Why an SSD's Empty-Drive Speed Does Not Describe Repeated Writes

A replacement write does not imply a physical overwrite

The host addresses logical storage blocks. Inside the SSD, those addresses are mapped to NAND locations; Micron's mapping discussion explains that distinction. A replacement can therefore occupy a different physical location while the host continues using the same logical address. The previous contents can become invalid without their cells immediately becoming reusable.

NAND cannot simply overwrite an already programmed location with arbitrary replacement data. KIOXIA's September 20, 2023 research opening distinguishes page-level reads and writes from block-level erasure. A physical erase block contains multiple pages, so reusing one old page involves a larger erase unit.

Kingston's March 2019 explanation describes the resulting reclamation work: preserve still-valid data in another block, erase the original block, then reuse its space. This is the garbage-collection mechanism relevant to the comparison.

References: [M1][K1][K2]

Filesystem free space is not an erased-block inventory

A filesystem's free-space figure describes space available to that filesystem. It does not identify which physical NAND blocks are already erased. Kingston distinguishes a filesystem marking a file deleted from the controller's physical reclamation work. Two drives showing the same filesystem free space have not thereby established the same internal state.

A host can also communicate that particular logical data is no longer needed, when that information is supported and conveyed through the storage path. KIOXIA's April 2023 managed-flash brief explains how such information can reduce the need to relocate unwanted data. Marking data invalid and physically erasing its cells remain different events.

Consequently, a report saying that files were deleted has not established a fresh physical write state. Nor does deallocation information establish secure sanitization. The separate encryption and cryptographic-erasure guide linked below addresses the evidence needed for that disposition decision.

References: [K1][K3]

Two four-page blocks: one copy or three copies

Consider two invented physical NAND blocks, A and B, each containing exactly four equal-sized pages. A has one valid page and three invalid pages. B has three valid pages and one invalid page. Neither contains an already-erased page. Stipulate sufficient erased destination space elsewhere to preserve all valid data, with no other activity or accounting overhead.

Reclaiming A requires copying its one valid page to one destination page before erasing all four pages of A. The erase produces four reusable pages in A, but the copy consumed one previously reusable destination page. Across the source and destination space together, the net gain is 4 - 1 = 3 reusable pages.

Reclaiming B requires three page copies, consuming three erased destination pages before its four-page block can be erased. The net gain is 4 - 3 = 1 reusable page.

Each case erases a four-page block. What differs is how much still-needed data must move and how much destination space that movement consumes. Counting only the four newly erased source pages would hide the difference. The example makes internal preservation work visible even though none of those copies represents new host data.

This is bookkeeping with stipulated pages, not a firmware selection policy, whole-drive write-amplification estimate, endurance prediction or measured speed ratio. Three copies versus one does not imply that one real drive is three times slower. It shows why an erased-block count and the work required to replenish that count are different quantities.

References: [K3]

Reclamation can compete with new host writing

Copies of valid pages require reads and additional programming; reclaiming the original block requires an erase. These operations use device resources alongside the work needed for incoming data. KIOXIA's research opening identifies erase activity as a possible cause of reduced throughput and longer latency. Its managed-flash brief also explains that moving more valid data can delay incoming work.

The buyer should therefore distinguish a result obtained while ample erased space was available from one obtained while the drive was replenishing that space under continued writing. The latter includes internal work the former may expose much less.

This mechanism does not specify when a particular controller performs reclamation or how much of its cost reaches the host. Firmware scheduling and the workload affect that result. Kingston's historical controller policy and KIOXIA's BOOSTER research results cannot be assigned to an unspecified SSD.

References: [K2][K3]

Fresh state and steady state answer different questions

SNIA's September 2011 glossary describes fresh-out-of-box storage as having ample pre-erased blocks. An empty filesystem on a previously written drive does not establish that condition. A fresh-state measurement can accurately describe its stated starting circumstances while leaving later repeated writes unmeasured.

SNIA's historical test terminology defines preconditioning as preparation for steady-state measurement and distinguishes preparation using the test workload itself. Its steady-state context concerns relatively stable behavior for the specified workload and measured quantity. A report needs to explain the basis for that description; elapsed time or a workload name alone does not establish it.

For a recurring rewrite requirement, evidence after relevant workload history addresses a different buying question from an initial write into fresh space. Neither state label guarantees every future result. A change in workload also changes what the existing evidence establishes. These historical definitions help interpret supplied reports; they are not a current certification claim or a conditioning procedure.

References: [S1][S2]

Compare two fictional reports with the same workload name

Suppose a buyer needs repeated random writing for a recurring scratch-data stage. Two fictional reports name the same workload W. Stipulate that request size, access distribution, concurrency, measurement boundary and other measurement conditions match apart from the declared prior-write state. Every number and report detail here is invented; no test has been performed.

Report A gives 600,000 IOPS and explicitly identifies a fresh starting state. It supplies no result after prior writing under W. Report B gives 450,000 IOPS after declared preparation under W. It identifies its steady-state criterion, supporting observations and measurement window, and reports response evidence from that same operating point.

A has the larger number, but only B supplies the stipulated workload-conditioned basis relevant to continued rewriting. That makes B's evidence more directly applicable to this question. Suitability still depends on whether its result meets the buyer's actual rate and response requirement. A remains an accurate statement of its stipulated fresh-state result; the missing later-state evidence does not prove a defective drive.

The comparison cannot establish A's conditioned rate, B's fresh rate or a percentage penalty caused by garbage collection. These are different fictional offers, and the two reported values do not isolate that mechanism. The useful follow-up for A is evidence under the relevant prior-write state, rather than another fresh-state headline or a claim that the filesystem was empty.

A larger TBW rating would not close that gap either. Rated write quantity cannot describe the current supply of erased NAND blocks or establish a sustained rate. The linked endurance guide owns the DWPD/TBW comparison and mapping-unit amplification example; the performance guide owns IOPS, throughput and latency interpretation.

Carry the selected hardware into its physical installation

When the storage and platform owners have selected the equipment, Leviathan's rack-and-stack service can address an agreed package of equipment mounting, IT cabling, labeling and physical installation records. The service page and project inquiry are linked below.

For that discussion, share a nonconfidential equipment summary, released rack and connection plans, site constraints and required physical handoff. Identify the storage-selection owner and equipment supplier. Workload evaluation, SSD conditioning and firmware tuning require separately assigned scope; a physical installation record does not establish a storage-performance result.

References: [L1]