HardDiskIndexFAQ

SSD endurance: TBW, DWPD, and how long an SSD really lasts

By Harry Saarinen ·

In this guide: 17 sections
  1. How long an SSD lasts at your write rate
  2. What to check before you buy a used SSD
  3. The wear is readable off a used drive before you pay for it
  4. Percentage used, available spare and media errors answer three different questions
  5. The trap: “8TB written” is an odometer, not a capacity
  6. TBW and DWPD are the same number twice
  7. Who genuinely does need to care
  8. Write amplification, and why it is never one
  9. SLC caching buys speed with a second write of every byte
  10. Power-loss protection changes what a synchronous write costs
  11. The wear unit is a physics event, not a counter
  12. JESD218 is four requirements at once, and most ratings are extrapolated
  13. What end of life actually does, and who implements it
  14. What the endurance-to-destruction runs proved, and what they did not
  15. Unpowered retention, and why an SSD is a bad archive
  16. Data center QLC and the cheap QLC that made its reputation are different products
  17. What has to be measured, and how this site records it

An ordinary laptop or desktop SSD will not wear out before you replace it: a 1 TB drive rated for 600 TBW takes 55 years to reach that rating at 30 GB of writes a day. For most buyers the endurance rating is not what ends an SSD’s life - the capacity is. You will replace a 1 TB drive because 1 TB stopped being enough, decades before you write the 600 terabytes its warranty covers. The exceptions are the jobs that write all day - a video scratch disk, a camera recorder, a write-heavy database, a ZFS log or cache device, a build server - where the same drive can be used up in months or a few years, and for those you buy a much bigger drive or an enterprise one on its DWPD rating. Either way, an SSD reports its own wear: on Windows CrystalDiskInfo shows it, so you can ask a seller for it before you pay.

How long an SSD lasts at your write rate

The reference drive for the sums in this guide is a typical 1 TB consumer TLC NVMe drive, rated 600 TBW with a five-year warranty. TBW, terabytes written, is the total the warranty lets you write. Divide the rating by what you write in a day:

Daily host writes Roughly what that is Days to 600 TBW Years
10 GB A laptop: browsing, mail, office documents 60,000 164
30 GB An ordinary desktop 20,000 55
50 GB A gaming desktop: frequent installs, large downloads 12,000 33
100 GB Developer workstation, VMs, big builds 6,000 16
500 GB Video scratch disk, sustained 1,200 3.3
2 TB Database, capture, cache tier 300 0.8

The five-year warranty expires first in every row above the fifth. In the top four rows the drive is obsolete, dead of something unrelated, or thrown out for being too small long before the flash is tired, and the correct amount of time to spend comparing TBW figures for that machine is none. Who genuinely does need to care covers the bottom two.

Other drives scale from here, because consumer ratings follow capacity. Samsung’s 870 EVO and Sandisk’s SN7100 are rated 300 TBW at 500 GB, 600 at 1 TB and 1,200 at 2 TB, so halve the years for 500 GB and double them for 2 TB. QLC runs lower: Samsung’s 870 QVO is 360 TBW at 1 TB, on a three-year warranty.

Do not trust the row you think you are in. Read the host-write counter twice, a week apart, and divide; on Windows CrystalDiskInfo prints it as Total Host Writes. Most people guess high: a 60 GB game install is memorable, and three hundred days of 4 GB background writes are not.

Two qualifications, in opposite directions. On a consumer drive TBW is a warranty limit, not a failure point - the destruction runs took consumer drives far past their ratings (an Intel rated just under 22 TB wrote about 750), which is a handful of samples from 2015 rather than a promise. And crossing it voids the warranty whether or not the drive works. Western Digital’s endurance white paper is blunt about the other side: “once the NAND inside the SSD has exhausted its rated endurance there is no longer any expectation that the data will be retained for a specific length of time”. Retention past the rating is not shortened. It is unspecified, which is a different and worse thing.

What a laptop, a desktop, games, photos and video really write

Playing a game, viewing a photo and watching a film are reads, and reads do not wear flash. What counts is what lands on the drive, and the quickest sum is the rating divided by the terabytes you write in a year. First copies, on the reference drive:

Habit Per day Per year Years to 600 TBW
Games: a new 100 GB install every week 14 GB 5.2 TB 115
Photos: 2,000 RAW files a week at about 75 MB, the size Apple gives for a 48 MP ProRAW 21 GB 7.8 TB 77
Video: an hour of 4K a day from a camera at 100 Mbit/s 45 GB 16.4 TB 36.5
Video: an hour of UHD 30p a day in ProRes 422, 265 GB by Apple’s table (April 2022) 265 GB 96.7 TB 6.2

Real writes run higher. Valve’s Steamworks documentation gives the case of a 10-byte change inside a 25 GB pack file: Steam builds a whole new copy beside the old one, so a tiny patch writes about 25 GB. Editors add proxies, caches and exports. Only the last row comes near the warranty, and there capacity arrives first: a 1 TB drive holds under four hours of that footage, and a 4 TB drive rated 2,400 TBW, as the 870 EVO and SN7100 are, lasts about 25 years at the same pace.

Worked examples, with illustrative counters. A laptop. A 500 GB drive rated 300 TBW shows 11 TB of Total Host Writes after three years. That is about 3.7 TB a year, and 300 ÷ 3.7 is about 80 years.

A gaming desktop. A 2 TB QLC drive rated 720 TBW, as the 870 QVO is, shows 30 TB after three years: 10 TB a year, 72 years to the rating. The three-year warranty is what runs out, and the slowdown on long writes is what you would notice.

A used drive for either. A 1 TB NVMe drive rated 600 TBW arrives showing 60 TB written and a Percentage Used of 9. At 10 TB a year the remaining 540 TB is 54 years. Wear is not the risk here; the warranty is, because you cannot count on one with a used drive.

What to check before you buy a used SSD

The checks in the order they pay off; later sections explain each. The first seven are for any laptop or desktop drive.

  1. Know what the machine takes. The maker’s document says what fits; which SSD to buy goes through M.2, SATA and length.
  2. Measure your write rate before you read any rating. Read Total Host Writes twice, a week apart, and find your row in the write-rate table. In the top four rows, ignore endurance and compare price per terabyte at NVMe SSDs, SATA SSDs or all SSDs; Trends shows how that price has moved.
  3. Get the report before you pay, not after. The whole CrystalDiskInfo window with the serial showing, or the raw output: nvme smart-log plus nvme endurance-log on NVMe, smartctl -x -l devstat on SATA, smartctl -x on SAS. A “100 percent health” screenshot is a tool’s rounded summary of a field it guessed the meaning of.
  4. Convert the host writes yourself and check them three ways - against the rating, the power-on hours and the interface bandwidth. Divide by 1,953,125 for NVMe data units and 1,953,125,000 for 512-byte LBAs. Anything implying more than 600 MB/s sustained on SATA is a wrong unit, not a remarkable drive.
  5. Read the direction of every wear attribute before its value. A raw 3 on attribute 202 is three percent used; a 3 on a drive reporting life remaining is three percent left. Where the drive offers the standardized Percentage Used Endurance Indicator - devstat page 7 on SATA, log page 0x11 on SAS - use that.
  6. Leave a consumer drive 20 percent empty, with working TRIM. On a 1 TB drive that is the 37.4 percent over-provisioning of an enterprise mixed-use or write-intensive part, for free, and it cuts the write amplification bound by two thirds. The check is one command on Windows or Linux.
  7. Compare fixed prices. The tables open on prices you can pay; tick “Buy It Now only” beside a table to set aside the auctions that also carry a Buy It Now price.

For a job that writes all day, or a used enterprise drive:

  1. In the bottom two rows of that table, buy bigger, or shop enterprise on the capacity number. Check the model for the exact rating. Start from the SSD listings or 1.6 TB parts.
  2. On a data center NVMe drive, run the OCP plugin. nvme ocp smart-add-log gives Capacitor Health and the erase-count spread. Capacitor Health well below 100 on an enterprise SSD is a disqualification, not a discount.
  3. Match the drive to the write size. A large QLC part rated on 16 or 32 KiB IU-aligned writes is superb for media and poor for a busy database; nvme id-ns shows the preferred write granularity.
  4. For a SLOG or any sync-write target, buy power-loss protection. nvme id-ctrl reporting vwc of 0 is the one-command test.

On November 2, 2026 a used 1 TB M.2 NVMe drive had a median Buy It Now price of $140 in this site’s US catalog against $249 new, item price only: 90 percent of the writing for 56 percent of the price, if the report shows a tenth of the rating used. A used enterprise SSD at 10 percent of its endurance has more writing left than a new consumer drive has in total, but it is no longer reliably cheaper: a used 3.84 TB 2.5-inch SATA part’s median was $186 per terabyte against $203 for a new 4 TB, and a used 1.92 TB part’s $206 against $195 for a new 2 TB. That trade is only visible if you read the odometer.

The wear is readable off a used drive before you pay for it

This is the part that decides whether a second-hand SSD is a bargain, and how you read your own. The general used-drive checks live in buying used drives on eBay; this section is the wear counters, Windows first. The dumps are illustrative: one consistent set of example figures, not a capture from one drive.

On Windows: CrystalDiskInfo, then the maker’s tool

CrystalDiskInfo is the first way: free, with NVMe support on Windows 10 or later. Connect the drive directly, because a USB adapter can hide SMART. Read three parts of its window:

  • Health Status. On NVMe the percentage is 100 minus Percentage Used, in its source code, so 97 means 3 percent used. On SATA it reads a different attribute per maker: B1, Wear Leveling Count, on a Samsung client drive; CA, Percent Lifetime Used, on a Micron. Caution appears at 10 percent life or less on many models. Its author (July 18, 2026) calls the figure an estimate from the maker’s rating that does not convert into days or months, and says to back up first.
  • Total Host Writes, in gigabytes, where the drive reports it. Divide by the rating: 30 TB on a drive rated 600 TBW is 5 percent. On NVMe the code counts its gigabytes in 1,024s, so the figure reads about 7 percent below the same writes in a rating’s decimal units.
  • The attribute table. NVMe lists Percentage Used, Available Spare and Media and Data Integrity Errors, which answer different questions. IDs and raw values are hexadecimal, and reading a drive on Windows shows how to read across.

The maker’s tool comes next, on its own make. Samsung Magician shows Drive Health and Total Bytes Written on Samsung drives only, and says features vary on Samsung OEM drives, such as one pulled from a laptop. Sandisk’s Dashboard, per the SN7100 sheet, monitors health and firmware; Crucial Storage Executive monitors temperature and overall health; Kingston describes its SSD Manager as a health and firmware application. A green verdict has limits, as the destruction runs showed with Samsung’s own software.

Windows has its own reading, for NVMe only. On Windows 11, open Settings, System, Storage, Advanced storage settings, Disks & volumes, then Properties on the drive, for Estimated remaining life and Available spare; Windows 10 has it under Manage disks and volumes. Microsoft says it does not monitor SATA SSDs, and that using up the estimate does not by itself mean the drive has failed.

smartctl is the second way, and the rest of this section uses it and Linux’s nvme tool, because they print what a summary hides. It runs on Windows too, without sudo, from an administrator prompt, and smartctl -a on an NVMe drive prints the fields below.

NVMe: the counters are standard, the two headline fields are not

The SMART / Health Information log page, 02h, is defined by the NVMe specification. The counters mean the same thing on every drive. The two fields everybody actually reads do not.

$ sudo nvme smart-log /dev/nvme0
critical_warning                    : 0
available_spare                     : 100%
available_spare_threshold           : 10%
percentage_used                     : 3%
data_units_written                  : 47,281,940
host_write_commands                 : 688,192,404
power_cycles                        : 412
power_on_hours                      : 21,904
unsafe_shutdowns                    : 38
media_errors                        : 0

percentage_used is, in the specification’s words, “a vendor specific estimate of the percentage of NVM subsystem life used”. At 100 the estimated endurance is consumed “but may not indicate an NVM subsystem failure”, and the value may go past 100. Western Digital says its own figure is “based on the actual PE cycles of the NAND utilized in the drive, not based on host writes”, which is why it can diverge from your arithmetic and is the more trustworthy of the two.

data_units_written counts host data in thousands of 512-byte units. It does not include metadata, and a value of 0 means the counter is not reported, not that nothing was written. Convert it, then cross-check it against the hours:

TB written = data_units_written ÷ 1,953,125
           = 47,281,940 ÷ 1,953,125
           = 24.2 TB
24.2 TB ÷ 21,904 hours     = 1.1 GB per hour = 26.5 GB per day
26.5 GB ÷ 1,000 GB capacity                 = 0.027 DWPD

Against a 600 TBW rating that is 4.0 percent of the budget, near enough the 3 percent the controller reports, and a drive averaging under three hundredths of a drive write a day for two and a half years sits on the ordinary desktop row of the table, whatever the seller said.

media_errors should be zero, but it is not purely a wear signal: the specification counts “uncorrectable ECC, CRC checksum failure, or LBA tag mismatch”, so a CRC failure on the link counts too. critical_warning is a bit field: bit 3, All Media Read-Only, is the end-of-life transition, and bit 4, Volatile Memory Backup Failed, is the power-loss protection failure flag.

The endurance group log is the page nobody runs

The Endurance Group Information log page, 09h, is in the standard NVMe specification, not an OCP or vendor extension, and nvme endurance-log reads it. Its Media Units Written “includes data bytes written by both the host and the controller (e.g., due to garbage collection)”; its Data Units Written counts host writes only. Dividing one by the other gives lifetime write amplification straight out of the standard:

media units written  53,258 × 10^9 bytes = 53.3 TB
data  units written  24,208 × 10^9 bytes = 24.2 TB

lifetime WAF = 53.3 ÷ 24.2 = 2.2

Its Endurance Estimate is the drive stating its own TBW, “assuming a write amplification of 1”, in units of 1,000,000,000 bytes: a value of 600,000 is 600 TB, and one that does not match the model’s datasheet is worth a second look. A zero in any of these fields means not reported.

OCP data center drives publish the rest

Data center drives built to the OCP Datacenter NVMe SSD Specification add an extended SMART log at C0h, read with nvme ocp smart-add-log: Physical Media Units Written, Capacitor Health, Media Dies Offline, Unaligned I/O, and the minimum and maximum user data erase counts. That minimum and maximum are the most informative pair on a used drive. A narrow gap at a high average is a drive used hard and managed well; a wide gap is a drive whose hot blocks took the punishment.

OCP compliance is a claim, not a guarantee of every behavior. Micron says its 6550 ION meets most but not all of the requirements, and nvme ocp unsupported-reqs-log is where a drive lists the ones it does not meet.

SATA attributes are a minefield, and the standard names the reason

SMART attribute numbering above the low IDs is not standardized. The smartctl manual page says it in one sentence: “the conversion from Raw value to a quantity with physical units is not specified by the SMART standard”. The names smartctl prints come from a per-model database, not from the drive. CrystalDiskInfo shows the same IDs in hexadecimal: 202 is CA, 233 is E9, 246 is F6.

$ sudo smartctl -A /dev/sda
ID# ATTRIBUTE_NAME          FLAGS    VALUE WORST THRESH FAIL RAW_VALUE
  5 Reallocated_Sector_Ct   -O--CK   100   100   010    -    0
  9 Power_On_Hours          -O--CK   095   095   000    -    21904
173 Ave_Block-Erase_Count   -O--CK   097   097   000    -    48
202 Percent_Lifetime_Remain ----CK   097   097   001    -    3
246 Total_LBAs_Written      -O--CK   100   100   000    -    47281940512

What those identifiers carry across that database:

ID Default meaning with no database match What it is on other drives
202 Data_Address_Mark_Errs Percent_Lifetime_Remain on Crucial and Micron; Exception_Mode_Status on Samsung data center parts
231 Temperature_Celsius SSD_Life_Left
233 Media_Wearout_Indicator Flash_Writes_GiB, NAND_Writes_1GiB, Lifetime_Nand_Writes, Total_LBAs_Written, Percent_Lifetime_Remain, Remaining_Lifetime_Perc
241 Total_LBAs_Written, in 512-byte LBAs Host_Writes_32MiB, Host_Writes_GiB, Lifetime_Writes_GiB
246 none Total_LBAs_Written on Micron and Crucial; SLC_Writes_32MiB on Silicon Motion controllers; Timed_Workld_RdWr_Ratio on Samsung data center drives

Attribute 233 is the most overloaded, appearing as Media_Wearout_Indicator in twenty families and as at least ten other things elsewhere, several of them byte counters rather than percentages. Attribute 246 is the trap the dump above is standing in: on Micron and Crucial it is host LBAs and the conversion below is right, while on a Silicon Motion controller it is SLC_Writes_32MiB, a NAND-side counter in 32 MiB units. A bare integer in the RAW_VALUE column looks identical either way.

Read the direction before the number. On Crucial and Micron client SSDs the drive database records attribute 202 as “norm = max(100-raw,0); raw = percent_lifetime_used”, so the raw column is percent used and the normalized column, Current in CrystalDiskInfo, is percent remaining - in the dump, VALUE 097 and raw 3 are the same fact. Reading the wrong column turns a nearly-new drive into a dead one. Micron and Crucial SATA drives also expose 247 Host_Program_Page_Count and 248 Bckgnd_Program_Page_Cnt, so (247 + 248) ÷ 247 is their write amplification.

The unit conversions, and the sanity check that catches a wrong one

512-byte LBAs → TB   raw ÷ 1,953,125,000
32 MiB units  → TB   raw ÷ 29,802
1 GiB units   → TB   raw ÷ 931.3
1 GB units    → TB   raw ÷ 1,000

So the 47,281,940,512 sectors in the dump are 24.2 TB - the same drive as the NVMe example, because an NVMe data unit is exactly 1000 sectors. Check every conversion against power-on hours, because a wrong unit shows at once:

24.2 TB over 21,904 hours    = 1.1 GB/hour        plausible
same raw read as 32 MiB units:
47,281,940,512 ÷ 29,802      = 1,586,000 TB
1,586,000 TB ÷ 21,904 hours  = 72 TB/hour
72 TB/hour ÷ 3,600 s         = 20 GB/second       impossible on SATA

SATA 6 Gbit/s tops out near 600 MB/s. The unit assumption was wrong, not the drive.

SATA Device Statistics and SAS log pages are standardized

On SATA, prefer the Device Statistics log (General Purpose Log address 0x04) to the attribute table wherever the drive supports it:

$ sudo smartctl -l devstat /dev/sda
Page Offset Size         Value  Description
  1  0x010  4            21904  Power-on Hours
  1  0x018  6      47281940512  Logical Sectors Written
  1  0x020  6        688192404  Number of Write Commands
  7  0x008  1                3  Percentage Used Endurance Indicator

Page 7 holds exactly one entry, Percentage Used Endurance Indicator, which counts up like the NVMe field and means the same thing on every drive that reports it. Sectors written divided by write commands is the mean write size - 68.7 sectors, about 34 KiB, here; near 4 KiB on a drive with a large indirection unit is a warning.

SAS SSDs keep a standardized wear figure too, just not as a SMART attribute: SCSI log page 0x11, parameter code 0x0001, is the “percentage used endurance indicator”, from 0 as new to 100 at the maker’s projected end of life. Host writes are in log page 0x19, and smartctl -x on a SAS device reads both, which makes smartctl -x the one command to ask a SAS seller for. Enterprise drives covers the rest of what a SAS pull needs.

Backing write amplification out of two attributes

Where a drive gives only an erase count and a host-write count, the two still estimate write amplification: 48 erase cycles on the example drive’s 1.10 TB of raw NAND is 52.8 TB, and 52.8 ÷ 24.2 TB of host writes is 2.2. Treat it as an estimate: the erase count may or may not include pSLC blocks.

Percentage used, available spare and media errors answer three different questions

These get conflated, including by tools that summarize them into a single health percentage. Percentage used is a wear estimate: it counts up, it may pass 100, and the drive may keep working past it. Available spare is an inventory of the reserve block pool, falling as blocks are retired; a drive can be at 10 percent used and 50 percent spare, which means it has retired a lot of blocks early for reasons that are not wear. Media errors are an event count of things the correction engine could not fix, some of them nothing to do with flash.

Only an OCP drive promises how the first two relate: ENDUD-3 makes Percentage Used “track linearly with bytes written”, and EOL-6 makes it reach 100 percent before Available Spare falls below its threshold. Nothing guarantees either on a consumer drive.

MTBF and AFR are not endurance figures. Western Digital calls MTBF “mostly unrelated to endurance”: a failure rate predicts random electrical failure across a population, while an endurance rating predicts one drive’s ability to read back what was stored on it. Kioxia’s CD8 brief carries a 2,500,000-hour MTTF with the vendor’s own caveat that it “is not a guarantee or estimate of product life”.

The trap: “8TB written” is an odometer, not a capacity

A listing titled “NVMe SSD - only 8TB written!” is not an 8 TB drive. The number is host writes - wear - and the seller is quoting it because it is low. Against a 600 TBW rating 8 TB is 1.3 percent of the budget, and if the same drive reports 12,000 power-on hours - 500 days - it averaged 16 GB a day, or 0.016 DWPD on a 1 TB drive.

What the title says What it means
“600 TBW” The rating. What the drive is licensed to do
“8TB written”, “8TB host writes” The odometer. What it has already done
“8TB”, alone, near a model number Almost certainly capacity

A seller who quotes host writes has read the SMART data, which is a mild positive signal. A seller quoting “100 percent health” has read a vendor tool’s rounded summary, and the destruction runs show what that is worth on a drive with thousands of reallocated sectors. Where a title gives both a capacity and a wear figure, this site reads the one marked “written” or “host writes” as wear and takes the size from the other, so the price per terabyte is figured on the drive’s capacity rather than its odometer.

TBW and DWPD are the same number twice

Two ratings, one fact, and every drive is sold with whichever one flatters it.

TBW - terabytes written - is the total volume of host writes the rating covers, normally as a warranty limit. DWPD - drive writes per day - is how many times you could overwrite the whole drive daily for the length of that warranty and still land inside the total. Consumer drives are sold on TBW because the number is large; enterprise drives on DWPD because it describes the duty cycle the buyer is choosing between. On a laptop or desktop you can stop at TBW.

The conversion is division:

DWPD = TBW ÷ (capacity_TB × 365 × warranty_years)
TBW  = DWPD × capacity_TB × 365 × warranty_years

On the reference drive that is 600 ÷ (1 × 365 × 5) = 0.33: a third of the drive, every day, for five years. Where a vendor publishes both numbers they agree closely: Solidigm’s D7-P5810 brief gives an 800 GB SLC drive at 50 DWPD over five years and 73 PBW, and 50 × 0.8 × 1825 is exactly 73,000. So a missing number is recoverable, as on Kioxia’s CD8 series briefs, which publish DWPD and a five-year warranty and no TBW.

The ratings drives actually carry, converted both ways:

Drive Rating Arithmetic The other number
1 TB consumer TLC, 5 yr 600 TBW 600 ÷ 1825 0.33 DWPD
2 TB consumer QLC, 5 yr 400 TBW 400 ÷ 3650 0.11 DWPD
3.84 TB read-intensive, 5 yr 1 DWPD 1 × 3.84 × 1825 7,008 TBW (7.0 PBW)
3.2 TB mixed-use, 5 yr 3 DWPD 3 × 3.2 × 1825 17,520 TBW (17.5 PBW)
800 GB SLC, 5 yr 50 DWPD 50 × 0.8 × 1825 73,000 TBW (73 PBW)

The spread between the cheapest consumer QLC part and a genuinely write-intensive one is a factor of four hundred and fifty in DWPD.

Consumer vendors scale TBW linearly with capacity, so DWPD is a property of the product line. Enterprise families do not follow that rule. Within one family DWPD rises with capacity, because the bigger part has more parallel dies to spread the same duty cycle across. Solidigm’s D5-P5336 product brief, one family on one 192-layer QLC die, runs from 0.42 DWPD at 7.68 TB (5.9 PB written) to 0.60 at 122.88 TB (134.3 PB), forty-three percent more daily duty cycle.

Enterprise endurance tiers, and what the capacity tells you

Every SSD has some over-provisioning - flash the controller keeps for itself - for free, because 1024 GiB of flash is 1,099.51 GB sold as 1,000 GB. Enterprise parts hide more, and how much shows in the capacity:

Marketed capacity Raw NAND Over-provisioning Positioning
512 GB 512 GiB 7.37 percent Consumer
1 TB / 2 TB / 4 TB 1024 / 2048 / 4096 GiB 9.95 percent Consumer
480 / 960 / 1,920 / 3,840 GB 512 / 1024 / 2048 / 4096 GiB 14.53 percent Enterprise read-intensive
400 / 800 / 1,600 / 3,200 GB 512 / 1024 / 2048 / 4096 GiB 37.44 percent Enterprise mixed-use or write-intensive

A 400, 800, 1,600 or 3,200 GB enterprise SSD is mixed-use (3 DWPD) or write-intensive (10 DWPD or more), and only the model’s own datasheet says which. A 1.92 or 3.84 TB drive is usually read-intensive, about 1 DWPD. Kioxia’s CD8-V ships 800 GB to 12,800 GB at 3 DWPD (product brief, Rev. 1.2, 2023) on the same 112-layer BiCS TLC as the CD8-R at 960 GB to 15,360 GB and 1 DWPD: same flash, same warranty, different amount hidden. A 1.6 TB SAS SSD and a 1.92 TB one can be the same silicon, and the capacity ladder has more models and the rest of buying one used.

What the rating is a guarantee of, and what it is not

The conversion assumes three things: capacity is the marketed decimal figure (drive capacity explained), both ratings count host writes rather than what lands on the NAND (write amplification), and both are measured against a defined workload (JESD218).

TBW is not a wear counter. It is the write volume at which four separate guarantees are still simultaneously true. Seagate’s technical paper TP618, which reproduces JESD218 Table 1 with JEDEC’s permission, lists them: the drive must still present its advertised capacity, still meet the uncorrectable bit error rate for its class, still be inside the functional failure requirement, and still retain data with the power off for the required time at the required temperature. A drive can be past its TBW and still writing happily, because the guarantee that broke first was probably retention, which is invisible until you unplug the drive and come back later.

At least one vendor makes the counter contractual. Crucial’s warranty claim window is three or five years from purchase, depending on the product, or “before writing the maximum total bytes written (TBW) as published in the product datasheet and as measured in the product’s SMART data, whichever comes first”. On that drive the SMART field is the legal instrument, which is a good reason to know exactly what it counts.

Who genuinely does need to care

A short list of jobs write all day, and on the same 600 TBW drive the years come down fast:

Job Host writes per day Time to 600 TBW
Video capture or a CI build server about 500 GB about 3.3 years
Recorder for sixteen cameras at 4 Mbit/s each 691 GB 2.4 years
L2ARC fed flat out at the current OpenZFS default’s nominal rate about 2.9 TB about 7 months
ZFS SLOG taking 200 MB/s of sync writes, eight hours a day 5.76 TB 104 days

Video is the one such job on a desktop, and the plain answer is capacity: a 4 TB drive rated 2,400 TBW takes about 13 years to reach its rating at 500 GB a day. Camera recorders never stop writing and nobody watches the drive; sizing one covers the hard drives rated for that load. Databases write small and random, which is the enterprise test workload run on a drive rated against the client one, so the NAND wears faster than the host counter implies. CI build servers write a working tree and an object cache per job and then discard them, exactly the turnover that drives write amplification up.

SLOG. A separate log device takes the pool’s synchronous writes - NFS with sync, iSCSI without writeback, databases opening files O_SYNC. The OpenZFS manual page says the size threshold for large writes is ignored when a SLOG is present, so every synchronous write lands on the log device. It also needs power-loss protection.

L2ARC. Every OpenZFS release from 2.2.7 (December 2024) through 2.4.4, the latest (August 21, 2026), sets l2arc_write_max to 32 MiB per feed, with feeds a second apart (OpenZFS zfs.4 manual page for each release, checked October 15, 2026). Fed flat out that is 2.9 TB a day, or 207 days to 600 TBW. That is the nominal rate, not a forecast: a feed writes only what the ARC is evicting. At the old 8 MiB default, which Ubuntu 24.04 LTS still ships on OpenZFS 2.2.2, the same sum gives about 2.3 years. NVMe caching covers both devices and whether an L2ARC earns its place at all.

For the server jobs, buy an enterprise drive on DWPD, where the over-provisioning above is what you pay for.

Write amplification, and why it is never one

The controller cannot overwrite a page in place. Flash is written in pages - 16 KB is typical - and erased only in whole blocks of hundreds of pages, so reusing a block that mixes stale and valid data means copying the valid pages elsewhere first. Write amplification factor is NAND bytes written divided by host bytes written, and those copies are why it is never one.

If garbage collection erases a block in which a fraction u of pages are still valid, write amplification is 1 ÷ (1 − u). Under uniformly random writes every block is on average as full as the drive is, so u is 1 ÷ (1 + OP), where OP is over-provisioning, and the factor comes to (1 + OP) ÷ OP.

That is an upper bound, because it assumes garbage collection picks a block at random. Real greedy collection picks the emptiest block and does better, but its published curves fall the same way. The bound, tabulated:

Over-provisioning Non-selective WAF bound Saved by the previous ten points
10 percent 11.0 -
20 percent 6.0 5.0
30 percent 4.33 1.67
40 percent 3.50 0.83
50 percent 3.00 0.50

The first ten points of over-provisioning are worth more than the next fifty combined, which is why every drive has some and why the expensive ones stop at about 37 percent.

The Cai et al. NAND reliability survey states the lifetime model manufacturers use, and every lever in an endurance argument is one of its five terms:

Lifetime (years) = PEC × (1 + OP) ÷ (365 × DWPD × WA × R_compress)

On the reference drive - a 1,500-cycle die assumed, 9.95 percent over-provisioning, 0.33 DWPD, a measured write amplification of 2.2, no compression - it gives 6.2 years, just past the five-year warranty. Double the write amplification and it is 3.1 years; run the drive at 1 DWPD and it is 2.1.

Your own free space is over-provisioning, if TRIM reaches the drive

Over-provisioning is the ratio of physical bytes to live bytes, and you control the denominator. A 1 TB consumer drive has 1,099.51 GB of NAND. Keep some of it empty, with a filesystem that issues discards and a path that carries them, and the controller sees what it would see on an enterprise part:

Live host data Effective over-provisioning Non-selective WAF bound
1,000 GB, completely full 9.95 percent 11.1
900 GB 22.2 percent 5.5
800 GB 37.4 percent 3.7
500 GB 119.9 percent 1.8

Keeping a consumer 1 TB drive twenty percent empty puts it at the 37.4 percent over-provisioning of an enterprise mixed-use or write-intensive part and cuts the write amplification bound by two thirds. To make it permanent, partition to less than full capacity or, on NVMe, create a smaller namespace; a shrunk partition needs the freed range discarded.

None of it works unless TRIM reaches the drive, because without it the controller believes every block ever written is still live. On Windows, Microsoft says NTFS has it on by default unless an administrator turned it off, and in an administrator terminal fsutil behavior query DisableDeleteNotify reads 0 for NTFS where it is on. On Linux fstrim.timer runs it weekly. In lsblk -D, a DISC-GRAN and DISC-MAX of zero mean discards are not getting through; USB enclosures are the usual culprit, since passthrough needs UASP and a bridge chip that implements it, and several RAID layers swallow it too.

The indirection unit sets a floor the datasheet assumes you will respect

Large QLC drives map indirection units much bigger than a 4 KiB sector - 16 KiB on most Solidigm D5-P5336 capacities and 32 KiB on the 122.88 TB one. A host write smaller than the IU, or misaligned to it, forces a read-modify-write of the whole unit:

16 KiB unit, 4 KiB aligned write: read 16, merge 4, write 16; floor 16 ÷ 4 = 4
32 KiB unit, 4 KiB aligned write:                           floor 32 ÷ 4 = 8

Solidigm footnotes its endurance figures “IU-Aligned Endurance”, so smaller or misaligned writes are not covered by the number on the datasheet. OCP requirement NVMe-OPT-7 makes a drive report its IU in the NPWG field, which nvme id-ns shows, and the OCP log’s Unaligned I/O counter records how often the host got it wrong.

Garbage collection, wear leveling and refresh pull in different directions

Garbage collection wants hot and cold data in separate blocks. Wear leveling deliberately mixes them, moving cold data into worn blocks, and each move is a write no host asked for. Refresh reads, corrects and rewrites aging blocks before their errors outrun the correction engine; in simulation the Cai survey reports it raises SSD lifetime by an average of 9x. None of it runs on a shelf, which is the archive argument.

A controller that compresses is the one exception to a write amplification of at least one: SandForce claimed a typical 0.5, and the destruction runs measured 28 percent less written to the NAND. Nothing on the modern consumer market makes that claim.

SLC caching buys speed with a second write of every byte

Your TLC or QLC drive usually writes into a region of its NAND running in pseudo-SLC mode - fast, one bit per cell - acknowledges the write, and folds the data into full-density cells later. An SLC-mode block holds a third of a TLC block and a quarter of a QLC one, so most consumer drives size the cache dynamically from free space:

1 TB QLC drive Free space Raw cells available Maximum pSLC cache
Empty ~1,000 GB ~1,000 GB ~250 GB
Half full ~500 GB ~500 GB ~125 GB
90 percent full ~100 GB ~100 GB ~25 GB

A nearly-full drive has almost no cache, which is why an SSD that felt fast for a year feels slow once you fill it. When the cache runs out, writes go direct to the full-density cells: a drive writing 5,000 MB/s into cache and 150 MB/s direct, with 100 GB of cache, averages 222 MB/s over a 300 GB copy.

The fold also costs endurance. Counted in media bytes, a cached byte is written twice, so SLC caching contributes a factor of 2 to write amplification for every byte that is cached and folded. Counted in block erases it is four times the erases of a direct TLC write, most of them in one-bit mode, and no vendor publishes a pSLC-mode cycle rating for its own die, so the net effect on drive life cannot be calculated from the outside. Large sequential writes that bypass the cache are gentler on a drive than the same bytes dribbled in small enough to be cached.

DRAM-less drives page their own mapping table

A controller mapping at 4 KiB needs about 0.98 GB of map per TB, which is why the Cai survey gives the industry ratio as roughly 1 GB of DRAM for every 1 TB. A DRAM-less NVMe drive borrows host memory instead, through the Host Memory Buffer, and Linux caps that at 128 MiB per controller by default (max_host_mem_size_mb):

Capacity Full map wants Linux grants Fraction of the map resident
1 TB ~0.98 GB 128 MiB about one seventh
4 TB ~3.91 GB 128 MiB about one twenty-ninth

On scattered small random writes every map miss is a flash read, and map updates are themselves writes that land on the NAND: a drive that has to page its own mapping table writes more metadata to flash than one that does not. Which SSD to buy weighs that for a home PC.

Power-loss protection changes what a synchronous write costs

Power-loss protection is also an endurance feature. On a drive with real PLP there is no volatile write cache to flush, so nvme id-ctrl reporting vwc of 0 is a free host-side test for it. On a consumer drive every fsync is a real flush to NAND - small, immediate and amplified. On a PLP drive the same fsync is acknowledged from a capacitor-backed buffer, and the controller can coalesce what it eventually writes into a full indirection unit.

The capacitors age, and the OCP log reports it as Capacitor Health, a percentage where 100 is the passing hold-up energy threshold at manufacture (a healthy new drive typically reads above 100) and 1 is the minimum for a reliable shutdown; FFFFh means no PLP at all. A used enterprise drive reading well below 100 has aged capacitors. Intel and Solidigm data center SATA drives report the same check in attribute 175 (some at 201) and Samsung’s in 201, and NVMe Critical Warning bit 4 is set once the capacitors have failed their own test.

The wear unit is a physics event, not a counter

A flash cell is a transistor whose threshold voltage moves with charge stored behind a thin insulating oxide, and the only way through an insulator is a field high enough that electrons tunnel. That tunneling is the wear: a high-field current through an oxide generates traps in it, and every program and every erase does it again. Reads do not - Western Digital’s endurance paper states that they “do not meaningfully affect the isolation elements”. So the unit of wear is the program/erase (P/E) cycle: not hours, not power cycles, not bytes read.

Traps let the stored charge leak out faster, so as the cycle count rises, the time a cell holds its state falls. Endurance and retention are one specification quoted at two points on the same curve, which is why the same die can be rated for a client drive that must hold data a year unpowered and an enterprise drive that need only manage three months.

Shrinking planar cells made it worse - a late planar 1x-nm MLC cell holds roughly 100 electrons, according to the Cai survey - and measured endurance fell with every node:

Technology Measured endurance, P/E cycles per block
Older SLC about 150,000
5x-nm planar MLC (50-59 nm) about 10,000
1x-nm planar MLC (15-19 nm) about 3,000
1x-nm planar TLC about 1,000

3D NAND reversed the trend, mainly because the tunneling oxide in a charge-trap cell breaks down less readily, and the survey reports endurance up by more than an order of magnitude.

What a current QLC die is really rated at

Every extra bit per cell doubles the voltage bands that must fit in the same window, so margins narrow and the die takes fewer cycles. The bottom of the 100 to 1,000 P/E cycles still quoted for QLC belongs to early planar QLC, and it is wrong for the 3D QLC sold today. No vendor publishes the die rating in a finished drive, so it has to be derived. Solidigm’s D5-P5336 at 122.88 TB is rated 134.3 PBW, measured IU-aligned:

134,300 TB written ÷ 122.88 TB capacity = 1,093 full drive writes

That is at least about 1,093 cycles, and if the vendor set the rating with the JEDEC guidance formula and its factor-of-two guard band, twice that at a write amplification of 1: 2,186 cycles.

So data center 3D QLC sits between about 1,100 and about 2,200 P/E cycles, depending on how much guard band the vendor applied. The same sums on the 2 TB, 400 TBW consumer QLC drive give about 200 to 400 - a factor of five and a half lower, on nominally the same four-bits-per-cell technology, which says more about which bins go into which product than about QLC.

JESD218 is four requirements at once, and most ratings are extrapolated

JEDEC’s JESD218 defines what a TBW rating means - the current revision, JESD218C.01, was published in August 2026 and costs non-members $90 - and its free companion, JESD219A.01, defines the test workloads. JESD218 Table 1, reproduced in Seagate’s TP618, sets four requirements per class:

Client Enterprise
Workload JESD219 client workload JESD219 enterprise workload
Active use 40 °C at 8 hours per day 55 °C at 24 hours per day
Powered-off retention 1 year at 30 °C 3 months at 40 °C
Functional failure requirement no more than 3 percent no more than 3 percent
UBER no more than 1 per 10^15 bits no more than 1 per 10^16 bits

The functional failure requirement is the same for both; the class difference is retention, temperature, duty cycle and error rate. One error per 10^15 bits is one per 125 TB read, and one per 10^16 bits is one per 1.25 PB. Solidigm’s D7-P5810 brief states a test target of 1E-17, ten times the JESD218 enterprise figure; enterprise vendors design past the spec and say so, which the consumer market does not.

Most published TBW figures were never measured

JESD218 allows direct verification of the whole TBW and, when that cannot finish within 1000 hours, extrapolation. A drive taking a sustained 1 GB/s for 1000 hours writes 3.6 PB, enough for a consumer drive and nowhere near a 134 PBW QLC part, so the large ratings on the market are extrapolated from accelerated stress, not measured to destruction. The guidance formula for setting a rating carries a guard band for imperfect wear leveling:

TBW  <  (capacity × NAND cycling capability) ÷ (2 × WAF)

It also makes a consistency check. Assume a 1,500-cycle die under the reference drive’s 600 TBW and it solves to a write amplification of 1500 ÷ 1200, or 1.25.

A 600 TBW rating on a 1 TB TLC drive is exactly what a 1,500-cycle die, the JEDEC guard band and a write amplification of 1.25 under the client workload produce. The die rating is an assumption, not a disclosure, but the rating is not arbitrary.

The two test workloads are opposite in kind

The client workload is a replayed trace of real I/O recorded over an extended period on a PC; the enterprise workload is a synthetic mix of random writes heavily biased to 4 KB and 8 KB, according to Western Digital’s endurance white paper. A client TBW was measured against a real desktop trace with plenty of sequential traffic in it, so a database, a metadata vdev or a busy mail spool on a client drive burns the budget faster than the counter suggests.

What end of life actually does, and who implements it

Consumer drives mostly do not document this. The Intel 335 Series was designed to go read-only when its media wear indicator ran out and to brick itself on the next power cycle, and it did exactly that: stop writing, let the owner copy the data off, refuse to be trusted again after a reboot.

For data center NVMe drives, end of life is specified. OCP requirement EOL-5 switches the device to read-only when Available Spare reaches 0 percent, which it must report while spare blocks remain, so reads keep working after writes stop. The policy is Feature C2h, EOL/PLP Failure Mode, readable with nvme ocp eol-plp-failure-mode; every legal value ends in read-only at end of life, and requirement ROWTM-1 makes read-only the factory default. RETC-3 forbids throttling on the endurance metric, so a worn OCP drive is not allowed to slow down as a warning.

What the endurance-to-destruction runs proved, and what they did not

The best-known experiment, by The Tech Report, ran six consumer drives until they died, concluding on March 12, 2015. The original articles are no longer on the publisher’s site; the figures survive in the Internet Archive and in third-party summaries.

Drive Final host writes
Kingston HyperX 3K 240 GB, incompressible data 728 TB
Intel 335 Series 240 GB about 750 TB
Samsung 840 Series 250 GB, TLC just under 1 PB
Corsair Neutron GTX 240 GB about 1.2 PB
Kingston HyperX 3K 240 GB, compressible data past 2.1 PB
Samsung 840 Pro 256 GB over 2.4 PB

Consumer drives massively exceed their ratings, on a sample of six: the 840 Pro’s 2.4 PB on a 256 GB drive is 9,375 full drive writes. Compression explains only part of it: of the two identical HyperX drives, the one fed compressible data “wrote 28% less to the NAND”, which carried it to about 1.1 PB, and the further petabyte the author put down to luck in the flash. End-of-life behavior depended entirely on the vendor: the Intel went read-only and bricked itself as designed, the Kingston refused writes after 728 TB and then did not respond after a reboot, and the others simply died.

The SMART warnings were not reliable, and this is the finding that matters for buying used. The Corsair, Intel and Kingston drives all warned before they died. Samsung’s own software “pronounced the 840 Series and 840 Pro to be in good health before their respective deaths”, and the 840 Pro reached 2.4 PB with over 7,000 reallocated sectors while the vendor tool called it healthy.

What it did not prove: six drives is not a population, they were planar MLC and TLC from 2012 and 2013, a generation that no longer exists, and the workload was largely sequential, the kindest case for write amplification. Conservative ratings are a statement about warranty policy, not a license to plan past the rating on a drive you depend on.

Unpowered retention, and why an SSD is a bad archive

Nothing refreshes a flash cell’s charge while the drive is unpowered, and it leaks. The JESD218 requirements - 1 year at 30 °C for client, 3 months at 40 °C for enterprise - apply at the end of rated life, on a drive that has consumed its full TBW; a drive at 3 percent used is nowhere near that boundary. The widely circulated table showing retention collapsing to a week or so is the enterprise row, extrapolated to high temperatures, on a drive at 100 percent of its endurance.

The OCP data center specification sets different numbers again: at least 1 month at 40 °C powered off at end of life (RETC-1), at least 7 years powered on (RETC-2), and fully functional after at least 5 years new in the box at 25 °C (SLIFE-1).

Seven years powered against one month unpowered is a factor of eighty-four. For archives:

  • More bits per cell means less retention headroom. QLC is the worst archive flash, and it is what cheap large drives are made of.
  • You cannot tell from the outside. A drive stored three years may read back perfectly or come back with uncorrectable sectors, and nothing on its label, its SMART data at the time of storage or its price predicts which.

Powering a stored drive up occasionally is a reasonable precaution rather than a guarantee, because nobody documents whether a consumer controller scrubs during idle. A hard drive holds its magnetic domains for decades on a shelf, with mechanical risks of its own, and tape is the medium designed for the job. Keep multiple copies on more than one medium, checksummed and verified on a schedule; an SSD is an excellent working drive and a poor box in the attic - see HDD or SSD.

Data center QLC and the cheap QLC that made its reputation are different products

The reference part is Solidigm’s D5-P5336, built on 192-layer QLC, with a 2 million hour MTBF, 3 months of powered-off retention at 40 °C and fewer than 1 uncorrectable sector per 10^17 bits read: an error rate specification ten times better than the JESD218 enterprise requirement. Solidigm’s own worst case - the 122.88 TB drive taking 32 KB 100 percent random writes around the clock for five years - retains about 5 percent of its life. At 4 KB random writes it retains about 10 percent.

Three things to carry to a listing page. Data center QLC and consumer QLC are different products - about 1,100 to 2,200 cycles against 200 to 400. The indirection unit is the whole risk: fed 4 KiB random writes, a large QLC drive runs at four to eight times the amplification its rating assumed. And a model number does not pin the flash: Micron’s consumer SSD pages say in the small print that it “reserves the right to transition between NAND series during future production cycles”, so two used drives with the same label can have different endurance - one more reason to read the drive’s own counters rather than look up the model.

What has to be measured, and how this site records it

Not published: the P/E rating of the NAND in a finished drive, the garbage collection policy, the pSLC-mode cycle rating and, usually, the SLC cache policy, so steady-state write amplification and sustained-write speed have to be measured rather than looked up. Vendor-specific: Percentage Used, unless the drive is OCP compliant. Unknowable: retention for a specific used drive, at its wear level and at the temperature of the closet it sits in. Extrapolated: most published TBW figures. This site’s catalog shows a figure where a source states it and marks the rest not stated.

Related guides

Affiliate links:we earn from qualifying eBay purchases. eBay's live price is final. Full disclosureAffiliate disclosure:We are a member of the eBay Partner Network and earn a commission from qualifying purchases made through links to eBay on this site. Prices and availability are captured periodically and may have changed - the live price is always the one shown on eBay.