Since NIC card receives data from memory in parallel, does it perform the transmission over the link serially? if yes, how?
If you think about it you already know the answer. To spur you along it is easier to think of optical cabling than copper to think through the example. With optical cabling there is one core for Tx, and one for Rx. This alone should be enough to answer your question, but with Wave division multiplexing it muddies the waters.
In answer to your question all information sent across an Ethernet network is sent serially. Multiplexing of any sort is sending multiple serial communications over the one core and is done by sending different channels over different wavelengths of light. However a fibre connection from an end device such as a server will generally only use one wavelength (e.g. 850nM for SX) to send all data serially over the wire. The same is true for copper connections. For up to FastEthernet 4 wires are needed as you have one pair for TX and another pair for Rx as they are electrically balanced.
Gigabit Ethernet over copper is slightly different as the signals are modulated over what is known as 4D-PAM5. What this means is that there are 4 signals over 5 voltages and 2 bits are sent at a time, however this is still classed as serial transmission, only with more information density. http://www.hardwaresecrets.com/how-gigabit-ethernet-works/
The reason many people gloss over this stuff is because if you actually look at the real details and not oversimplified models it's a massive rabbit hole.
Firstly Ethernet is not just one type of interface but a whole family of them with the fastest variant being FORTY THOUSAND times faster than the slowest.
Secondly many modern interfaces don't fit nicely into either the traditional parallel or serial model. Many modern interfaces use a multi-lane arrangement where multiple connections carry data in parallel but each "lane" has separately managed bit timings and the receiver adds small delays to make up for timing skew and bring the data back into sync.
Generally frames are first transferred from main memory to a buffer in the medium access controller. The interface used will depend on the application, an on-chip MAC is likely to use a parallel interface. Old off-chip MACs also used parallel interfaces but nowadays serial interfaces are increasingly used. 10 gigabit and faster interfaces typically use multi-lane PCIe which isn't really either a traditional parallel or a traditional serial interface.
The buffer inside the MAC is needed to decouple the timings on the Ethernet interface from the timings on the host interface.
The MAC decides when to transmit a frame (using CSMA/CD if needed) and transmits it at wire speed to the PHY. For 100 megabit and 1 gigabit systems the MAC to PHY interface is typically a narrow (2-8 bits) parallel interface, though serial interfaces are also sometimes used. For 10 gigabit and 25 gigabit I believe a fast LVDS serial interface is typically used. For higher speeds I believe a multi-lane interface built up out of fast LVDS serial links is used.
The PHY takes the data from the MAC and re-encodes it for the wire. 10BASE-T uses manchester encoding. 100BASE-TX first passes the data through a 4B5B encoding process. This is then further encoded using "MLT-3" to reduce the effective frequency. 1000BASE-T uses multiple pairs in parallel, uses a multi-level encoding scheme to send multiple bits at the same time on the same pair and uses echo and crosstalk cancellation techniques to use all four pairs in both directions at the same time.
On fiber at speeds up to 25 gigabit data is normally sent serially on a single wavelength.
At higher speeds there are a dizzying array of different options for transmission over fiber. Some use multiple separate optical lanes, either carried on separate fibres, some use multi-level PAM4 encoding and some use coherent optical technology where the phase of the light is modulated.
From Que Publishing
Serial Versus Parallel Information Transfer
Parallel Information Transfers
Parallel transfers use multiple "lanes" for data and programs, and in keeping with the 8 bits = 1 byte nature of computer information, most parallel transfers use multiples of 8. Parallel transfers take place between the following devices:
- CPU and RAM
- CPU and interface cards (see Chapter 8)
- LPT (printer) port and parallel printer
- SCSI port and SCSI devices
- ATA/IDE host adapter and ATA/IDE drives
- RAM and interface cards (either via the CPU or directly with DMA)
A serial transfer uses a single "lane" in the computer for information transfers. This sounds like a recipe for slowdowns, but it all depends on how fast the speed limit is on the "data highway."
The following ports and devices in the computer use serial transfers:
- Serial (also called RS-232 or COM) ports and devices
- USB (Universal Serial Bus) 1.1 and 2.0 ports and devices
- Modems (which can be internal devices or can connect to serial or USB ports)
- IEEE-1394 (FireWire, i.Link) ports and devices
- Serial ATA (SATA) host adapters and drives