Network Interface Cards (NICs) 101

[aaab]

Also if you have a crappy switch a good network card is probably not going to make much difference.

 

[ubercake]

I guess another network topic would be switch performance. I wonder how many of us are ruining our internal network speeds with low-cost ($20) Gigabit switches. I have at least two of them in my house. Is there a difference (in a home environment) between buying the $100 switch or the $20 switch? Maybe in the "...103" article?

 

[Someone Somewhere]

I have a super cheap old unmanaged switch on my desk... still get ~0.4ms round trip through it and another switch.

Don't think I've ever really tried to push it in terms of bandwidth.

 

[gamebrigada]

Didn't read the full article, writer doesn't seem very knowledgeable on the topic, lots of wrong assumptions. Cat6 is not the standard cabling in everyone's home, cat5e is. A lot of manufacturers cheap out and do half duplex cat5e because you'll never notice the difference. 8P8C and RJ45 are entirely different connectors.

Cat6 is capable of 10Gb speeds, but it cannot achieve 100Gb just with simply a new NIC. Cat7A cabling is required for 100Gb at a maximum distance of 49 ft. Theoretically the cable will be able to do 100Gb at 328ft with 32nm or 22nm circuits. Cat6a, cat7 and cat7a are extraordinarily expensive and difficult to work with. Unshielded cat6a is as thick as your thumb. Cat7 is thinner due to using shields, usually a foil on each pair and a stranded shield on the outside. Cat7a is even thicker than Cat6a due to it being a combination of Cat6a and cat7.

 

[CerianK]

A few notes/opinions/errata, since I work with 1000s of point-to-point network connections that require 24/7 up-time with zero packet loss:

PCI (and PCI-e) probably should be considered LIF (Low Insertion Force), not ZIF.

Some network interface card/driver combinations will drop packets several times per day. PCI, older PCI-e, and some newer PCI-e with older drivers are more prone to this issue.

Manually setting the Speed and/or Duplex is seldom required. Doing so without careful consideration can further interfere with the integrity of network communications when an issue with controller, firmware, driver or cable is present as a root cause issue.

Many (if not most or all) newer network interface controllers support Auto-Crossover, just in case you need to create a point-to-point connection between two PCs and do not have a cross-over cable handy.

Though seldom needed, a few motherboards that have dual network ports on board also support auto-switch if only one network cable is plugged in and its controller chip fails, to re-route communications through the other chip.

 

[falchard]

... This topic completely went over my head. For offloading CPU load, I don't think it matters anymore. I have great difficulty maxing even 1 core and I am using an AMD FX-8320. The only task I can do that will is CGI calculations, and I am not doing network operations then since I don't have a render farm.

 

[Crunch140]

"4" in your picture is a Transformer (also referred to as Magnetics). The EEPROM is at U1.

 

[Crunch140]

Are you, Tom's Hardware, sure about the following:

"If the data is received from the Ethernet cable, the controller is responsible for stripping the first three layers of encapsulation prior to handing the data off to the computer."

It is my understanding that the NIC is only responsible for the Layer 1 ("frames") of the IP protocol model (Layers 1 & 2 of the OSI reference model) and that the OS itself (whether Windows, Linux, etc.) is responsible for Layers 2 & 3 of the IP with some Layer 4 processing, also.

Are you, Tom's Hardware, sure about the following:

"If the data is received from the Ethernet cable, the controller is responsible for stripping the first three layers of encapsulation prior to handing the data off to the computer."

It is my understanding that the NIC is only responsible for the Layer 1 ("frames") of the IP protocol model (Layers 1 & 2 of the OSI reference model) and that the OS itself (whether Windows, Linux, etc.) is responsible for Layers 2 & 3 of the IP with some Layer 4 processing, also.

Are you, Tom's Hardware, sure about the following:

"If the data is received from the Ethernet cable, the controller is responsible for stripping the first three layers of encapsulation prior to handing the data off to the computer."

It is my understanding that the NIC is only responsible for the Layer 1 ("frames") of the IP protocol model (Layers 1 & 2 of the OSI reference model) and that the OS itself (whether Windows, Linux, etc.) is responsible for Layers 2 & 3 of the IP with some Layer 4 processing, also.

PHYs are responsible for the signaling at L1. The PHY is integrated into most 1Gbs NICs. The rest is not so straight forward. It depends on the OS, and the Protocol stack. A standard NIC does indeed have jobs up to L3-4 (See LSO, and checksum offloads). The application however has to have knowledge of L3, so no a NIC does not just simply strip off the headers. The Protocol stack has to have knowledge of L2 in 'most' cases, so no the NIC doesn't simply strip the L2 header (see ARP).

 

[Kewlx25]

... This topic completely went over my head. For offloading CPU load, I don't think it matters anymore. I have great difficulty maxing even 1 core and I am using an AMD FX-8320. The only task I can do that will is CGI calculations, and I am not doing network operations then since I don't have a render farm.

Offloading is not as useful for most situations, but NICs that do not have decent offloading are missing much more important features that are much more complex and costly than offloading.

Kind of like a novelist who can't spell. Spelling is not important because your word processor will spell for you, but if someone is good at writing, their time spent writing should have helped them learn to spell. A big red flag.