News Computing at the Speed of Light: Lightmatter Mars SoC Bends Light to Process Data

[Admin]

Lightmatter revealed the details of its revolutionary silicon photonics Mars SoC that uses lasers for computation.

Computing at the Speed of Light: Lightmatter Mars SoC Bends Light to Process Data : Read more

 

[Endymio]

While "computing at the speed of light" has a jazzy ring to it, the fact is that electrical impulses move at roughly the same speed as light anyway. The purported advantages of this chip appear to come mostly from the fact that, computationally, this seems to be essentially an analog tensor core, rather than digital. Nice, if you're building an AI cluster ... but I don't see this replacing a general-purpose CPU any time soon.

 

[everettfsargent]

"In fact, adding more photonic compute units to the array increases efficiency - you get quadratic performance scaling compared to power consumption increases. In other words, adding a unit will give you four times more performance compared to the amount of additional power consumed. "

... should be ...

"In other words, adding a unit will give you THREE times MORE performance compared to the amount of ADDITIONAL power consumed. "

!00% is already baked into 1X and 400% is baked into 2X (quadratic), therefore reword that sentence because 400% - 100% = 300% MORE performance compared to the amount of ADDITIONAL power consumed, 2X -1X =1X.

You all can't even do 2-bit calculations correctly!

Oh and I can't wait for my 16-bit tensor overlords sometime after 2100 CE. Unicorn hardware.

 

[Endymio]

... should be ...

"In other words, adding a unit will give you THREE times MORE performance

Actually, it's correct if the word "more" was simply meant to be "the". As in "four times THE performance".

 

[everettfsargent]

Actually, it's correct if the word "more" was simply meant to be "the". As in "four times THE performance".

Still doesn't work!

"4X THE performance compared to the amount of ADDITIONAL power consumed, 2X -1X =1X. "

Double the power quadruple the performance, now that makes perfect sense.

 

[recycledelectrons]

the fact is that electrical impulses move at roughly the same speed as light anyway.

I disagree.

Real electrical wires are not ideal conductors or even ideal resistors. They also have some capacitance. That makes the circuit an RC circuit, also called a RC delay circuit. It takes time to charge the capacitor.

 

[OpticsinComputing]

The numbers presented were worse than what can be achieved with conventional digital electronics at equivalent bit depth.

 

[Chung Leong]

Nice, if you're building an AI cluster ... but I don't see this replacing a general-purpose CPU any time soon.

I imagine the tech can be used to massively boost ray-tracing performance.

 

[Endymio]

I disagree. Real electrical wires are not ideal conductors or even ideal resistors.

And? You don't think this photonic chip is an ideal vacuum either, do you? Assuming its effective IOR is similar to a fiberoptic thread, then signals and both it and an electrical wire are going to move at approximately the same speed: 0.6-0.7c.

 

[margrave]

While "computing at the speed of light" has a jazzy ring to it, the fact is that electrical impulses move at roughly the same speed as light anyway. The purported advantages of this chip appear to come mostly from the fact that, computationally, this seems to be essentially an analog tensor core, rather than digital. Nice, if you're building an AI cluster ... but I don't see this replacing a general-purpose CPU any time soon.

If "roughly" means around 60%, then sure. Electrical signals generally move around 60%-ish of light speed. But people here have been picky about 4x more vs. 3x more. So i'll be picky by saying 60% is not "roughly the same as".

 

[grimfox]

I wonder if there is a security benefit given the "alien" nature of the light processor. Could you use something like this to create a sort of firewall between a secure storage (say for biometric data) and the host system?

I also see some benefits in networking. Imagine having your long haul fiber drop directly into a processor and then get distributed to sub fiber networks without having to convert down to electrical signals and back up to light again at each node. Could shave a few MS of latency off your transcontinental or transoceanic transmissions.

Similarly, would it be more immune to radiation? Space probes with GHz processors or satellites that could survive in close orbit around Jupiter and Saturn for long periods.

Seems like there could be some potential there. Hopefully they can keep developing the tech and unlock some of it.

 

[Endymio]

So i'll be picky by saying 60% is not "roughly the same as".

You're missing the fact that "light" doesn't move at the speed of light, either. Not in a fiberoptic cable or photonics device ... or anything else other than a pure vacuum. The actual velocity depends on the IOR of the transmissive medium.

Imagine having your long haul fiber drop directly into a processor and then get distributed to sub fiber networks without having to convert down to electrical signals and back

That's actually the primary application for photonics chips, at least for the last decade or so.

 

[shady28]

While "computing at the speed of light" has a jazzy ring to it, the fact is that electrical impulses move at roughly the same speed as light anyway. The purported advantages of this chip appear to come mostly from the fact that, computationally, this seems to be essentially an analog tensor core, rather than digital. Nice, if you're building an AI cluster ... but I don't see this replacing a general-purpose CPU any time soon.

That's largely incorrect, as it is affected by the medium.
Also conventional transistors require time to overcome impedance, ie they require a capacative charge to build up and up before a transistor can 'switch' which results in more latency.

https://inst.eecs.berkeley.edu/~ee42/fa01/LectNotes/42_24.pdf

 

[Endymio]

That's largely incorrect, as it is affected by the medium.

Which is what I said. Both electrical and electromagnetic (i.e. light) signals are affected by their propagation medium, and both have (in the mediums in question) roughly equal velocities.

Also conventional transistors require time to overcome impedance, ie they require a capacative charge to build up and up before a transistor can 'switch' which results in more latency.

And photonics lambda switching and mems waveguides also add latency due to switching time. The point simply was that the "speed" of light isn't the factor conveying the advantage here.

 

[frogr]

Can it play Crysis?