TL;DR HDD stores data in spinning disks with magnified regions. SSD stores data in charge traps; it writes in pages and erases in blocks.
DRAMs or registers are considered volatile memory; the information gets lost when unplugged from power. We need non-volatile memory to persist and store data for the long term. The two common kinds of non-volatile memory are HDD and SSD. We'll cover some basics of it.
This video illustrated the whole thing pretty well with animations.
The disk/platter is made of metal. There could be multiple platters. The functional magnetic part is a 120nm thin layer of Cobalt Chromium Tantalum Alloy (Combination of elements Co, Cr, Ta). The material contains small magnetic domains that can be manipulated via external magnetic fields.
There's a head that flies 15nm above the platter. Current creates a strong magnetic field that forces the magnetic domain to align. The platter spins so that the head can reach any location on it.
There are > 500,000 tracks (concentric circles) on the platter. The tracks are divided into sectors. In each sector, the information is structured in the following way:
Information is encoded as the change of magnetic field across neighboring magnetic domains. A change signifies 1 and no change means 0.
There's also a DRAM component that buffers data and aids the control components.
Recent development focuses on further increasing the data density, such as pre-heating the magnetic domains with lasers.
Watch this video for fancy animations.
SSD uses flash memory which is used by phones and tablets as well. There's no mechanical part in it.
In some ways, it's similar to DRAMs because it stores information in memory cells. But those memory cells can persist data even without a power supply. Let's look at an individual memory cell.
The memory cells stores some quantity of electrons in charge trap flash. They can hold electrons for decades.
A cell can hold different levels of electrons and represent different bits of information. More bits of information make data storage cheaper but slower to extract the information.
In earlier years, floating gate transistors were used to store electrons. That was changed to a charge trap in recent years.
Operations
Memory cells are stacked vertically to form strings. The structure is called a V-NAND (Vertical-NAND). It was named NAND because the way the transistors are connected is similar to a NAND gate.
A string can be thought of as a column where you can read one value at a time. By duplicating the strings many times, you get a block where each individual row is a page. (Sources say that a string has around 100 cells stacked together. Suppose that's the case, with the simple duplication above, a block will only have exactly 100 pages. Since a block can have 256 pages, I would tend to speculate that by duplicating strings you get physical rows and multiple rows form a block. This is too much detail/speculation into the physical layout. It suffices to say that a block contains multiple pages.)
With a block, you can read one page at a time.
A page is the smallest unit that can be programmed. Common page sizes are 4KB/8KB/16KB.
A block is the smallest unit that can be erased. Erasing data requires a high amount of voltage and it's hard to restrict the impact to an individual page. This has complications that we'll discuss in the next section.
Just to give a sense of the scale for a 2TB SSD:
The block size above is 2MB. Block size can be as high as 8MB. In that case, the page size might be 16KB and there are 512 pages per block.
Although it's faster than HDD, SSD has its own complications and challenges.
One major constraint, as we mentioned, is that you can't erase an individual page; you need to erase an entire block. There is actually no way to “overwrite” a page. To “overwrite” a page, you are really reading out content that is still valid, erasing the whole block, and reprogramming the data into it. This causes write amplification where you only intend to write 8KB of data but behind the scene, 4MB of data needs to be written.
So it is usually preferred to write to clean pages rather than “overwriting” an existing page that's no longer in use. Old pages that are no longer needed are garbage collected at some point.
It's obvious at this point that SSD requires a much more sophisticated controller to manage all this. It needs to decide where the page of data should go and keeps track of a lookup table for address translation (user address → physical address). The lookup table is stored in a DRAM component, which is used to buffer read/write data as well.
The controller can decide to write to multiple chips at the same time, maximizing the throughput.
While quantum tunneling sounds amazing, the barrier that gets tunneled by electrons wears out little by little. A flash memory cell can only sustain a limited number of Program/Erase cycles. This limited the lifespan of an SSD device.
To extend the lifespan of the device, the controller needs to be smart about where to write the next page to evenly distribute the load.
Due to the NAND structure, to read the value of a single cell, you need to apply a voltage high enough for neighboring cells to pass through, causing some disturbance to the neighbors. After some number of such disturbances, the number of electrons trapped in the cells is no longer so reliable. Even for read-only loads, the controller might decide to erase the blocks and move the data elsewhere, causing further write amplification.
| HDD | SSD | |
| Cost | $ | $$$ |
| Latency | ~5ms | ~0.05 ms |
| Throughput | ~100MB/s | ~500MB/s |
| Lifespan | 3~5 years on average | 10 years (may be much shorter due to heavy load) |
https://www.youtube.com/watch?v=wtdnatmVdIg
https://www.youtube.com/watch?v=5Mh3o886qpg
https://www.youtube.com/watch?v=r2KaVfSH884
https://www.extremetech.com/extreme/210492-extremetech-explains-how-do-ssds-work