The cost of four bits lives in the voltage window
QLC stores four bits per cell, which requires resolving sixteen states inside the same threshold-voltage range. Margins between adjacent states shrink, tolerance to voltage drift, read disturb and temperature swing falls with them, and programming needs finer multi-pass verification. The trade is plain: QLC buys cost per gigabyte and areal density, and pays in endurance, write rate and retention margin.
Read the steady-state curve, not the peak
Multi-pass verification means a QLC program operation takes markedly longer than TLC, and that gap only shows itself fully under sustained load. Short bursts can still be partly hidden by concurrency, but once writing continues to the point where background reclaim has to compete with foreground traffic for media time, the rate falls back to what the medium itself can give. The metrics that matter when evaluating a QLC product are therefore the steady-state write curve and latency distribution after hours at full load, not the peak bandwidth on the datasheet header — and whether the host workload leaves a long enough idle window for background work to tidy the medium often determines real behaviour more than any single drive parameter.
Workloads that genuinely fit
- Q7 (J760Q, J760QD, PCIe 4.0, U.2/U.3): object storage, backup and archive. Large sequential writes with sparse reads sidestep exactly what QLC is bad at.
- Q9 (J960Q, J960QD, PCIe 5.0): AI checkpointing and very large capacity arrays. Checkpoints are big sequential writes read back infrequently — heavy per event, but well spaced.
The inverse holds too. Metadata, journals and KV cache are small high-frequency writes and should not land on QLC; that is what the X7 and X9 series exist for. Both QLC lines are currently at product definition/EVT, so the split above is planned positioning, with capacity points and performance values to be settled in the subsequent SKU definitions.
Second-order effects of very high capacity
The portfolio plans for a maximum single-drive capacity of 122.88TB. Pushing toward that class scales rebuild time, the blast radius of a drive failure, and background refresh overhead in proportion. Dual-port parts such as the J760QD and J960QD add link redundancy on the array side, but that addresses path failure, not device failure. Density gains have to be evaluated together with availability design at the array layer, or the cost saved comes straight back during a single rebuild window.