing without first transferring back into the electrical domain

This isn't entirely true.

We do know in a general way how to perform computation using photons and non-linear optics.

Electrons are involved as they are an essential part of a non-linear optical medium. But the information being processed is not converted to the electrical domain, and there's no electricity involved. The electrons remain bound to their atoms, not mobile like current charge carriers.

> Electrons are involved as they are an essential part of a non-linear optical medium. But the information being processed is not converted to the electrical domain

That's a matter of hair-splitting. More importantly, it hasn't been used to make practical gates.

"Nonlinear optics" has had plenty of important successes in (electrically controlled) amplification and frequency shifting, and those victories make our lives better every day in the way they've improved fiber optics, especially CWDM.

However all-optical switching is still pretty much vaporware. I think it's really disengenuous to lump optical switching in with the broader field of nonlinear optics -- this basically just hides the glaring lack of progress in all-optical switching among the victories in neighboring fields. So we get a lot of advocacy of the form "hey you should believe all-optical switching will happen some day because this neigboring field that doesn't make gates is under the same umbrella as us!"

The range of light wavelengths that can reasonably be used for computation is in the hundreds (or thousands) of nanometers. Even the most optimistic approaches for confining such light to small areas, plasmonics, can only realistically give you something like a 10-fold size reduction, putting you at tens of nanometers. The transistors and wires in advanced process nodes have been smaller than that for quite some time, now.

Moore's Law is about density. Photon gates cannot compete with electron gates in terms of density.

Never, ever believe in dense all-optical computers. The deBroglie wavelength of light is huge. There could be niche applications, but their markets are microscopic.

For X-Rays? This technology is sure decades away from commercialization, but it could be the way forward in a time when conventional silicon has reached its physical limits.

I wonder why people always think just one step ahead. Science is about incremental advances EVERYWHERE to eventually allow for something completely new. If everyone had your attitude we would still beat each other with wooden rods and use candles to light our houses, because you know... The de Broglie wavelength of light is just tooooo large.

When we have cheap, manufacturable coherent X-ray sources, I will reconsider.

I think some people don’t get that a nonlinear interferometer IS a switching element, its use is AS a logic element. In the late 70’s, Toni Bischofberger and YR Shen put a liquid crystal inside an interferometer and observed switching effects using a single shot ruby laser. The theory had previously been worked out, and there were initial experiments at Bell Labs. Shen et al used liquid crystal because they could vary the nonlinearity greatly by temperature: they were exploring that entire parametric space. Cut to 2020. So they reduced the size to the 100 micron level, so what? Electron gates are at the 10nm level in production, and experimentally, at the few atoms level. While sub diffraction optical effects have been reported, one end of the wire is still a few microns in size. So perhaps someone can find an application for these Si gates, but not in switching logic.

That's a great point, I hadn't thought of it that way before. (Well, at least for practical light wavelengths in the visible or near IR/UV).

What if we could use electron energy levels in a molecule to mediate a light-based gate? (Sorry in advance for imprecise terminology...)

Lets say we have a molecule with two energetic orbital states. An electron in the Ground state (G) can be lifted to Orbital 1 (O1) by photon with wavelength A, and can be lifted from there to Orbital 2 (O2) by a different photon of wavelength B. Photon B alone is too energetic to lift from G->O1, and not energetic enough to lift directly from G->O2, where photon A is exactly enough to lift from G->O1. So the gate would be set by photon A raising the orbital from G->O1, and the gate is read by sending photon B which is only absorbed if A has already been absorbed. Compose multiple such gates by designing a pair (or more) of molecules where the wavelengths of A & B are swapped, so that the first molecule's B photon is the second molecule's A photon, etc. Maybe recover energy from excited orbitals by passing a laser of each wavelength through all the molecules at the end of each 'cycle', inducing the molecule to emit any absorbed photons in a useful direction to be recycled.

Thoughts?

> The current work allows for optical switches that take up much less space than previous attempts. This advance opens the way for direct on-chip integration as well as super-resolution imaging.

Nanoseconds are still infinitely slower than "instant" that you can get with destructive interference. That thing is no faster than a modern fet.

Why the pessimism? This is a massive incremental leap that seems likely to be interated upon. Before this paper I thought there was hardly a chance for photonic circuits using silicon - now it looks rather achievable! Do you foresee some sort of nanoscale constraint?

The thing is we already have constructive/destructive interference as a foundation for optical logic.

The second laser source in their scheme is what is driven electronically. Turning a laser on, and off, on nanoseconds scale, is not that easy, or smart thing to do.

So, why would you do a CW modulation of the laser to modulate another laser, if you can modulate that second laser directly?

MZI modulators are extremely simple, small in comparison to just any existing optical computing devices, and, most importantly, already on the way to mass manufacturing for use in a single beam 100G ethernet.

With the correct materials, in theory you could slow down light enough to control its flow through a circuit with a series of nano reflective materials and its overall carrying capacity for speed of information would far outmatch electrons which require a medium to travel through which ultimately renders it many times slower than the speed of light depending on the properties of the material. Even super conductors have a far greater resistance for an electron than a room full of air would on a the flow of a photon. If I had to venture the guess I would say we mastered the electron as we could control and produce it very easily, we have just begin to explore the exotic nature of photons and the potential impacts on technology they may hold

> With the correct materials, in theory you could slow down light enough to control its flow through a circuit with a series of nano reflective materials and its overall carrying capacity for speed of information would far outmatch electrons which require a medium to travel through which ultimately renders it many times slower than the speed of light depending on the properties of the material.

That's an extremely load-bearing "in theory".

You modulate laser A with laser B for the same reason, on a chip transistor, you modulate current A with current B. The end goal is to power a series of switches/circuits from a small number of powered lasers external to the chip.

Yeah, but why if you can have a single laser for the entire device?

The first cars were slower than horses however that is no longer the case. The prototype is not the point, it’s what can be achieved from developing the technology. It will eliminate bottlenecks in places too tiny to place a modern fet

Green light is 500nm. We are already at 5nm with current commercial electronic circuits. In integration light will never beat the electrons. It will never be faster. You won't put more gates on chip.

> We are already at 5nm

Not that it matters a lot but this is marketing, there is nothing physically measuring 5nm on a die

But what about other aspects, such as for example radiation hardening for electronic devices used in space ? Would this make them more or less susceptible to radiation ?

Do we need to use visible light? Often when speaking of light people mean the entire electromagnetic spectrum. X-rays can be 10 picometers, but would that be too much energy?

Destructive interference may be instantaneous, but a measurement of such certainly isn't. In the low luminance shot noise limited regime it might be significant.

This can be used as a low pass filter against shot noise, but so can be a purely electronic device, which will do it without an extra laser, and do it better.

As for amplification role, 2 nanoseconds would too be too bad in comparison to existing optical amplifier, for which bandwidth is practically unlimited (many terrahertz.)

As for its utility as a more compact alternative to optical amplifiers for use cases where you don't need much bandwidth, you still have that second laser input, instead of which you could've put a still much smaller electronic amplifier, which would still be needed for the final optical->electrical conversion.

Whenever I hear such breakthroughs, I always say wake me when a product is made from this

Reply on news.ycombinator.com