• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

NVIDIA is Preparing Co-Packaged Photonics for NVLink

AleksandarK

AleksandarK

News Editor
Staff member
Joined
Aug 19, 2017
Messages
3,012 (1.07/day)
During its GPU Technology Conference (GTC) in China, Mr. Bill Dally—NVIDIA's chief scientist and SVP of research—has presented many interesting things about how the company plans to push the future of HPC, AI, graphics, healthcare, and edge computing. Mr. Dally has presented NVIDIA's research efforts and what is the future vision for its products. Among one of the most interesting things presented was a plan to ditch the standard electrical data transfer and use the speed of light to scale and advance node communication. The new technology utilizing optical data transfer is supposed to bring the power required to transfer by a significant amount.

The proposed plan by the company is to use an optical NVLink equivalent. While the current NVLink 2.0 chip uses eight pico Joules per bit (8 pJ/b) and can send signals only to 0.3 meters without any repeaters, the optical replacement is capable of sending data anywhere from 20 to 100 meters while consuming half the power (4 pJ/b). NVIDIA has conceptualized a system with four GPUs in a tray, all of which are connected by light. To power such a setup, there are lasers that produce 8-10 wavelengths. These wavelengths are modulated onto this at a speed of 25 Gbit/s per wavelength, using ring resonators. On the receiving side, ring photodetectors are used to pick up the wavelength and send it to the photodetector. This technique ensures fast data transfer capable of long distances.



And that is not even all of it. To integrate everything, NVIDIA needs to combine it with chip packaging technology. Pictured below is a setup where the GPU and memory combination sits on a silicon interposer, which itself is placed on top of an organic package. This organic package is connected to the photonic integrated circuit (PIC). Further, there is an electrical interface chip (EIC) that converts the electrical signals coming from the GPU, using resonators, to optical signals. Optical interconnect needs a similarly built switch on its side.

Using these techniques, NVIDIA says that in the future there could be systems with 1000s of interconnected GPUs, each of which talks to another using a speed of light. It is only a matter of time before the technology reaches the market, however, there are no guarantees as this is only a research project for now.


View at TechPowerUp Main Site
 
can send signals only to 0.3 meters without any repeaters
Click to expand...

I have wireless monitors in mind.
 
Interesting audio's had optical for awhile I don't understand why we haven't seen optical transfer from GPU to display at this point.
 
Amazing news , great to see Nvidia pushing relentlessly the barriers !

DeathtoGnomes said:
I have wireless monitors in mind.
Click to expand...

What ? You need to read the article again none of these technologies ( NVlink 2.0 / Photonic NVlink ) is wireless .

InVasMani said:
Interesting audio's had optical for awhile I don't understand why we haven't seen optical transfer from GPU to display at this point.
Click to expand...

Because the communication speed and bandwidth requiremements are not high enough to make it worth going optical over traditional connection . It would simply add cost for 0 benefits .
 
Last edited:
Gundem said:
Same
Click to expand...
I would rather see USBc single cable connections for power, display and accessories than optical and power being two.
 
Gundem said:
Same
Click to expand...

Optical for audio is being replaced by copper, anyways. TOSLINK uses plastic fibers, iirc, and those are more resistant to kinks and breaking. The downside is it's basically maxed out in terms of bandwidth, afaik (sub-1Gbps, which makes it unusable even for HD resolutions, nevermind 1080p, 1440p, or 4k).

The optical fiber usually utilized in networking is usually glass and a bit more fragile. Also transceivers for those super high speed comms are usually quite a bit more expensive. Only recently have 10GbE SFP+s come down in price, and the max 16 Gbps is barely going to cut it for 4k60HDR... that is if there is no overhead at all.

Photonics for off-chip communications probably are the future, but they've just been much more expensive and difficult to implement vs new advancements in metal interconnects. There will come a tipping point, assuming metal cannot scale any further.

There is also the problem of existing infrastructure (something which limits replacing HDMI or DP). Just look at the number of people pushing and clawing for 10GbE, instead of fiber SFP+.

Steevo said:
I would rather see USBc single cable connections for power, display and accessories than optical and power being two.
Click to expand...
Ideally (IMO), we'd see the original Thunderbolt. Lightpeak, which only Sony used (though was codeveloped by Intel and Sony).

Copper for power, optical for data. Better to use the more ideal mediums for each respective task.

That being said, you'll probably get your wish if DP-VESA ever gets a foothold in the consumer TV market. HDMI is limited by their adherence to their connector form factor. DP just ditched it for DP2.0, which uses the USB-C connector for it's highest bitrates. The older DP plugs will still chug along at lower bitrates, but they are no longer the future.
 
jeremyshaw said:
Optical for audio is being replaced by copper, anyways. TOSLINK uses plastic fibers, iirc, and those are more resistant to kinks and breaking. The downside is it's basically maxed out in terms of bandwidth, afaik (sub-1Gbps, which makes it unusable even for HD resolutions, nevermind 1080p, 1440p, or 4k).

The optical fiber usually utilized in networking is usually glass and a bit more fragile. Also transceivers for those super high speed comms are usually quite a bit more expensive. Only recently have 10GbE SFP+s come down in price, and the max 16 Gbps is barely going to cut it for 4k60HDR... that is if there is no overhead at all.

Photonics for off-chip communications probably are the future, but they've just been much more expensive and difficult to implement vs new advancements in metal interconnects. There will come a tipping point, assuming metal cannot scale any further.

There is also the problem of existing infrastructure (something which limits replacing HDMI or DP). Just look at the number of people pushing and clawing for 10GbE, instead of fiber SFP+.


Ideally (IMO), we'd see the original Thunderbolt. Lightpeak, which only Sony used (though was codeveloped by Intel and Sony).

Copper for power, optical for data. Better to use the more ideal mediums for each respective task.

That being said, you'll probably get your wish if DP-VESA ever gets a foothold in the consumer TV market. HDMI is limited by their adherence to their connector form factor. DP just ditched it for DP2.0, which uses the USB-C connector for it's highest bitrates. The older DP plugs will still chug along at lower bitrates, but they are no longer the future.
Click to expand...
They should honestly have more all in one PC's where the guts are built right into the display. That would shave a lot of distance to the display right there and make a hell of a lot more things more reasonably feasible consequently. Time for that bulky CRT style display to become new again like everything old.
 
I'll take two...if the licensing is cheap ;) . So does this eventually lead to consumer cards with some kinda special ability to connect to each other to boost performance... wouldn't that be interesting? Wonder what it might be called?
 
siki said:
Why are you so exited? I dont fully understand what it all means, dumb it down please.
Click to expand...

I am not *that* excited.

Light (photons) can be used for data transfer, you need emitter and receiver. That's it. It's very fast and advanced, if built well.

Idea is, also, ages old - because it's relatively simple idea to come by. Implementation is the question. Would I like to see optronic of photonic computer or whatever you like to call them? Sure! Would be great!

This is, however, something like optical cable between GPUs (they don't go much in depth, so this is just simplification) instead of copper one.

It will mean literally nothing to me (and 99% people here), as we have only one GPU in computer. For super-computers, it will probably mean something-to-lot, depending on implementation.

No, Jensen didn't discover light, similar as he didn't discover 50-year old ray-tracing...

To be fair, the more one technology is used, the faster it will advance. But we're probably many years away from optical computing, which would be super-great.
 
siki said:
Why are you so exited? I dont fully understand what it all means, dumb it down please.
Click to expand...

They've gone to plaid.
 
RH92 said:
What ? You need to read the article again none of these technologies ( NVlink 2.0 / Photonic NVlink ) is wireless .
Click to expand...
While I do try to be funny in one way or another, sometimes you have to read between the lines, optical is in a way wireless, as in not actual <insert material ( copper, alum, etc.) > wire.

Fluffmeister said:
They've gone to plaid.
Click to expand...
I can't stop, its too dangerous.
 
This actually sounds like a very small step into what could turn out to be an evolution in computing, away from electrons and towards full blown photonic computing. Perhaps Nvidia is testing the waters here. I'd also expect things from Intel and AMD in the coming years.

In addition to the speed and power advantages, I think it could also represent some interesting prospects in the mobile market. I mean, let's say there is the possibility that ultra low power PC's could run on solar power. Wouldn't it be interesting if you could skip the whole inefficient conversion process to electrons and directly feed your PC photons? Could be pretty neat.
 
DeathtoGnomes said:
While I do try to be funny in one way or another, sometimes you have to read between the lines, optical is in a way wireless, as in not actual <insert material ( copper, alum, etc.) > wire.
Click to expand...

No issues with trying to be funny but it's pretty hard to judge from your text . More importantly no optical ( as depicted in the article ) is in no way wireless , indeed there isn't any cope/alum/etc wire but instead there is a physical hollow cylinder ( read wire ) made out glass fibres which conducts the optical signal ( read light ) . Without this PHYSICAL glass wire the communication between the transmitter and the receiver would be impossible .


Now wireless optical communication exists for very short distances or/and space and military application but since it requires powerfull laser emitters ( for relatively long distance ) etc you can imagine that this is very far away from making it into your living room from an economical and even physical perspective .
 
Last edited:
RH92 said:
Amazing news , great to see Nvidia pushing relentlessly the barriers !



What ? You need to read the article again none of these technologies ( NVlink 2.0 / Photonic NVlink ) is wireless .



Because the communication speed and bandwidth requiremements are not high enough to make it worth going optical over traditional connection . It would simply add cost for 0 benefits .
Click to expand...
Alright, but what about at shorter distances like perhaps for mGPU? Perhaps shorter what about tiny optical connections between the PCIE bus links or across one end to the other!!? There has to be area's where they could insert optical tech in short distances at a lower cost to actually speed things up. I guess that's really what Nvidia is going towards here. That said longer runs will come in due time. They'll probably eventually have cables with a handful of chips built into them every so many feet or inches zipping optical along like a particle accelerator of sorts.
 
RH92 said:
No issues with trying to be funny but it's pretty hard to judge from your text . More importantly no optical ( as depicted in the article ) is in no way wireless , indeed there isn't any cope/alum/etc wire but instead there is a physical hollow cylinder ( read wire ) made out glass fibres which conducts the optical signal ( read light ) . Without this PHYSICAL glass wire the communication between the transmitter and the receiver would be impossible .


Now wireless optical communication exists for very short distances or/and space and military application but since it requires powerfull laser emitters ( for relatively long distance ) etc you can imagine that this is very far away from making it into your living room from an economical and even physical perspective .
Click to expand...
Uhm, I don't want to be the party crasher, but both light and RF radiation are the same thing, just different wavelenghts and different propagation media. They both obey the same laws of physics.
Simply, optics provide better spectrum efficiency.
 
RH92 said:
No issues with trying to be funny but it's pretty hard to judge from your text . More importantly no optical ( as depicted in the article ) is in no way wireless , indeed there isn't any cope/alum/etc wire but instead there is a physical hollow cylinder ( read wire ) made out glass fibres which conducts the optical signal ( read light ) . Without this PHYSICAL glass wire the communication between the transmitter and the receiver would be impossible .


Now wireless optical communication exists for very short distances or/and space and military application but since it requires powerfull laser emitters ( for relatively long distance ) etc you can imagine that this is very far away from making it into your living room from an economical and even physical perspective .
Click to expand...

I assume he meant it's "wireless" as in the fiber cable is a non conductive medium. You can use light to communicate without any wires.
 
Could be interesting for GPU / CPU interconnects, probably even for memory connections without relying on caching as much. Even just these first simple integrations would create a lot of demand I would imagine, boosting this technology even further.

Even though this idea has been floating around for at least 10 years now, this is way more detailed than they usually come out with. So I'm cautiously optimistic.
 
Last edited:
I'm sure they'll start with shorter runs of it and work at slowly perfecting it further make longer distances more feasible in some form or another. It probably looks a bit far fetched right now today, but give it 5 to 10 years and it could be a entirely different topic matter as more circuitry shrinks down. There is still a lot of chips made on much higher nodes than CPU/GPU chips that have a lot of catching up to do. I could see it integrated into the PCB in some form or another to speed up traditional copper traces. I'm wondering if they can 3D print integrated tiny optical circuits into the PCB itself. If they could use it between CPU and memory or GPU and memory that has some enormous implications.
 
light controllers were produced by Intel's r&d a decade ago, give or take a year. they never came to market for various reasons. this technology is not new.
 
You must log in or register to reply here.
Back
Top