Comments Locked

23 Comments

Back to Article

  • cosmotic - Tuesday, September 18, 2018 - link

    No 2.5BASE-T or 5GBASE-T?
  • DanNeely - Tuesday, September 18, 2018 - link

    "Intel does not explicitly deny that the X550 works with 2.5G or 5G routers/switchers, yet the company does not validate the controller with such equpment (at least officially), so nothing is guaranteed."
  • Kevin G - Tuesday, September 18, 2018 - link

    Doesn't look like it according to the spec sheet. NbaseT for 2.5 Gbit and 5 Gbit speeds was actually ratified after 10 Gbit Ethernet. The big difference outside of the raw data rate is that 10 Gbit uses a more robust error correction scheme where as 2.5 Gbit and 5 Gbit leverage the same that is used in 1 Gbit.

    Another feature that the X550 is missing is AVB/TSN support. One of the nice things about the Aquantia cards is that I can use it as an audio-over-the-network interface. I've never moved more than 32 audio channels but according to bandwidth calculators on my 10 Gbit card, I should be able to move 1152 in both directions.
  • kobblestown - Wednesday, September 19, 2018 - link

    "The big difference outside of the raw data rate is that 10 Gbit uses a more robust error correction scheme where as 2.5 Gbit and 5 Gbit leverage the same that is used in 1 Gbit."

    Can you corroborate this? They both use the same transmission scheme as 10G at lower (1/4 and 1/2) signalling rate. And since they were introduced later - why would they gimp the ECC? Or maybe a lesser ECC is sufficient for the narrower spectra? I couldn't find enough details with Google.
  • Kevin G - Wednesday, September 19, 2018 - link

    It mainly boils down to 128 DSQ support which is not required with the 2.5 Gbit and 5 Gbit links due to their lower clock rate. 128 DSQ is also why 10 Gbit Ethernet is also why it consumes more power than one would expect vs. the slower Ethernet speeds.
  • Nevod - Thursday, September 20, 2018 - link

    You are not entirely correct about the error correction scheme. Actually, scheme used by NbaseT is a bit more robust than used by 10GbE, although by a small margin. It is, however, much more straightforward.

    Nbase-T uses LDPC error correction coding for all bits of message, and sends final encoded data using PAM-16 modulation, 4 bits per symbol.

    10GbE splits data to be sent into 7-bit bins, and only encodes 4 lesser bits with LDPC. LDPC data is concatenated with actual data, and is also split into 7-bit bins. These bins are then encoded using DSQ-128, which means that in a pair of PAM-16 symbols, only 128 combinations are used out of 256 possible, for the next sent bin the other 128-combination constellation is used. Thus, only the lesser bits, which are more susceptible to noise (as they correspond to lesser variations of signal level) are encoded. This scheme has probably been selected to reduce complexity of LDPC coder.

    1GbE uses somewhat similar Trellis encoding. IIRC, 8 bits are encoded as 5 PAM-5 symbols, thus only 256 out of 625 symbol combinations are allowed. Upon receive, an disallowed combination is recovered to closest allowed one. CRC is used to check for remaining errors, there is no further error correction than Trellis code.

    Much higher power consumption of 10GbE is due to much longer digital filters used at much higher symbol rate.
    10GbE has to use digital filters for echo cancellation(EC, subtracting signal transmitted at the given pair from signal received on it to permit full duplex operation), near-end crosstalk cancellation (NEXT, subtracts signals transmitted on other pairs from received on given pair), and far-end crosstalk cancellation (FEXT, subtracts signals received on other pairs from received on given pairs). For 1GbE only echo cancellation is mandatory, although I suppose better adapters do employ all the filters and are more robust to cable quality. For 10GbE, per each pair, EC filter has about 800 taps, NEXT filters have 3*300 taps (one set per each other pair), and FEXT have 3*100 taps, roughly. All these filters operate at 400Mhz. While 1GbE, having symbol rate 6.4 times lower, works with just 125 taps of EC per pair at 62.5 Mhz.
    Even if we consider an 1GbE adapter that has all the filters, it needs 1/6.4 the complexity at 1/6.4 speed of 10GbE. As such, digital filters for 10GbE need 40.96 times the power needed for 10GbE. That's not even considering much higher precision needed.
  • GreenReaper - Tuesday, September 18, 2018 - link

    It's supported "in Linux only", at least on the X550-AT2 controller:
    https://ark.intel.com/products/84329/Intel-Etherne...

    Moreover LR-Link's spec sheet for these cards claim support as well:
    http://www.lr-link.com/Upfiles/down/LREC9812BT-LRE...
  • CheapSushi - Tuesday, September 18, 2018 - link

    Just buy the ones from Aquantia instead. Much cheaper and they offer 10G, 5G and 2.5G, 1G, etc since they can go down the line when needed. The 5G is just $69.98 and sometimes even cheaper. And these are newer controller designs too.
  • piroroadkill - Wednesday, September 19, 2018 - link

    Without NBASE-T, these are worthless. Forget 'em
  • mooninite - Tuesday, September 18, 2018 - link

    $275 for two ports? LOL

    You can buy a motherboard with two X550 ports for $100 more. 10GB pricing is ridiculous. This is one market segment that needs cheap Chinese parts to flood the market and drive 10GB prices down.
  • GreenReaper - Tuesday, September 18, 2018 - link

    Double your price, double your pleasure? :-)
    Intel's own suggested price for the controller itself is $79.61, so that is part of it.

    Surely the cost of 10GbE will fall in time (although perhaps tariffs will add to that).
    But I imagine business users who truly *need* two ports right now will not be put off.
  • sor - Tuesday, September 18, 2018 - link

    Great. Now we just need devices that will let us connect these NICs together for a reasonable price. I’m not talking about that one or two low port count switches someone can name.
  • CheapSushi - Tuesday, September 18, 2018 - link

    Well, you could just connect two computers together without a switch; say your NAS and your main rig.
  • dgingeri - Tuesday, September 18, 2018 - link

    I got my single port x520 card for $89 and the dual port for $169. 10Gbase-T is just too expensive.
  • bcronce - Wednesday, September 19, 2018 - link

    Too expensive? My single port 1Gb i210 was $70 and dual port 1Gb i350-T2 was $130. The 10Gb cards are barely more expensive.
  • mooninite - Wednesday, September 19, 2018 - link

    The i3xx cards are a rip-off. Mine is collecting dust after it is incompatible with a Supermicro mobo of mine and RMAs didn't resolve it. Intel NICs are over-hyped, average devices with just as many problems as other NIC brands.
  • bcronce - Wednesday, September 19, 2018 - link

    Depends on what you mean by "overhyped". The interrupt rate, driver support, and packet processing is top notch. My firewall can process line rate 64bytes packets below 20% CPU usage on a quad core, while doing traffic shaping.

    114MiB/s transfer over SMB through the firewall is about 1% cpu and 150 interrupts per second, while maintaining microsecond latency.

    Not sure if I found the issue you were having, but I only got two hits when looking on the subject. I saw a Supermicro issue where the integrated NIC had a flaw, not i350's fault, and fake i350s having compatibility issues.
  • CheapSushi - Tuesday, September 18, 2018 - link

    The single port 10G NIC's by Aquantia start at $99.98 on Amazon and can often be found a bit cheaper. And these are brand new controllers. When they first launched they had a deal for about $65 I believe. So, no, this is not a good, competitive price.
  • abufrejoval - Wednesday, September 19, 2018 - link

    Low cost 10Gbase-T switches (actually NBase-T) are there now, but they make a racket.

    Buffalo BS-MP2008/2012 offer 8x or 12x NBase-T at around €50/port.

    But the variable speed fans are designed to cool 31/40 Watts for 8/12 ports with far too little regard for noise.

    I fixed that, which is the long story.

    tl;dr

    1st generation 10GBase-T is around 10Watts/port, which is the main reason it rarely found its way to the desktop: People wanted it in the same form factor and “the same cool” as 1GBase-T which is less than 1Watt/port.

    Aquantia and NBase-T is all about pushing that down towards 3Watts or so (hard figures seem impossible to get), both using Energy Efficient Ethernet and the slower 2.5 and 5Gbit signal rates. That figures nicely with the 40Watt power consumption of the 12-port Buffalo part, which I believe uses three AQR407 chips (three big heat sinks).

    But the two fans that come with the unit produce way too much noise, even in idle. After a couple of hours in the same room, it was I or ‘it’. So I opened it and had a look and touch at the fans and heat-sinks to gauge heat and noise sources. One fan was even noisier than the other but mechanically there seemed nothing wrong in terms of balance or obstruction with either.

    The problem seems to be that the rotor is pulsed to produce the variable speeds and varies the frequency not the strength of the pulse. That may be required for reliability at variable speeds, but it’s terrible for noise.

    The obvious solution would be large slow-moving fans or even bigger heat sinks, but you can’t fit those in the chassis and I don’t have a metal workshop. So I took an option I cannot recommend to anyone, because you are losing insurance and risk the life of your kids apart from any warranty!

    I replaced the fans with Noctuas of equal size and have those run off the 12Volt power rail at constant speeds using the stronger resistor they supply it. With that resistor in place, the sound no longer stands out from background noise. Running off the original fan connectors (careful, pinout is different!) the Noctua fans stop spinning in idle mode (around 3 of 12Volts), which is why I went with the constant speed variant.

    It has half the rated air-flow of the original fans, so in all likelihood it’s insufficient, especially if one should fail. But I configured both the switch and all NICs to “green” Ethernet and then stress tested it running iperf3 at 10Gbit for an hour on the four hosts I have. The chassis never went more than hand warm, so I hope it’s going to be ok.

    Unfortunately, there is no temperature sensor that can be read with the switches GUI, nor will it report any data on the fans, even if the fault indicator LED turns on as soon as you disconnect one of the original fans.

    I hope one of these days a manufacturer simply does the obvious and produces a double height variant with low-noise cooling: No switch needs to be noisier than the computers it serves!
  • mooninite - Wednesday, September 19, 2018 - link

    The problem I've seen with new switches is that they are poorly designed. If you open up a brand new 2.5/5/10G switch the actual PCB board is about 1/3rd of the size of the case that has a stupid, high-noise fan, blowing into a very empty case. The entire 10G market needs an overhaul with better designed and better priced parts. I'm not sure why this market segment is so lack luster.
  • oRAirwolf - Wednesday, September 19, 2018 - link

    Not bad if you need an rj45 10gbe card for a freenas server. Freenas has no support for the aquantia cards, so options are limited for rj45. I used an Intel x540-t2 in mine for a while and it worked fine, but I just switched to a mellanox connectx-3 sfp+ card, as it produces way less heat and works without installing drivers.

    I'm not really sure if these LR alone a cards are any better of a buy than a bootleg Intel card with no yotta mark from eBay. You aren't getting a warranty either way in America, so might as well go with the eBay option for less money and pray it works.

Log in

Don't have an account? Sign up now