Write Amplification

This card uses a byte ratio: media bytes programmed divided by host-written bytes over the same interval. State the boundary before quoting a value. SNIA's dictionary also defines an operation-count WAF; a larger number of smaller operations is not necessarily a larger number of bytes. [1] [3]

NAND updates usually go to another physical location. The superseded copy becomes invalid, but erasing its block may still require copying other, live pages elsewhere. Those relocation writes add to the new host data. Garbage collection reclaims space; wear leveling can move data to distribute erase wear even when that is not the cheapest reclamation choice. [2]

The outcome depends on placement and workload. Agrawal and colleagues examine cleaning policies and the conflict between reclamation efficiency and wear balance. Their simulator explains mechanisms and tradeoffs; it does not give a universal WAF for current retail SSDs. [2]

Count what the counter actually covers. Payload relocation, mapping metadata, parity and other controller work must not silently appear in one numerator and disappear from another. Copyback can reduce transfers across the flash interface without reducing programmed bytes; lower traffic or faster execution is not itself evidence of lower byte WAF. [2]

Device-side compression can reduce the media payload enough for byte WAF to fall below one. Compression in the host instead changes what reaches the drive. The SNIA computational-storage case study supports the possibility, not a claim that all SSDs compress data or achieve its performance results. [5]

NVMe SMART / Health Data Units Written counts host data, excluding metadata. It is not a NAND-program counter, so it cannot supply the numerator by itself. Base 2.3 expresses it in rounded groups of 1,000 512-byte units; zero means not reported. [4]

NVMe does have separate standardized media-write accounting. Where supported and reported, the Endurance Group Information log's Data Units Written excludes internal controller writes, while Media Units Written includes them, such as garbage collection. Both use rounded billion-byte units. Compare deltas from the same group and interval; do not mix a group numerator with an unrelated controller or namespace denominator. [4]

Rounded counters limit short-window precision. Unchanged nonzero readings do not prove there were no writes, and a one-unit increase is not an exact one-unit transfer. Keep raw start/end values, collection times, scope and reset history; an unavailable zero is not a measured zero. These are consequences of the encoding, not a device experiment. [4]

Worked example: take supported, unsaturated Endurance Group counters from the same group and interval, with no reset. Data Units Written rises from 100 to 102 and Media Units Written from 200 to 204. Their delta ratio is 2, but rounding alone allows host bytes strictly between 1 and 3 billion and media bytes strictly between 3 and 5 billion. The compatible byte ratio is therefore strictly between 1 and 5, not an exact 2. These are arithmetic bounds, not a confidence interval or a device measurement. A longer observation window reduces this quantization uncertainty. [4]

For disjoint, compatible intervals, add the media-write deltas and host-write deltas separately, then divide. Averaging the interval ratios without their host-byte weights can produce a different answer. If a window has no host writes, its local ratio is undefined; do not discard its internal writes when forming a longer-window total. [7]

A read-only host workload need not mean read-only NAND activity: refresh or read reclaim can program data. When the host-write denominator is zero, retain physical-write volume and elapsed time rather than reporting a zero WAF. [8]

Application, database and device ratios use different boundaries. A database may rewrite logical records during compaction before its output reaches the device. Treating that output as both application input and host traffic hides work between layers. A product of layer ratios is meaningful only when the intermediate byte quantities, observation window and scope match. [6]

A useful report names the model and firmware, workload and preparation, written address range, compression conditions, interval, raw counters and included write classes. Report latency and durability evidence separately. Use the linked analyses for finite-block examples, over-provisioning denominators and read-reclaim mechanisms.