[MadVikingGod]

So the findings here do make sense. For sub 5m cables directly connecting two machines is going to be faster then having some PHY in between that has to resignal. I'm surprised that fiber is only 0.4ns/m worse then their direct copper cables, that is pretty incredible.

What I would actually like to see is how this performs in a more real world situation. Like does this increase line error rates, causing the transport or application to have to resend at a higher rate, which would erase all savings by having lower latency. Also if they are really signaling these in the multi GHz are these passive cables acting like antenna, and having a cabinet full of them just killing itself on crosstalk?

[Palomides]

high speed links all have forward error correction now (even PCIe); nothing in my small rack full of 40Gbe devices connected with DACs has any link level errors reported

[kazinator]

They looked at the medium itself, not the attached data link hardware.

Look at the graphs. The fiber has a higher slope; each meter adds more latency than a meter of copper.

This is simply due to the speed of electromagnetic wave propgation in the different media.

https://networkengineering.stackexchange.com/questions/16438...

Both the propagation of light in fiber and signal propagation in copper are much slower than the speed of lightin vaccuum, but they are not equal.

[jauntywundrkind]

There's also hollow core fiber, which is pretty close to speed of light in a vacuum. 2.0e8 m/s for fiber, 2.3e8 m/s for copper, pretty close to the full 3e8 m/s for hollow core.

No glass, just some reflective coating on the inside of a waveguide (hollow tube).

https://azure.microsoft.com/en-us/blog/how-hollow-core-fiber...

[timewizard]

> are these passive cables acting like antenna

With both ends connected to a device? No.

Aside from that you've got a linear scrambler into balanced drivers into twisted pair. It's about as noise immune as you can get. Unless you put the noise right up next to the cable itself.

[Hilift]

Storage over copper used to be sub optimal but not necessarily due to the cable. UDP QUIC is much closer to wire speed. so 10 GB copper and 10 GB fiber are probably the same, but 40+ GB fiber is quite common now.

[laurencerowe]

> So the findings here do make sense. For sub 5m cables directly connecting two machines is going to be faster then having some PHY in between that has to resignal. I'm surprised that fiber is only 0.4ns/m worse then their direct copper cables, that is pretty incredible.

Surely resignaling should be the fixed cost they calculate at about 1ns? Why does it also incur a 0.4ns/m cost?

[cenamus]

Light speed is ~3ns per metre, so maybe the lowered speed through the fibre?

Speed of electricity in wire should be pretty close to c (at least the front)

[bhaney]

> I'm surprised that fiber is only 0.4ns/m worse then their direct copper cables

Especially since physics imposes a ~1.67ns/m penalty on fiber. The best-case inverse speed of light in copper is ~3.3ns/m, while it's ~5ns/m in fiber optics.

[p_l]

DACs don't cause problems, but twisted pair at 10Gig is a PITA due to power and thermals

[somanyphotons]

What allows DACs to avoid the power/thermal issues that twisted pair has?

(My naive view is that they're both 'just copper'?)

[tcdent]

PHYs are going away and fiber is going straight to the chip now, so while the article is correct, in the near future this will not be the case.

[sophacles]

The chip has a phy built into it on-die you mean. This affects timing for getting the signal from memory to the phy, but not necessary the switching times of transistors in the phy, nor the timings of turning the light on and off.