Name: MCF Optical Transceivers for AI DC
Uploaded: 2025-11-15T01:12:44.000Z
Duration: 9 min 51 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

Adverbs of manner

So this is a diagram actually showing multiple forum, but

Adverbs of degree

really just to show how complex the AI data center these days become.

Participle clauses

Basically you're actually looking at multiple segment of the networking actually needed, enabling the AI compute to work.

So this is actually have a scale up component, scale out component and scale across component just on the back end of AI networking.

And you still have to take care of the front end.

That's where all the data center service is interfaced to the end user request.

So I won't dive in too much there, just set the background.

So here just show you how intensive the fiber infrastructure needed for these AI data center.

This is actually some of the pictures showing one of the AI data center.

You can see the dock above all the compute.

You actually have layers of layers of pathways for fibers, cables to actually lay out and pass through from one part of the computer, the other part of the compute.

Sometimes going from one side of building to the other side of building, even between buildings in campus.

So the fiber number now sheer fiber number inside the data center started to become huge.

Even though each segment could be just up to like 12 kilometer.

But there are literally millions of these cables actually running between point A and point B.

So you add it up, actually start to become million kilometers right there.

So this really become very large segment of fiber deployment.

And to address these problems, certainly multi core fiber is one of the elements or tools now hyperscalers start to consider.

And like two by two is a really simple example.

But that's a very prime option for using multiple fiber address the pain point for these hyperscalers.

So you can actually have some key benefit just by deploying two by two versus a single core fiber.

You can cut down the number count immediately by four and you can actually reduce the number interface between fiber.

And you can greatly reduce the weight of these fiber.

And you can also provide much higher density in terms of layout these fiber cables.

Some of these customers tell us just the sheer weights of these fiber cable can be 100 pound per foot.

If you think about it, it will actually require very intensive foundational build up for each of these buildings.

So as a result not only cost but also space, but

also the time actually to get this AI data center built.

So there are these area of cost or pain point.

You don't realize what fiber actually can do.

Or you can help these hyperscaler to actually do a better job.

And additional benefit I mentioned, I think Alan mentioned earlier about ieee because if you go to the next level of Data speed from 200 git per lane now to 400 git per lane, you actually start to hit dispersion problems.

And traditionally we use cwdm, basically coarse WDM wavelengths to take care of this use less fiber.

But now if you just use cwdm, you're going to start to see dispersion problems actually even just for one or two kilometers.

So these are the area I think we really had to start a plan.

And one more additional benefit is we all know right now the channel counts of these data center optics start to go up a lot.

Right now we're really eight lanes and people already started talking about 16 lanes, even 64 lanes.

So the more you go, the fiber interface also start to become a bottleneck.

So this additional benefit, you can actually reduce the footprint just on the fiber coupling side.

You may not see it about when we build transceivers, we had to worry about these things.

Here's just to show why we had to build transceivers.

Because you can actually build these boxes to do single mode to multimode sorry to multi core converters.

Sometimes in industry people call it FIFO box, But

these are come with expense of additional costs, additional box, additional reinforcement, additional cost to actually build these boxes in the infrastructure.

So again, all our AI data center customer they are focusing on to produce the largest basically token per second or token per dollar token per watts for customers.

So any space allocate to other things they can allocate for compute.

So this has become one of the bottlenecks.

So we are basically here to develop a native so called multi core fiber native transceiver.

So customer does not know how to do conversion.

They can directly plug multicore fiber directly into transceiver.

So so you can actually have much cleaner deployment at much lower cost and space efficient.

And at OFC this year we actually demonstrated we can build now 800 gig transceivers with multiple fibers.

So this is actually built inside of transceiver with a multi core fiber converter.

So we can actually terminate it directly at the receptacles of the transceiver with a multi core input and output battery industry first 1,6T multicore fiber transceivers.

Now these are just to show you a snapshot of, you know, how we build these transceivers and actually produce really really good performance on the transmitter and receiver.

So customer basically can deploy these transceiver without worry about additional link budget, additional pain.

They need have to deploy these multi core fiber.

And this is a chart to show if you really want to scale the data center, you actually had to go a lot bigger.

Now these days, not only just one building, you're talking about these multi gigawatt level AI data centers.

You actually had to have many buildings in a campus site and sometimes you had to go a few kilometers and at times you even need to string together a few buildings about 10 to 20 kilometer apart.

So these are the area also start to see fiber issues.

And you can actually again meaningfully address a dispersion problem by Deploy MCF.

Because you can almost go four times longer distance with just a single wavelength like 13:10 instead of using a CWDM typically.

Now these days deploy in this kind of networks and higher density.

We are talking about for scale up AI data center right now.

Like I said earlier, the scale up portion is where you have the highest bandwidth density actually coming out of the compute.

If you're looking at like Nvidia GB200, they call it Mel 72.

Each one of these servers, basically each one of these Nvidia GPUs they actually have 7.2 terabit per second coming out of each of these server chips.

So literally you have 10x more bandwidth.

So what that means is if you are running at 200 git per lane, you literally had to have 10 like 36 or 32 lanes coming out of these transceivers.

So literally if you're looking at how much lane coming out, you really start to worry about how much density you need.

So we are talking about you can have 4x improvement if you use multicore fiber as a again coupling elements and then directly coupled to multi core fiber.

So this additional area you can actually use multicore fiber for this kind of applications.

And then looking at the multi core fiber deployment status, standardization is important.

Earlier commented ITU traditionally has been the standard body to write the spec for fiber.

But unfortunately I think multicore fiber right now I heard about starting a project but not yet get started.

So this is the kind of area I think the community need to come together and really start to work on it.

And I think the ecosystem building is important.

So this kind of need to get the whole end to end infrastructure in place.

Fiber transceivers, termination, splicing connectors.

So all these need to come together and for deployment because we do anticipate because of AI growth in the next couple years, they will start to have one or two hyperscalers.

Really started thinking about deploy in the next 23 years.

So if ITU are not acting fast enough, then maybe start to have activities like multi source agreement.

So at least we start to get a commonality of this is called a networking elements, including fiber to be somewhat standardized.

So this is the activity we're anticipating coming up.