Glossary

QLC flash

QLC (quad-level cell) flash is NAND flash that stores four bits in each memory cell by distinguishing 16 voltage levels. It holds a third more data per cell than TLC flash, at the cost of lower write endurance and slower native writes, and is used mainly in high-capacity drives for read-intensive data.

Why QLC flash matters for capacity tiers

For most of flash's history, large-capacity tiers stayed on disk because flash cost too much per terabyte. QLC narrowed that gap far enough for flash to take on data that is large, read often and rewritten rarely: data lakes, AI training corpora, media libraries and content repositories. It is also what has pushed single data centre drives above 120 TB, changing how many servers, racks and watts a petabyte of flash requires.

How QLC stores four bits

A NAND cell's value is set by its threshold voltage. Four bits need 2⁴ = 16 distinguishable states. The usable voltage window is broadly similar across cell types, so QLC fits 15 gaps between states into the space where TLC fits 7, making each gap roughly 7 ÷ 15 ≈ 0.47 as wide.

Narrow gaps mean a smaller drift of stored charge moves a cell into a neighbouring state. QLC controllers therefore rely on stronger error correction and on read retry, re-sensing a page with shifted reference voltages when the first decode fails. Fresh cells usually decode first time; as cells age, more reads take the slower path.

Penta-level cell (PLC) flash, five bits in 32 states, is the next step. It adds 25% per cell over QLC (5 ÷ 4) while again roughly halving the gap between states, so each added bit buys less capacity and costs more margin.

Write behaviour and endurance

Programming a cell to one of 16 levels takes more, finer pulses than programming it to one of 8, so native QLC writes are slow. Drives run part of their NAND in single-bit mode as a write cache that absorbs bursts quickly and folds the data into QLC blocks later. A write stream long enough to fill that region proceeds at the native QLC rate, and folding writes the data a second time.

QLC drives carry lower drive-writes-per-day ratings than TLC drives, often below one. At large capacities a low rating still covers a lot of data: a 61.44 TB drive rated at 0.3 DWPD for five years accepts about 0.3 × 61.44 × 365 × 5 ≈ 33,638 TB of writes, as set out under flash storage endurance.

What QLC flash means for petabyte-scale data

Fit is decided by write pattern. Datasets written once and read many times, such as training corpora, data lake history, media archives under active use and analytics sources, spend a small share of QLC endurance. Workloads that rewrite constantly, such as metadata journals, caches, scratch space and frequently replaced checkpoints, wear QLC quickly and run into its slow native writes; those roles sit better on TLC.

Growth rarely strains endurance. A data lake adding 1 PB a year onto 10 PB of QLC writes about 1.33 PB including 9 + 3 erasure coding, roughly 0.13 of its capacity per year or about 0.0004 drive writes per day, far inside even the lowest QLC ratings.

Bulk loads run at sustained speed. A multi-petabyte migration onto QLC outlasts every drive's write cache within minutes, so the realistic ingest rate is the native QLC rate multiplied by the number of drives, well below burst figures.

Reads hold up better than writes. QLC sequential read throughput is close to that of TLC drives on the same interface, because reading avoids the slow programming step. Read-heavy pipelines such as AI training streams and analytics scans therefore lose little on QLC, while ingest-heavy stages feel the difference.

Mixed media in one server is common. Capacity on QLC alongside a smaller set of TLC drives for metadata and write-intensive services lets each class do the work it is rated for.

Failure domains get larger. One drive above 120 TB holds what several large hard drives hold, so each failure triggers reconstruction of more data. Flash read speed shortens the work, and the protection layout determines how exposed the remaining data is while it runs.

Density changes the facility arithmetic. Fewer drives and servers per petabyte reduce rack space and power, while IOPS per terabyte fall below those of smaller TLC drives, though still far above disk. Read latency also spreads wider as drives age and read retry becomes more common, which matters most for small-object, latency-sensitive services.

QLC capacity and Scality RING

Scality RING is software on standard x86 servers, so the choice between QLC and TLC drives is made in the hardware configuration of each deployment, and erasure coding schemes are set per storage class. Across its customer base Scality reports about 6 exabytes under management, with up to 300 billion objects in a single RING.