Chip Startups Using Light Instead of Wires Gaining Speed and Investments

Computers using light rather than electric currents for processing, only years ago seen as research projects, are gaining traction and startups that have solved the engineering challenge of using photons in chips are getting big funding. From a report: In the latest example, Ayar Labs, a startup developing this technology called silicon photonics, said on Tuesday it had raised $130 million from investors including chip giant Nvidia. While the transistor-based silicon chip has increased computing power exponentially over past decades as transistors have reached the width of several atoms, shrinking them further is challenging. Not only is it hard to make something so miniscule, but as they get smaller, signals can bleed between them.

So, Moore's law, which said every two years the density of the transistors on a chip would double and bring down costs, is slowing, pushing the industry to seek new solutions to handle increasingly heavy artificial intelligence computing needs. According to data firm PitchBook, last year silicon photonics startups raised over $750 million, doubling from 2020. In 2016 that was about $18 million. The challenge is that many large machine-learning algorithms can use hundreds or thousands of chips for computing, and there is a bottleneck on the speed of data transmission between chips or servers using current electrical methods.

less rare earths needed ?

[Micah NC]

If this could get us out of the Chip crisis and our dependency on China, I'm all for it !

Re:

[Fly Swatter]

The second paragraph of the summary gives it away: they are talking about chip interconnects. So instead of copper circuit traces there will be optical traces between chips. So instead of simple to deposit metal it will be glass or most likely plastic to beam the light around.

Perhaps they can just line all the chips up in a row and their inputs and outputs line up with no interconnect tracing needed. But that would be too simple.

Re:

[Z80a]

That could help quite a bit with the multi die designs the industry is moving to, like the ryzen chips etc..

Re: less rare earths needed ?

[ShooterNeo]

Yeah absolutely. The light is faster if you use hollow core optical fiber and doesn't radiate etc. You could probably build AI accelerator chips/mega GPUs with 10-100 chiplets interconnected by optical fiber. They would be too expensive for home use though but would have unbeatable performance for AI inference and video rendering.

Re:

[Z80a]

It do sound to me like something that could be potentially cheaper than the current interconnect boards being used at the moment.

Re:

[HiThere]

As for pricing... that's true right now, but will it still be true in five years? There doesn't seem anything inherent in the design that requires any increased cost. The problem is that it's new, and they're still optimizing it.

Re:

[ShooterNeo]

The reason they would be too expensive is literally they would have too much top tier silicon in them. A modern consumer CPU or GPU has 1-3 chiplets and costs $800+, probably several hundred to make, while one with 100 chiplets would probably cost $100 a chiplet just in manufacturing costs, or $10,000.

Re:

[account_deleted]

Comment removed based on user account deletion

Re:

[WhackAttack]

Most silicon chips are made in Taiwan, not China.

Re:

[HiThere]

The problem with that is that China considers Taiwan to be a part of it, and is taking slow steps to reacquire it. Things like building artificial islands in the China sea, allowing it bases that are a lot closer to Taiwan. The US will object, but I don't think that over the decades this will be sustained. More practical is to relocate the fabs into areas that aren't threatened.

Re:

[WhackAttack]

Luckily new fabs that can manufacture semiconductors on par with TSMC are currently being built in the US.

Re:

[Micah NC]

Yes, largely so, but the rare earths are often said to come from China.

This Is Excellent News

[theshowmecanuck]

I was afraid we would never be able to see all the cool flashing lights when we are finally able to have some android heads to open.

Re:

[account_deleted]

Comment removed based on user account deletion

das blinkenlights

[bugs2squash]

I don't see how this speeds things up much for non-quantum type computing. Can signals really travel at a significantly higher data rate over, say, an inch of optical fiber than they can over an inch of wire ?

Re:

[Anonymous Coward]

I'm sure Dave over at EEVblog could explain it in a much more round about way, but...

Electrical signals do NOT travel at the speed of light (c). c is the fastest they *can* travel (vacuum). Through any other medium (wires, PCBs, chip layers, etc.) they are slowed down by the affects of capacitance, inductance, and resistance.
https://www.protoexpress.com/b... [protoexpress.com]

Even light doesn't travel at c through glass/plastic/air.

At current CPU frequencies (actually, since CPUs started to get north of about 500KHz) radio

Re:

[bugs2squash]

I should have been clearer, can data travel at a significantly higher rate (ie. bits/s rather than meters/s) over a short optical fiber than it can over a short piece of microstrip ? I would have though it is not a factor anyway since the light is not actually used in any computation in a conventional CPU

Re: das blinkenlights

[samwichse]

In copper, the signal travels about 1/3C. In a hollow fiber containing a vacuum, this would be 1C, with air in there at STP, it would still travel about 0.9996c. In a solid glass fiber, it would be about 2/3C.

At 4ghz, the signal travels about 2.5cm/cycle in copper. So you can imagine tripling that would be pretty significant.

Re: das blinkenlights

[Fons_de_spons]

High frequency electrical signals bleed through the isolation due to capacitive and inductive coupling to other signals and the ground plane. It becomes a big mess. Optical had less issues here. It is the interface to the electrical part that is complicated here.

Re:

[alvinrod]

As I recall the electrical current in CPUs travels at somewhere between half and two-thirds the speed of light, so if it's faster it wouldn't be by much. I'm assuming that the reason for going this route is the difficulty with continuing to shrink the technology being used now. They're already at the point where elections can start jumping across wires when we don't want them to. Either this is easier to scale down to even smaller feature sizes or it doesn't experience the same problems at those sizes that

Re:

[OrangeTide]

It's slower. Refractive index of a silicon waveguide is 3.5. A third versus two-thirds the speed of light isn't a big deal when we're talking about centimeters versus kilometers. If you can throw out capacitance, resistance, reactance, and parasitic losses typical of high speed electronic systems then photonics start to look pretty appealing. As for shrinking down your electronics, a single mode waveguide can only be so small. Perhaps one day we can have photonic logic gates that are in super-position and r

Re:

[account_deleted]

Comment removed based on user account deletion

Re:

[ceoyoyo]

Yes. The actual propagation speed is going to be faster (maybe a factor of 2) but you might lose that in your conversions. In a decent setup the light won't encounter significant resistance though, which means no heat.

The biggest problem with electrons though, is that when you accelerate them they radiate, and the harder you accelerate them, the more they radiate. So as you speed up your signalling you get more radiative losses, crosstalk, weird coupling problems of all kinds. The design work that goes into

Move along, nothing to see.

[Impy the Impiuos Imp]

and there is a bottleneck on the speed of data transmission between chips or servers using current electrical methods.

Wait, this is just transmission between chips, not inside them?

Emahgerd...

Re:

[account_deleted]

Comment removed based on user account deletion

Re:

[account_deleted]

Comment removed based on user account deletion