Gregory's Server
@karotte Powerline network - near max cursed, somewhere between 10M and 200M depending on wiring and luck and the phase of the moon
|
13 comments
 @funkylab @gsuberland @karotte What about Thunderbolt? A huge pain in the neck, but super fast. And 10Gbase-R / 25Gbase-R (exponentially faster and not all all cursed)  @azonenberg @gsuberland @karotte I'm not very familiar with that, but I think the PHY is relatively similar to modern PCIe, so that would make it relatively uncursed. But the USB4 encapsulate-everything-in-everything has potential for much cursing, will see how that turns out!  @azonenberg @gsuberland the cursed thing about USB4v2 is that it uses PAM3 on the physical layer. @karotte @funkylab @gsuberland Yes, PAM3 is annoying. Especially because ~nothing COTS can do PAM3 so there's no hope of speaking it with an FPGA transceiver etc.  @azonenberg @karotte @funkylab @gsuberland just out of curiosity, what stops you from configuring the transceivers to PAM-4 and then just not using either the top or the bottom level? Do you get into trouble with level symmetry? @jaseg @karotte @funkylab @gsuberland For TX, I think it would work. For RX, I think you'd run into problems with asymmetry and the baseline wander correction putting your thresholds in the wrong place. @jaseg @karotte @funkylab @gsuberland (also even PAM4 transceivers are really rare, only like the biggest virtexes have them right now) @azonenberg @gsuberland @karotte The low-speed optical modes like that are relatively OK, it's basically linecode directly onto a laser that's either on or off, or PAM4. "infinite bw" assumption, pretty much. The actual long-distance (read: transatlantic) and high-rate (read: Tb/s) stuff is where the interesting stuff happens; 90 Gbd coherent IQ dual-pol xceivers, non-equal-prob'lity symbols and/or non-regular constellation diagrams, nonlinear-channel equalizers, FEC codewords of scary lengths… @azonenberg @gsuberland @karotte ( … and that's before you learn about the fact that in long-haul connections, you need to put in as much power as you can to get good OSNR at the output, but not as much that your mutual information per symbol drops due to phase smear due to fiber medium nonlinearity at high E-field strengths, but still want to do wavelength multiplex with only small guard bands between the carriers, but of course wavelengths' powers will mutually cause nonlin./phase distortion) @funkylab @gsuberland @karotte Oh I bet fiber gets cursed when you go fast enough. But at 10/25G its easy.  @azonenberg @funkylab @gsuberland @karotte I'd argue the FEC layer in 25GBASE-R is cursed due to all the transcoding and gearboxes to convert 80 64b66b codewords to one 5140b block to append the 140b parity resulting in the same 5280 bits. It's not pretty. 25GBASE-CR-S is nice and pure like 10GBASE-R but the FEC layer is required in everything else 25GBASE-R. The FEC itself isn't cursed just the wrapper around it.  @dlharmon @azonenberg @funkylab @gsuberland @karotte |
@gsuberland @karotte I must admit 10GBase-T feels about as cursed as 1OOOBase-T; less abuse of twisted pair to microwave frequencies, none of the original PAM-5 encoding, and more sensible FEC design, imho. I wonder where people would put V.90 – there's so much interesting stuff you have to pry from the devil's hot fingers in order to do 56 kb/s over a truly analog line, with a draconic bandwidth and power limit.