100G-LR4 optic cable teardown | Hacker News Reader

100G-LR4 optic cable teardown

social.afront.org

187 pointstodsacerdoti62 comments

"so there's a peltier cooler underneath the platform holding the lasers and first optical stage so that the optic can electrically heat or cool the lasers to keep them at the right temperature! "

I get that Peltier coolers are simple, solid state devices, but still. Between that and the prisms and lenses, the amount of complexity here is mind blowing.

I'd expect that sentence to come from a hard sci-fi novel about weapons on an orbital platform or something. Instead, it's a pinky-sized bit of tech working to send cat videos to our phones. Amazing!

I'm surprised they use flip chip tech in something this expensive. Flip chips tend to be sensitive to light and this is after all a laser module. Raspberry pi had lots of issues with photosensitivity in its flip chip wifi controller. Use flash nearby and it would reset.

If it were me I'd just spend the 10c for the plastic..

Also really surprised the laser colors are only 4.5nm apart!

The size is amazing by the way.. I worked with fiber tech in the late 90s and a 16-colour DWDM multiplexer was the size of a full rack!

Perhaps it's a heat thing, plus the flip chips are in a separate space from the lasers.

I'd guess it's wanting for laser frequency to not shift with temperature

The STM32 is not responsible for generating the laser frequency and will not produce much heat on it's own

> Flip chips tend to be sensitive to light and this is after all a laser module

If a laser shines through to the chip then your SFP is dead already. SFP module has a metal case, SFP socket has a metal casing and usually surrounded in more metal around, be it a switch or a server.

> Also really surprised the laser colors are only 4.5nm apart!

Sounds tiny when you are talking about wavelengths, but convert to frequency and it's 3THz in near-IR.

100G-LR4 is a technology that's on its way out, and this illustrates why. Muxing four 25G lanes is complex and costly and these days also unnecessary at 100G. 100G is going single lambda.

This is still the way to build 200G/400G/800G optics, though.

Tuning Tx/Rx for NRZ can be challenging. PAM4 is going to be fun.

800G would be muxing 8 100G lanes? That must be pricy!

It's 8×100G yeah, but they also use two fiber pairs so four lanes go on each.

The QSFP-DD modules are also bigger (they stick out further from the slot) and have more power available, so that probably helps somewhat with engineering.

It's incredible how advanced the technology is, that fully saturates PCIE5.0 x16. but at a distance 1 million times farther.

It's fascinating to think you can combine and split signals on the fly with prisms. It should be completely obvious but I've never thought about it, my mind is far too set into digital habits. By comparison electronics feel so slow.

It's the micron-precise alignment of the four beams that is astonishing. I can't see which part(s) is actively aligned to line up the beams.

Tom

Yes. There are optical switches out in the market that use prisms to switch lambdas without any optical-electrical-optical conversion.

Those are not prism. Prism implies it used as dispersion component, which is not the case. The thing does the real work there is the dichroic thin film coating on those prism shaped glass.

Am I reading right that there's a direct interaction between magnetism and optical polarization? I know that light polarization is a pretty wacky topic to start with, but having magnets in the mix makes it seem unlikely amounts of wacky to me!

Anyone got any reading on the topic?

That sounds very wrong. Not a physicist, but light does not interact with magnets, or you need to get into the superconducting strength levels before it does.

https://en.m.wikipedia.org/wiki/Optical_rotation

It seems the optical rotation happens due to molecular structure of specific materials.

Edit, found this: https://en.wikipedia.org/wiki/Faraday_effect

So a specific material plus magnetic field is needed. But still not sure how strong of a field is sufficient.

For more details on this specific application these pages might be more relevant:

https://en.wikipedia.org/wiki/Optical_isolator

https://en.wikipedia.org/wiki/Faraday_rotator

Magnet there is probably used as an isolator. The laser can only go out,but not come back.( yes, there will be laser coming back from reflection, which also interrupt the operation of the source laser.

Yep, the Faraday Effect!

The sibling comment has a link to the Wikipedia page, yet for a quick video explanation and demonstration Applied Science's video on the Faraday Effect was fantastic. It demonstrates it, explains some use cases, and also goes into how the material the light is passing through affects the strength of the magnetic induced rotation.

https://www.youtube.com/watch?v=XhU-nNiAgtI

That TOSA looks like alien technology you would find on a 1 million years lost starship.

The headline here should probably be updated as this is an optical module teardown, not a cable teardown.

Great stuff, I am also happy that people are sharing it using mastodon hosted sites. great way to cultivate focused communities.

A facinating teardown! I wonder how long it will be before we manage to get 100G over reasonable copper length. https://en.wikipedia.org/wiki/100_Gigabit_Ethernet says we can have 100GBASE-CR4 for 7m but no Cat8 yet.

That would probably be more expensive than fiber. It's time to move on from CatN cabling.

Two strand singlemode already costs far less per meter than ordinary cat6 utp nevermind cat6a or shielded stuff. Copper isn't cheap.

Maybe my future house should be wired for fiber...

See https://news.ycombinator.com/item?id=33393540

one of the challenges there is that you can't do PoE over fiber, so for common house use with things like security cameras and 802.11ac/ax access points, you still want copper.

it also has a better bend radius by far now. (cat6a is about 1 inch, g.657a3/b3 is <0.25 inches)

Not easy to pull though typical European ductwork in the wall of a house.

And cat6 is ? Bend radius can be a problem with fiber but you can get bend-insensitive ones.

The problem with fiber is really you need fusion splicer for termination and that's pretty pricy

Couldn't you get away with mechanical splices? For short runs like this the extra losses shouldn't matter too much?

probably ? No idea about details.

I guess "just" terminating them would be fine if you just installed it once but at least looking at tutorials it does look like a PITA compared to CAT connectors

A lot of tutorials are old. Mechanical splices have improved very quickly in the past 5 years.

Talking with my friend who is an home automation installer, he says they now just use mechanical crimp splices and they don't take them longer than, say, prepping and terminating a coax cable. He says it's definitely shorter now than the average time they spend trying to untwist, straighten, and line up wires in a cat connector.

They used to use connectors requiring polishes that took 30 minutes

You can do without splices at all, just fish the cable through the wall with connectors and all.

The duplex LC connector can be disassembled, and with a bit of tape to keep things together you can get both fibers through much more easily.

If you're lazy, you can just use a premade cable that's longer than you need. A slightly prettier solution is a keystone jack.

mechanical splices should absolutely be avoided if at all possible, we live in the era of $900 made in china core alignment fusion splicers that are not terrible.

Yes, any level of bend-insensitive fiber has a better bend radius than cat6a. a3/b3 is much better.

Cat6a min bend radius is about an inch.

g.657a3/b3=~0.2 inch

g.657a2/b2=~0.39 inch

g.675a1/b1=~0.59 inch

as for connectors, not anymore. They have simple ones that are just mechanical crimp.

IE you cleave it, crimp it, and that's done.

0.15-0.25db loss, which is fine for a home run.

also, LC connectors/etc are small enough that you can just fish them (they are less than half the size of a cat connector), or disassemble them and re-assemble.

I have a fusion splicer, but i just fish completed cables with the connector covered.

Also, these days you can get a reasonable fusion splicer on amazon for ~750 bucks that is 0.025db average loss.

There are less and less reasons to run cat these days.

QSFPTEK just emailed me this morning offering me multimode 100G-SR4 optics (100M, so same as ethernet) for 35 bucks a piece.

They are regularly only 50 bucks a piece now anyway.

100G is obviously crazy for anyone's home need, but the point is the speed at which things like it are becoming cheaper is really high right now.

Copper's loss tangent is high. There's no getting around that. PAM4-based 50GBASE-CR DACs have similar length.

Twinax has more bandwidth than STP. We're at the physical medium's limits. Sure, you can pump the voltage but that quickly hits diminishing returns. Best to use copper for sub 1 meter and fiber otherwise.

No point.

10G DAC is already hot enough, bulky and too thick. With SFP modules it's even hotter.

I'd guess never. Twinax maybe, but "Cat8" would probably be more expensive than fiber above few metres already. Fiber is cheap, it's fancy electronics, lasers and photodiodes that make the thing expensive, but once you're there per metre price is cheap.

>pretty much everything else is voltage regulation for taking the 3.3V from the switch and regulating it down to lower voltages to drive the lasers

Why doesn't the switch deliver the voltages directly that are needed for the lasers?

Because the interface this module plugs into is not specific to optic data transmission. Also, remember how LEDs and lasers want constant current. A regulator circuit would be required anyway.

Very cool thanks for sharing, those Retimer chips must be quite something to handle the data flow 24/7 without ever failing.

Does Mastodon not allow longer posts or does the OP thinks the story would be better in different posts?

The poster puts over 1100 characters in one of the posts, while the default Mastodon server limits to 500 characters. It may be more about images - I think there's a default maximum of 4 images per post, so overall it kind of makes sense to tie each post's text with the matching image(s).

I'd love to see the same for a CSR4 module, those are very weird.

Since CSR4 uses a MTP-12 cable, I'm assuming it's effectively just 6 x tx/rx pairs compressed into a small package?

(i.e. no LWDM going on in the transceiver?)

Reply on news.ycombinator.com