An SSD can write substantially more data to NAND than the host sends to the device. The extra traffic appears when the controller must relocate still-valid pages before reclaiming flash blocks that contain invalid data.
This internal movement is write amplification. It is a consequence of the mismatch between fine-grained logical updates and NAND erase constraints, not an extra write issued by the application.
NAND pages cannot be overwritten in place
NAND flash is programmed in pages but erased in larger erase blocks. A page that already contains programmed data cannot generally receive an arbitrary in-place replacement. The controller writes the new version to another available page and marks the old physical page as stale in its mapping state.
Repeated logical overwrites therefore leave erase blocks containing a mixture of valid and invalid pages. Invalid pages no longer hold current host data, but their physical space is not reusable until the containing block is erased.
The flash translation layer maps host logical block addresses to physical NAND locations, allowing this relocation to remain hidden behind a conventional block interface.
Garbage collection creates internal writes
A block with invalid pages becomes a candidate for reclamation. If it also contains valid pages, the controller cannot erase it immediately without losing current data.
The controller first copies those valid pages elsewhere, updates its mapping metadata, and then erases the old block. The reclaimed block can return to the free-block pool.
Those relocation writes consume NAND program operations even though the host did not request new logical data for the relocated pages. A host workload that writes 100 GB can therefore cause more than 100 GB of physical NAND programming.
Write amplification is commonly expressed as the ratio of physical media writes to host writes. A ratio near 1 indicates little extra movement. Higher values indicate that internal relocation or related media-management work is adding write traffic.
The exact accounting exposed by a device is implementation-specific, so telemetry fields from different SSD families are not always directly comparable.
Free space changes reclamation cost
Garbage collection is easier when the controller has a healthy supply of erased blocks. New writes can be placed without immediately reclaiming partially invalid blocks, and the controller has more freedom to select blocks with fewer valid pages.
As usable free space becomes scarce, that freedom shrinks. Reclamation may have to process blocks containing a larger fraction of live data. More valid pages must then be copied to free one erase block.
This relationship is one reason SSDs reserve physical capacity beyond the logical capacity visible to the host. Over-provisioned space gives the controller working room for placement, wear management, and garbage collection.
Host-visible unused space can also help when the device is informed that the corresponding logical blocks no longer contain required data.
Deallocation removes stale logical ownership
Deleting a file does not inherently tell a block device that its old logical blocks may be discarded. The filesystem can reuse those addresses later while the SSD still treats their current mappings as valid.
Block deallocation changes that state. In NVMe, the Dataset Management command can mark logical block ranges as deallocated. The controller may then treat the previous contents of those ranges as no longer required by the host.
That information can reduce the amount of valid data that garbage collection must preserve. It does not guarantee an immediate physical erase, and controller behavior remains implementation-dependent.
Deallocation therefore changes the controller’s knowledge of logical validity rather than acting as a direct command to erase a particular NAND block.
Data lifetime mixing raises relocation pressure
Placement affects the amount of live data copied during reclamation. If short-lived and long-lived data occupy the same erase blocks, frequent invalidation of the short-lived pages can leave long-lived pages stranded among reclaimable space.
Reclaiming such a block requires moving the long-lived pages even though their logical contents have not changed. Repeated mixing can cause the same persistent data to move multiple times as neighboring data expires.
Keeping data with similar expected lifetimes closer together can reduce this effect. Zoned storage and placement-oriented interfaces expose more control over physical organization specifically because opaque placement inside a conventional SSD can create garbage-collection overhead.
The benefit depends on workload behavior and controller design; sequential host I/O alone does not guarantee low amplification if data lifetimes and internal placement still produce costly reclamation.
Write amplification affects endurance and latency
Every additional NAND program operation consumes part of the media’s finite program-and-erase cycle budget. Internal relocation therefore connects workload behavior to device endurance even when host write volume is unchanged.
Garbage collection also competes for controller channels, NAND bandwidth, buffers, and scheduling slots. Under sustained writes, foreground requests can overlap with reclamation activity. Latency can rise when the free-block pool is low or when substantial valid data must be moved.
The resulting delay is not necessarily steady. A device can absorb bursts into available free space and perform reclamation later, producing periods of low latency followed by periods with more internal work.
Large volatile or non-volatile caches can further separate host completion timing from eventual NAND activity. Short benchmarks may therefore capture cache absorption rather than steady-state media behavior.
Steady state exposes the media-management boundary
A freshly erased or lightly used SSD can have abundant free blocks and little immediate garbage-collection pressure. Its early write performance may not represent sustained operation after the logical address space has been populated and repeatedly updated.
Once invalid pages, valid pages, free blocks, and incoming writes reach a recurring balance, garbage collection becomes part of the continuing write path. At that point, host write rate, logical utilization, deallocation behavior, data lifetime distribution, over-provisioning, and controller policy jointly determine internal movement.
Write amplification is therefore a boundary between the host’s logical block model and NAND’s physical erase model. The host reports logical writes; the controller must continuously transform those writes into page placement, relocation, and block erasure while preserving current data.