Hacker News Reader

Comment by rrgok | Hacker News Reader

Dumb question: what is stopping us from creating TCP/IP over USB4v2 so that I can create a home network (LAN) of 120Gbps?

View on HN

Using a valuable rate-limited post to say: you can. USB4 mandates direct host-to-host connectivity. If you bought an Intel laptop with 11th-gen/Rocket Lake (March 2011) or better, the cpu has built-in 40Gbps USB4/thunderbolt that can do host-to-host, although not every laptop implements the actual PCIe side of that. This has been support in Thunderbolt since Linux 4.15[1] in 2017.

Alas more generally, things aren't super hot for USB4 devices & things like ethernet integration. USB4<->ethernet options are still very very limited. Paying multi-hundred dollars for 10Gbe is... unfortunate. We have a long way to go before USB4 IP starts coming to actual devices; the hubs shield us from the obviousness of the need by being, in effect, a far-away mini-southbridge with wild behavior-switching on each port. So it's easy to keep doing what we're doing. But eventually I hope we stop using bridge chips and have some re-usable PCIe tunneling or other just better actual USB4<->flash or USB4<->ethernet chips that really try to be modern. It'll be a while.

As another poster says, cable length is short. 2m with active cables. USB3 had a lot of great cables with active repeaters built in (for a while there my desktop was 10m away from my desk), and I admit, I'm surprised I haven't seen this happen yet for USB4, but it should. One of my desires from USB is to make a longer range USB spec. Please give me 20Gbps over 10m? But for people just wanting to plug their laptop or desktop into their NAS, or wanting to direct-attach a mini-cluster together... heck yeah, USB4 is there for you, today.

[1] https://www.phoronix.com/news/Linux-4.15-Networking

The max cable Length at that speed is pretty darn short.

What's stopping us from adding a repeater every couple of feet? Transistors aren't expensive and integrating it into the cable shouldn't be too difficult.

It will be infinitely more expensive and harder than just using fiber. Two transceivers is 350 bucks total for 2km single mode. 200 for multi mode if you want cheapest possible A 4 port switch is about 700 bucks.

You could also do 25gb x 4 and aggregate at probably 1/3 price. You could upgrade later, without changing fiber. You could also go with the switch but use 25g transceivers, and upgrade the transceivers later. This is about 60% cost.

In a year this will all be half that price. USB4 plus a ton of repeaters plus … is going to dwarf the initial cost of 100g fiber quickly. In 5 years, you will probably be able to run 800g over the same fiber if you are bored. A or aggregate 100g links. You would have to replace all the USB with new USB-9 2x2.4 SE stuff.

* USB-9 2x2.4 SE Gen 3 Plus :D

120Gbps repeaters will be expensive. Making circuits that fast is incredibly tricky, with power, heat, rf (and more) issues.

If you think it's cheap make one and get rich.

But it's most certainly not cheap or easy or low cost.

I'm sure it's possible, but 40Gbps fiber bundles probably win easily on cost per bandwidth over any significant distance.

Transistors that can pass 120gbps are probably expensive as hell, possibly into InP territory.

The cheapest way to do this is over OS2 fiber. Look up how much money you have to spend to get 2x 100gbps OS2 PHYs and 2x 100gbps QSFP28 cards.

> Transistors that can pass 120gbps are probably expensive as hell, possibly into InP territory.

That's a strange argument, given that USB devices and your USB ports and CPU need a multiple of the amount of transistors in a repeater to process these signals.

Internal busses can operate in parallel, at some fraction of the serial bandwidth. You only need really fast stuff at the PHY. An interleaver/deinterleaver circuit is not much more complicated than the kind of signal shaping hardware you would need multiple copies of in the cable.

USB is a tree so IDK seems hard to generally map that to a packet switched network

If 100G NICs were avalible you could do: computer -> usb4 -> 100GB NIC -> optical -> 100GB NIC -> usb4 -> computer

We already have things like thunderbolt-net

And we already have link aggregation, so you can have 120gbps "networking" over very short distances today, if you want it. All you need is four short, expensive cables. The problem is you will rapidly run into system-level limitations, given that 120gbps would consume more or less all of the I/O resources on many common desktop-class CPUs.

The biggest thing stopping you might be the cost difference between 100 feet of cat7 versus a 100-foot certified Thunderbolt 4 cable.

The longest true thunderbolt 4 cable you can have is 3 meters or a bit under 10 ft.

I guess you could try making some kind of thunderbolt 4 media converter and do it over fiber, but R&D would probably be pretty steep.

There are optical Thubderbolt 3 cables made by Corning/Wero up 100m ($$$$). While not TB4, I've used regular TB3 cables with TB4 devices without issue. I assume it would be similar for those optical cables given same bandwidth requirements. I've also seen TB4 5m cables being sold with the listing saying they are rated up to 30gbps instead of 40.

You can get a dirt cheap 56gbps infiniband card today and connect it to your laptop via a thunderbolt riser.

100gbps is not dirt cheap but it exists today and definitely cheaper than the first 120gbps USB cards when they’re released.

There was at least one model of Dell switch that used HDMI for 10G connectivity a few years back, so it seems doable. I think most people looking for cheap and fast home networking end up going TCP over infiniband, which is a project I've been meaning to look at sooner rather than later myself. You won't touch 120G though, but I don't think you'll find disks that fast anyway.

As others have said, it comes down to length, if you want 120Gbps you’ll need to go with fiber

USB is strictly master/slave, unlike ethernet.

Reply on news.ycombinator.com