New data transmission record | Hacker News Reader

New data transmission record

dtu.dk

45 pointsbelter20 comments

There is a point where bandwidth is so huge that with a bit of medium and a pair of repeaters, it becomes memory.

It already is? We already have a memory hierarchy from registers to L1/2/3 cache to main memory to ssd and finally to network. Each one increases in latency, sometimes massively

Reminds me of the Harder Drives video[1] where he made a program to store data in ICMP echo requests

[1] https://youtu.be/JcJSW7Rprio

That sounds really wild.

Banwidth != latency, however.

A truckload of TB disks will give you awesome bandwidth by driving it around to the destination.

The latency of a particular call, however, wouldn't make it usable.

Memory is hierarchical, but any latency beyond NVMe is non factor for treating remote data as memory

Agreed on light speed (latency). But the majority of the consuming applications need to process that data. If we have enough data to process, latency is only visible until the first initial bytes. After that it should look seamless to consuming application.

In what use cases all remote data is required in real-time?

not necessary all data, but random data access would make you (or your application) feel the latency.

It's very rare that you are able to just ingest data in a serial fashion For most scenarios, you need to jump between locations

Think an index, reading different files, etc.

Then the latency dominates

Oh yes. If you cram a lot of data in a laser beam reflecting on one of the Apollo reflectors, your average latency (the time it takes for the right part of the beam to arrive at the receiver) will be around 1 second.

If you do that with a Voyager, it will be about a day or so, IIRC.

> There is a point where bandwidth is so huge that with a bit of medium and a pair of repeaters, it becomes memory.

You are referring to the old "delay line memories" from yesteryear:

https://en.wikipedia.org/wiki/Delay-line_memory

I know that. Also magnetic drum memories behaved similarly (as do spinning disks, BTW) and were “main storage” on many machines from the past.

It's sort of interesting to think about how you would feed this with data.

So if there's a router box that on one side has the optical link that can transfer 1.8 pbit/s out then what's on the other side to get the data into the box? 180k 10gbit ethernet ports? Well, obviously not, so there would be ... 180 10 Tbit/s optical links?

And what would be the processor that could multiplex at that data rate?

If the latest intel 6.0ghz processor coult push out one cache line worth of data for each cycle, that would mean you need (1.8 * 10^15) / (64 * 6 * 10^9 * 8) = 586 of these processors! Could some custom silicon handle this or would you need some optical processing?

Wild stuff.

All the processing is necessarily done in photonics for these record transmissions. Even so, these kind of speeds are still impressive for photonics.

To answer your question more adequately, nothing is pushing out the data in or recieving it beyond a counter and an application-specific photonics chip that is able to process the frames enough to count them. Anything else assumes levels of processing that would make anyone interested in compute fight over it.

Right - I get it that in the experiment there was no real data being delivered, but trying to consider how one would implement it in an actually useful device for me puts the speed into context. It really is incredible.

The economicse of optical RAM might be nearer to the commercial tipping point. The power requirements per bit are already lower than those for semiconductor RAM.

"Optical RAM and integrated optical memories: a survey" Light: Science & Applications 9, Article number: 91 (2020)

https://www.nature.com/articles/s41377-020-0325-9

What was the distance it was transmitted over though? I don't recall seeing anything about the distance in the article (correct me please if I've missed it).

1mm, 1cm, 1m, 1km, 1000km?

Exciting, regardless though.

7.9km

The article links to their paper 'Petabit-per-second data transmission using a chip-scale microcomb ring resonator source'[1]

> We experimentally demonstrate transmission of 1.84 Pbit s–1 over a 37-core, 7.9-km-long fibre using 223 wavelength channels derived from a single microcomb ring resonator producing a stabilized dark-pulse Kerr frequency comb.

[1] https://www.nature.com/articles/s41566-022-01082-z

Oh ok I missed that part.

Thank you for correction!

Thank you for inquiring (or I hadn't realised I wanted to know too!)

(2022)

Reply on news.ycombinator.com