Glossary

NAND flash

NAND flash is non-volatile memory that stores data as electrical charge in arrays of memory cells and keeps it without power. It is the storage medium inside SSDs, NVMe drives and all-flash systems, named after the NAND logic arrangement in which its cells are connected in series strings.

Pages, blocks and erase-before-write

NAND cells are grouped into pages, typically 16 KB, and pages into blocks of several hundred pages. Reads and writes, called programs, operate on pages; erasure operates only on whole blocks. A programmed page cannot be overwritten, so new data goes to a fresh page and the old one is marked invalid.

The drive's flash translation layer maps logical addresses onto these physical pages. When free blocks run short, garbage collection copies the still-valid pages out of a partly invalid block and erases it, and those extra copies are write amplification. Wear levelling spreads erasures across blocks, and spare capacity hidden from the host gives garbage collection room to work. As layer counts rise, blocks grow, so each erase covers more data and small random writes leave more valid data to copy.

Voltage states from SLC to QLC

A cell stores bits by holding one of several threshold-voltage levels. Each added bit doubles the number of levels inside roughly the same voltage range, raising density and narrowing the gap between neighbouring levels.

Narrower gaps bring more raw bit errors, slower programming and lower endurance. Every read returns some raw errors, more as cells wear, age or sit next to heavily read neighbours, and drive controllers correct them with strong codes, commonly LDPC, retrying the read at shifted reference voltages when a first decode fails. Charge also leaks slowly out of cells, faster when hot, which limits how long an unpowered drive keeps its data readable.

Stacked layers in 3D NAND

Planar NAND stopped shrinking in the mid-2010s, once cells were too small to hold enough electrons reliably. Manufacturers began stacking cells vertically instead, and current generations exceed 200 layers. Density now grows through layer count and bits per cell together, which is how single drives reached tens of terabytes and then passed 100 TB.

TLC versus QLC under a write-heavy load

TLC flash, or triple-level cell flash, stores three bits per cell across eight charge levels and is the mainstream cell type in data-centre SSDs. To program a TLC cell, the controller nudges it into one of eight narrow bands with a series of small pulses, checking after each one, so writes take longer than on single-bit flash. Data-centre TLC drives are commonly sold in two grades, read-intensive drives rated around one drive write per day and mixed-use drives rated around three. Client drives hide TLC's slow programming behind an SLC-mode cache that fills during long writes; data-centre drives are specified at sustained, steady-state rates instead.

Against QLC, TLC gives up a quarter of the capacity per cell (3 bits against 4) in return for a higher cycle rating and steadier write latency. Under write-once data the extra endurance goes unused: training corpora, media archives and backup copies barely touch it. Under churn it matters. An object metadata store rewriting small records, an index, or a bucket of small objects that are replaced constantly can wear a QLC tier quickly, while TLC absorbs the same load with margin and holds its latency while doing it.

Write shape, deletes and firmware in a large fleet

The pattern of writes decides both wear and latency. Large sequential writes of data that is never modified fill whole blocks that are later erased as a unit, keeping write amplification near one. Small random overwrites scatter invalid pages across many blocks, and the same drive then wears faster and shows wider latency tails. Object storage writes whole objects without updating them in place, and erasure-coded fragments arrive in large pieces, a pattern that suits NAND well.

Deleted data becomes free space only after garbage collection, and a drive learns which pages are no longer needed through TRIM or Deallocate commands. Storage software that passes deletions down keeps more free space available and write amplification lower in clusters with heavy churn, such as short-retention backup repositories.

Firmware shapes behaviour as much as the NAND does. Two drives built on identical flash can differ widely in latency consistency, garbage collection and error handling, so in a fleet of thousands of drives the firmware level belongs in the hardware specification. Retention matters at the other end of the lifecycle: enterprise drives are rated to keep data for months unpowered, which suits online data and leaves long-term offline copies to media built for shelf life.

Scality RING and NAND flash

Scality RING works above each drive's flash translation layer. Wear levelling, garbage collection and error correction inside an SSD remain the job of its controller and firmware, and RING does not change them. What RING decides is which drives hold which data, through storage classes that each carry their own erasure coding scheme. For RING XP, Scality describes NVMe flash at the storage layer serving small objects in the millions and billions, and whether the NAND in those drives is TLC or QLC follows from the drive model selected for each server.