How an SSD actually works

In an ordinary transistor, a voltage on the gate decides whether current flows. A flash cell inserts a floating gate between the control gate and the channel — isolated on all sides by silicon dioxide, one of the best insulators there is.

Getting electrons in and out

Since the floating gate is insulated, electrons cannot simply be placed there. The drive applies a high voltage — far above normal operating levels — and electrons cross the insulator by quantum tunnelling: a genuine quantum effect where a particle passes through a barrier it lacks the energy to climb over.

Reading is gentle and does no damage. The drive applies a moderate voltage to the control gate and sees whether the cell conducts. Electrons on the floating gate repel the channel, so a charged cell needs more voltage to switch on. Measuring that threshold reveals the stored value.

Why cells wear out

Writing is violent. Every tunnelling event drags electrons through the oxide at high field strength, and each pass damages the lattice slightly — leaving defects that trap stray charge. Over thousands of cycles the oxide degrades until it can no longer hold electrons reliably, or the voltage needed to clear it becomes unattainable.

This is the fundamental difference from a hard drive. A magnetic region can be rewritten indefinitely; a flash cell has a finite number of erase cycles, and every single one uses a little of it up. Everything in the next four steps is, in some form, a response to this fact.

A cell does not hold a binary quantity. It holds an amount of charge, and that amount can be measured. If you can reliably distinguish four charge levels you have stored two bits; eight levels gives three; sixteen gives four.

Why each extra bit costs so much more than the last

The voltage range a cell can hold is fixed. Splitting it into two levels leaves an enormous gap between them — the measurement can be sloppy and still be right. Splitting the same range sixteen ways leaves gaps a fraction as wide, and now every source of error matters: manufacturing variation, temperature, charge leaking away over months, interference from neighbouring cells being written.

This is why writing gets slower as well as less durable. Placing charge accurately enough to land in a narrow band takes multiple careful pulses with verification between them, rather than one confident blast. A QLC write is several times slower than an SLC one on identical silicon.

The SLC cache trick

Consumer drives exploit this cleverly. A region of the TLC or QLC flash is operated temporarily in SLC mode — one bit per cell, wide margins, fast writes. Incoming data lands there at full speed, and later, while the drive is idle, it is repacked into dense multi-level storage in the background.

This is why a cheap SSD writes at 3,000 MB/s for the first thirty seconds and then collapses to 150. You have filled the SLC cache, and further writes go directly into QLC. Reviews that only measure short transfers miss this entirely; it is the single most important thing to check before buying a large, cheap drive.

Going upwards instead of smaller

Around 2013 shrinking cells further started making everything worse — fewer electrons per cell means less margin and worse endurance. The industry stopped shrinking and started stacking, building cells in vertical layers. 3D NAND now reaches over 200 layers, and because the layers can use a larger, more reliable cell, it improved density and endurance at once.

Cells are grouped into pages, typically 4 or 16 kilobytes, which are the smallest unit that can be written. Pages are grouped into blocks of a few hundred pages — often several megabytes — and a block is the smallest unit that can be erased.

What actually happens when you change a file

1. The operating system asks to rewrite a 4 KB block of a document.
2. The drive cannot overwrite that page, so it writes the new version to a completely different, already-erased page somewhere else.
3. It updates its internal map so that logical address now points at the new physical page.
4. The old page is marked stale. It still holds data and still occupies space, and it cannot be reused until its whole block is erased.
5. Later, garbage collection copies the still-valid pages out of that block elsewhere, erases the block, and returns it to the free pool.

Write amplification

Notice what step five costs. To reclaim a block containing one stale page and 255 valid ones, the drive must rewrite 255 pages of data that nobody asked it to touch. The ratio of flash actually written to data the host requested is write amplification, and it directly consumes the drive's finite endurance.

A well-designed controller with plenty of free space might achieve 1.1. A nearly full drive doing small random writes can exceed 10 — meaning every megabyte you save costs ten megabytes of wear. This is the single best reason to leave an SSD 10 to 20 percent empty.

TRIM: telling the drive what is rubbish

When you delete a file, the operating system simply marks its space free in the file system. The drive is told nothing, and so continues carefully preserving pages of a deleted file during every garbage collection — copying data that no longer matters, for as long as the drive lives.

The TRIM command fixes this: the operating system tells the drive which logical blocks are no longer in use, and the drive marks those pages stale immediately. Blocks can then often be erased with nothing to copy out at all. Every modern operating system issues TRIM automatically, and a drive without it degrades measurably within months.

The controller therefore never lets the host decide where data physically lives. It maintains a flash translation layer: a map from the logical addresses the computer uses to the physical pages actually holding the data. The map changes constantly, and the computer is never told.

Why a 512 GB drive contains more than 512 GB

Part of this is the familiar decimal-versus-binary discrepancy. The rest is deliberate: manufacturers fit, say, 512 GiB of flash and advertise 500 GB, keeping the difference as spare. Enterprise drives reserve far more — a drive with 1,024 GiB of flash might be sold as 800 GB, trading capacity for endurance and consistent performance.

Reading the endurance rating

Manufacturers publish TBW (terabytes written) or DWPD (drive writes per day). Both describe the same thing from different directions, and both are conservative warranty figures rather than the point of failure.

The practical conclusion is reassuring. A normal desktop user writing perhaps 20 TB a year would take thirty years to exhaust a 600 TB rating. Endurance is a genuine concern for databases and video editing, and essentially theoretical for ordinary use — SSDs are far more likely to be replaced for being too small.

How they fail, and the one real gotcha

A worn SSD usually becomes read-only rather than losing data: the controller detects it can no longer erase reliably and refuses further writes, leaving everything readable so it can be copied off. That is the graceful case, and it is considerably kinder than a head crash.

The genuine gotcha is different. Charge on a floating gate does slowly leak, and the rate rises sharply with temperature and with how worn the cells are. A drive left unpowered in a warm loft for several years can lose data, and a heavily worn one is far more vulnerable. An SSD is an excellent working drive and a poor choice for a backup left in a drawer for a decade — for which, ironically, a mechanical drive or tape is safer.

SATA and its AHCI command set were designed around a mechanical drive: one head, one operation at a time, and reordering to minimise arm movement. An SSD has no arm, no rotation, and can genuinely service many requests simultaneously across independent flash chips.

The numbers

Form factors, which are a separate question

The shape of a drive and the protocol it speaks are independent, and conflating them causes real purchasing mistakes. An M.2 slot is a physical connector; what it supports depends entirely on the motherboard.

The trap is buying an M.2 SATA drive for a machine with an NVMe-only slot, or the reverse. They look nearly identical, and the only reliable check is the notch pattern on the connector against the motherboard manual.

Heat is now a genuine consideration

Fast NVMe drives produce meaningful heat and will thermally throttle, sometimes dramatically, during sustained transfers. A PCIe 5.0 drive without a heatsink can drop to a fraction of its rated speed within a minute. The heatsinks on modern motherboards are not decoration, and a drive that benchmarks brilliantly for ten seconds may be unremarkable for a large file copy.

The headline speed is the least useful number

'Up to 7,000 MB/s' describes a large sequential transfer into an empty SLC cache — a workload almost nobody has. Real use is dominated by small random reads: loading an application, opening a project, booting. Those are measured in IOPS and in latency at low queue depth, and they are rarely advertised.

DRAM-less drives, and why they feel slower

The flash translation layer's map is large — roughly a megabyte of map per gigabyte of drive. Good drives keep it in a dedicated DRAM chip. Cheaper ones omit the DRAM and either store the map in flash or borrow system memory over the PCIe bus, a feature called HMB.

The difference rarely shows in a benchmark and often shows in use. Every random access may require a map lookup from flash first, which roughly doubles the latency of small operations. If a drive is unusually cheap for its capacity and speed, the missing DRAM is the most likely explanation.

Keep it partly empty

This is the most actionable advice here. A nearly full SSD has few free blocks, so garbage collection must work constantly and write amplification climbs steeply. Performance can fall by more than half, and wear accelerates at the same time.

1. Leave at least 10 percent free, and 20 percent if you write heavily.
2. Make sure TRIM is enabled — it is by default on every current operating system.
3. Do not defragment an SSD. There is no seek penalty, and it is pure unnecessary wear.
4. For large sustained writes, check the sustained figure rather than the peak.
5. Treat an SSD as a working drive, not as long-term unpowered archival storage.
6. Check SMART attributes occasionally; percentage used and reserved blocks give genuine warning long before failure.

Where each type still makes sense