Write Amplification
Write amplification is the ratio between how much data an SSD's controller actually writes to flash and how much data the host asked it to write — and it is never one, because NAND flash has a rule no other common storage medium shares: a cell can be programmed from erased to written in small increments, but it can only be erased back to its blank state in large, whole-block operations, never one small piece at a time.
Garbage collection is the process that forces this mismatch into the open, and it is unavoidable once a flash block fills up with a mixture of live and dead data. A page cannot be rewritten in place — updating data means writing a fresh page elsewhere and marking the old one stale, since an individual page cannot be erased without erasing its entire surrounding block. Over time, every block accumulates some stale pages interspersed with pages still holding live data the host actually needs. To reclaim a stale page's space, the controller must first copy every remaining valid page in that block somewhere else, then erase the entire block so it can be reused — and that copy operation is a write the host never requested, performed purely so the controller can recover space the host's own update pattern left behind.
The write amplification factor is simply the ratio of total flash writes to host-requested writes, and it is bounded below by one but has no fixed ceiling. A perfectly sequential, full-drive write pattern that never triggers garbage collection mid-stream can approach a write amplification factor near one — nearly every flash write corresponds to an actual host write. A random, small-write-heavy workload that constantly scatters stale pages throughout many blocks forces garbage collection to copy proportionally more valid data for every stale page reclaimed, pushing the factor substantially higher; factors of two, three, or more are common under write-heavy random workloads, meaning the flash physically absorbs two, three, or more times the wear that the host's own data volume alone would suggest.
Write amplification is the point where every NAND-level physical constraint this series has covered converges into a single, user-visible number. Each extra write the controller performs is not free: it consumes real incremental step pulse programming cycles, each one taking real time and each one pushing real charge through the same oxide whose wear this series has already described as finite. A workload with write amplification of three is not merely writing three times as much data to the flash chips — it is consuming roughly three times the erase-program cycle budget, three times the ISPP pulse-verify overhead, and three times the cumulative oxide stress of the equivalent write-amplification-of-one workload, all to store the exact same amount of host data.
Over-provisioning — reserving flash capacity the host never sees as usable storage — is the standard, most direct lever for controlling write amplification, because it gives garbage collection more room to work with before it must act under pressure. A drive with generous spare capacity can let blocks accumulate more stale pages before reclaiming any of them, meaning each reclaimed block tends to hold fewer surviving valid pages to copy, and the controller can choose which blocks to reclaim more selectively rather than being forced to act on whichever block happens to run out of free space first. A drive with little or no spare capacity is forced into reclaiming blocks aggressively and often under a near-full condition, where blocks typically still hold more valid data that must be copied — directly raising the write amplification factor and, by extension, directly raising the real wear rate the drive experiences for the same host workload.
| Workload pattern | Typical write amplification | Why |
|---|---|---|
| Large sequential writes | Near 1.0 | Blocks fill and become stale as whole units — little to copy |
| Random small writes | 2-4+ | Stale pages scatter across many blocks — more valid data copied per reclaim |
| Heavily over-provisioned drive | Lower, for a given workload | More room for stale pages to accumulate before reclaim is forced |
| Nearly full drive, little spare | Higher, for the same workload | Garbage collection forced to act under pressure, copying more live data |
Write amplification is the system-level ledger that every physical property of NAND flash this series has described eventually gets entered into — ISPP's pulse-verify cost, finite erase-cycle endurance, and threshold-voltage margin all compound in direct proportion to this one ratio. A controller's garbage collection algorithm and over-provisioning budget are, in the end, the only real levers for keeping that ratio close to the one-to-one correspondence the host assumed it was getting in the first place.