HardDiskIndexFAQ

HDD or SSD: where each still wins, and why the gap persists

By Harry Saarinen ·

In this guide: 14 sections
  1. The short answer, by workload
  2. A laptop, a desktop, a console, a backup drive: what each one wants
  3. What to check on a listing page
  4. HDD vs SSD price per terabyte in 2026
  5. Speed: random access is where flash wins by a thousand
  6. Sequential speed is much closer than people assume
  7. Bigger hard drives are slower per terabyte
  8. The flash tail: when an SSD read is as slow as a hard drive
  9. Both slow down when full, and only flash falls off a cliff
  10. HDD vs SSD lifespan and reliability
  11. HDD vs SSD for backup and archiving
  12. HDD vs SSD in a NAS: power, noise and vibration
  13. Per dollar, the hard drive wins on capacity and nothing else
  14. Putting an SSD in front of hard drives

Buy an SSD for anything you wait on - the operating system, programs, games and the files you are working on - and for any laptop, and hard drives for bulk storage nobody waits on: backups, a media library, archives and a NAS. That split has held for a decade, and the reasons are structural rather than fashionable: a 7,200 rpm hard drive needs about 12.5 milliseconds to reach a random sector where an NVMe SSD answers in 20 to 70 microseconds, while a terabyte of new 3.5-inch hard drive cost about a quarter as much as a terabyte of new M.2 SSD on this site’s figures for November 1, 2026. Both got much more expensive in 2025 and 2026, for opposite reasons, and the price gap between them has not closed.

The short answer, by workload

  • Laptop: an SSD. An M.2 NVMe drive if it has the slot, a 2.5-inch SATA SSD if it has only a drive bay. A hard drive is nearly never the right pick now.
  • Desktop or gaming PC: an NVMe SSD for the system, programs and games, or a SATA SSD if the board has no M.2 slot. Add a 3.5-inch SATA hard drive for what the SSD has no room for. For games, buy capacity before a faster tier.
  • Backup drive: hard drives, plural, in more than one place, and an external one is the easy start. Charge leaks out of an unpowered SSD; magnetism does not leak off a platter.
  • Console: current games stay on the SSD, and a USB hard drive holds what you are not playing.
  • NAS and bulk storage: CMR hard drives, because capacity is the one thing they win on and they win it by a wide margin.
  • Databases, virtual machines and many users at once: an SSD, because one actuator cannot serve many random streams.
Workload Choose Why
Any laptop SSD, M.2 NVMe or 2.5-inch SATA as the machine takes No head to crash when it is carried, a fraction of the idle power, no seek
Boot and applications NVMe SSD, SATA SSD if no slot Queue-depth-one random reads, where the gap is already 100× before any queue
Gaming library SSD, and prioritize capacity over tier Level loads are random reads; SATA to NVMe is a small further gain
Console games The internal SSD or the maker’s expansion; a USB hard drive for storage Current games will not run from USB
Video editing scratch NVMe, chosen on sustained write The SLC cache cliff is exactly this workload
Video library, single stream Hard drive Sequential playback needs tens of MB/s, not thousands
Backup target Hard drive, often external Large, sequential, cheap, powered occasionally
Cold archive Hard drives or tape, plural Charge leaks; magnetism does not
NAS bulk storage Hard drives, CMR Cost per terabyte, and rebuild behavior
Databases, VMs, many users SSD Uncorrelated random streams are the worst case for one actuator

Shingled recording is the trap on the hard drive side. A shingled (SMR) drive overlaps its tracks to fit more on the same platters, and in exchange it writes well only sequentially; retail SMR drives are drive-managed, so the firmware hides the shingling. That costs nothing on a sequential backup and a great deal in a NAS, where its random-write and rebuild behavior gives you the worst of both technologies at hard drive prices. CMR vs SMR covers how much capacity it buys, how to find out what a specific drive is and what to do when you cannot.

Figures from vendor manuals, JEDEC standards and Backblaze’s fleet data name their source and show the arithmetic.

A laptop, a desktop, a console, a backup drive: what each one wants

A laptop: a hard drive is nearly never the right pick now

Start by finding out what the machine has. In PowerShell, run as administrator:

Get-PhysicalDisk | Format-Table FriendlyName, MediaType, BusType

MediaType reads HDD, SSD or Unspecified and BusType reads SATA, NVMe or USB among others (Microsoft’s class reference). What the laptop can take instead is in its maker’s manual: a 2.5-inch bay, an M.2 slot, both, or a drive soldered to the board. The manual says which one fits; this site shows what each kind costs.

If it says HDD, replace it with an SSD. Three reasons, in the order you will feel them.

Seek. A 5,400 rpm laptop disk waits 5.56 ms on average for the sector to come round (Toshiba’s MQ04 manual, February 2018), plus a seek that Toshiba does not publish, which comes to about 17 ms an access and about 60 random reads a second. Samsung rates its 870 EVO SATA SSD at up to 13,000 at queue depth one (data sheet, January 2021). Starting Windows, opening a browser and searching your files are full of small scattered reads, so this is the upgrade an old laptop’s owner notices in every minute of use.

Shock. A running hard drive has a head flying over a spinning platter. Seagate rates its 7 mm BarraCuda laptop disks at 400 G while operating and 1,000 G at rest (datasheet, May 2020), and Toshiba’s manual gives the same two figures. An SSD has no head, and Samsung quotes 1,500 G for the 870 EVO. The pulse lengths differ (2 ms, 1 ms, 0.5 ms), so that is no clean ratio. The point is the gap inside the disk’s own rating: it is at its weakest while the laptop is open and in use, which is when a laptop gets knocked.

Power. This one is smaller than it is usually made out to be. Under load a laptop disk draws less than an SSD: 1.5 W on WD’s 1 TB Blue (product brief, May 2023), against 2.0 to 2.5 W on a 1 TB 870 EVO and 4.2 to 4.3 W on a 1 TB 990 EVO Plus NVMe drive (data sheet, October 2024). Idle is where the SSD wins: 0.45 W on Seagate’s 1 TB disk and 0.5 W on WD’s against 30 mW on the 870 EVO and 60 mW on the 990 EVO Plus, an eighth to a fifteenth. A disk can spin down to 0.1 W, and then you wait for it to spin back up.

The rest is capacity and recording. 2 TB is the most Seagate fits in the 7 mm height it sells for thin laptops; its 3 to 5 TB models are 15 mm, aimed at external boxes and all-in-ones. Every model on that BarraCuda sheet and on WD’s Blue brief is shingled, so a long write slows right down. The one exception worth naming is a laptop that never leaves a desk, with a second bay that holds only a media library, and a USB drive does that job with nothing to open.

The sum for a five-year-old laptop. On this site’s eBay figures for November 1, 2026, a new 1 TB laptop disk had a median of $94 and a new 1 TB SATA SSD $241. You do not have to replace like for like: the SSD has to hold what you use plus a margin, not what the disk held. With 300 GB used and 15 to 20 percent kept free that is about 375 GB, so a 500 or 512 GB SSD, with medians near $199 and $135 new and near $65 used, against about $55 for a used 500 GB disk. Keep the rest on an external drive. Which SSD to buy for a laptop or desktop covers M.2 or SATA, generation, length, DRAM and moving your system across; the listings are at SATA SSDs and NVMe SSDs.

A desktop or gaming PC: one SSD, and a hard drive if you need the room

A desktop has no shock or battery problem, so the hard drive is back on merit: it is the cheap room behind the SSD. The SSD carries the system, the programs and what you are playing, and some games now require it: Starfield’s Steam page lists “SSD Required” under both its minimum and recommended specifications. A 3.5-inch SATA hard drive carries the library, at $45 to $61 a terabyte new against $220 to $249 for an M.2 SSD. How much storage you need sizes each one.

Two things to read on the hard drive’s listing. The cheapest rows are mostly SAS: 871 of the 1,000 cheapest hard drive rows on this site on November 1, 2026. SAS is a server link that does not connect to a desktop’s SATA ports. SATA hard drives lists the SATA rows alone, and drive interfaces explains the rest. And a desktop disk is not always 7,200 rpm: Seagate’s 2 TB BarraCuda is, at up to 220 MB/s, while the 8 TB model’s manual quotes 6.0 ms of average latency, where 7,200 rpm would be 4.17 by the sum below, and up to 190 MB/s. Both manuals (March 2023) state SMR.

A console: the maker’s rules decide

Sony says PS5 games “can’t be played directly from USB extended storage”, though PS4 games can, and asks for SuperSpeed USB 5 Gbps or later at 250 GB to 8 TB (Sony support). Microsoft’s launch guide marks USB drives “No” for games optimized for Xbox Series X|S and lets them play Xbox One titles directly (Xbox Wire, September 24, 2020). So the SSD inside, a PS5 M.2 drive or an Xbox Storage Expansion Card holds what you are playing, and a USB hard drive holds the rest, which saves downloading it again. Usable sizes and Sony’s M.2 limits are in how much storage you need.

A backup drive: external, and more than one

An external hard drive is the cheapest copy you can unplug, and it comes in two shapes. A portable drive runs from its USB cable with no power supply (Seagate Portable Drive, 1 to 5 TB, September 2019). A desktop external needs its own adapter (Seagate Expansion Desktop, 4 to 18 TB on its 2021 sheet). Carry the first and leave the second on the desk: Seagate’s 3.5-inch BarraCuda manual gives 70 to 80 G operating and 300 G at rest at 8 TB, against the 400 and 1,000 of its laptop disks.

A portable SSD suits a copy that travels every day. Samsung rates its T7 at up to 1,050 MB/s and 1,500 G (data sheet, 2020), but new 1 TB USB SSDs had a median of $254 a terabyte against $75 for a 4 TB USB hard drive, and an SSD is a poor copy for a shelf. How much storage you need works out how many terabytes of copies your data wants.

What to check on a listing page

  1. Decide the slot before the drive. The machine’s manual says whether it has a 2.5-inch bay, an M.2 slot or both. An M.2 slot is not automatically NVMe - see drive interfaces before you buy the wrong key - so filter flash by interface, not by form factor: NVMe SSDs or SATA SSDs.
  2. Read the SAS note on the hard drive side. A note sits above any table with SAS rows, because a SAS drive does not connect to an ordinary PC. SATA hard drives lists SATA drives only.
  3. Read today’s price ratio off the live medians. Sort the SATA hard drive listings and the NVMe listings by price per terabyte and divide, or follow both on Trends. Dollar figures in this guide are dated; the listing pages and Trends carry today’s.
  4. Compare capacity points, not price bands. Each capacity has its own page: set 1 TB drives beside 8 TB drives and watch the price per terabyte fall away. Below roughly a terabyte a hard drive is all motor and casting, which is why the gap persists. The min and max filters bound price, not size.
  5. Decide recording mode before capacity on the hard drive side. Treat “not stated” as SMR if you are building an array. If shingling is fine for what you are doing, ignore it and buy the cheaper terabytes.
  6. Check the spindle where latency matters. The rpm filter separates 7,200 rpm drives from 5,400 rpm ones, an exact 1.39 ms of average rotational latency on every single access, forever.
  7. For flash, read the sustained figure and not the peak. Every consumer drive quotes its SLC cache speed; ask what it does after the cache.
  8. Know what the price per terabyte covers. It is the item price, with shipping shown beside it and never added, so add it to your own sum. The tables list working drives, at fixed prices by default; ?bids=show adds the bid-only auctions, marked.
  9. Check the power-on hours on arrival, on either medium. A used drive of either kind carries an odometer, and buying used drives on eBay is about reading it honestly.

HDD vs SSD price per terabyte in 2026

On this site’s figures a hard drive is the cheaper way to store data at every size in this table, and price per terabyte is the only measure on which it is cheaper. They are the median Buy It Now price per terabyte of single-drive listings on its US catalog, fetched on November 1, 2026, item price alone, with shipping kept separate.

These are eBay asking prices. For a new drive, check a retailer with stock as well: in September an 8 TB NAS or Exos disk was about $41 to $50 a terabyte on Amazon against $61 here, and a 2 TB NVMe drive about $175 against $220 (the dated shop prices). The ratio between the two comes out about the same. Trends follows the market from here.

What you would buy Size New, per TB Used, per TB
2.5-inch SATA hard drive (laptop size) 1 TB $94 $66
2.5-inch SATA hard drive (laptop size) 2 TB $72 $60
2.5-inch SATA SSD 1 TB $241 $135
2.5-inch SATA SSD 2 TB $196 $124
M.2 NVMe SSD 1 TB $249 $140
M.2 NVMe SSD 2 TB $220 $125
3.5-inch SATA hard drive 4 TB $53 $40
3.5-inch SATA hard drive 8 TB $61 $38
3.5-inch SATA hard drive 16 TB $45 $33
External USB hard drive 2 TB $105 $35
External USB hard drive 4 TB $75 $29
External USB hard drive 8 TB $52 $28
External USB SSD 1 TB $254 $219

In a laptop the gap is about 2.6 to 1: $241 against $94 at 1 TB, and 2.7 to 1 at 2 TB. In a desktop it is about 4 to 1, because the hard drive gets to be big: a 2 TB M.2 SSD at $220 against an 8 TB disk at $61 is 3.6 to 1, and against a 16 TB disk at $45 it is 4.9 to 1. An external drive costs more than the bare disk at small sizes, and by 8 TB the median external ($52) sits below the median internal ($61). Used takes about 44 percent off flash at 1 TB, 17 to 38 percent off bare hard drives and about half or more off external ones, and what a used drive hides is the subject of buying used drives on eBay. Data center prices run on other numbers, and they come later.

Why both got more expensive at once

The comfortable assumption that flash swings wildly while disk drifts steadily downward inverted in 2025 and 2026. TrendForce’s quarterly NAND contract-price figures and forecasts compound to roughly four times across three quarters, Q4 2025 to Q2 2026, and its forecast of September 30, 2026 adds a further 15 to 20 percent for Q4 2026. AI infrastructure demand outran fab capacity, and fab capacity arrives in lumps years after somebody decides to build it.

Hard drives rose for the opposite reason: three suppliers build almost exactly to demand, so when demand stepped up there was no inventory and no idle line. US hard drive street prices rose about 15 to 90 percent between March and August 2026 depending on the line, most for consumer drives. Nearline is now 87 percent of Seagate’s hard drive sales, and consumer retail is about 5 percent of Western Digital’s revenue. The retail shopper is buying the leftovers of a data center supply chain. The 2026 storage price guide traces each of those figures to its release; buy storage now or wait is where the forecasts live.

Why the gap persists

The price per terabyte has stayed far apart for fifteen years because the two technologies have different cost structures. Most of a hard drive’s cost is fixed: base casting, spindle motor, voice-coil actuator, preamp, controller board, helium seal and test time, plus a smaller cost per platter and per head. Capacity is platters times areal density, and raising areal density costs nothing in the bill of materials at all. Seagate’s own product manuals show it. The Exos X24 at 24 TB and the Exos M at 32 TB (ST32000NM004K) both carry 20 heads on 10 disks, on the same 7200 rpm spindle in the same helium-sealed enclosure:

Exos X24, 24 TB      20 heads / 10 disks   =  2.4 TB per platter
Exos M,   32 TB      20 heads / 10 disks   =  3.2 TB per platter

identical mechanism, +33.3% capacity, zero additional parts

An SSD’s cost converges on one number as capacity rises: the cost of a NAND die divided by the bits on it. In November 2025 Phison’s chief executive stated that the price of a 1 terabit TLC NAND die had gone from $4.80 to $10.70 in six months. A binary terabit is 0.1374 decimal terabytes, so that is $34.9 rising to $77.9 per terabyte of bare NAND - before the controller, the DRAM, the board, the test time, the firmware and anybody’s margin. A hard drive maker who improves areal density by 33 percent sells 33 percent more terabytes out of the same castings, motors and test slots. A NAND maker who improves bit density by 33 percent has to build the capacity to deposit and etch a third more layers first. One of these is free and the other is a fab.

That is also why the low end of the hard drive market disappeared rather than getting cheap. Nobody manufactures a new 250 GB 3.5-inch drive, because the motor, the casting and the test time cost the same as they do on a 30 TB one. Below roughly a terabyte the fixed cost dominates and the disk has little left to offer. On this site’s eBay figures a new 500 GB SATA SSD still asks two to three times a new 500 GB laptop disk, but used the two are within about $10 of each other, and in German shops in September the cheapest 1 TB SATA SSD, at EUR 109, undercut the cheapest 1 TB hard drive, at EUR 120 (the 2026 storage price guide). At a terabyte or less, price both.

Forecasts of a crossover have not survived. A widely cited 2021 analysis projected QLC flash and hard drives reaching price parity at $1.00 per terabyte “during 2026”. The date has arrived, neither medium is within an order of magnitude of a dollar, and flash still costs several times what disk does. The two media stopped competing for the same bytes. Flash took the bytes somebody waits on, disk kept the bytes nobody waits on, and both grew.

Speed: random access is where flash wins by a thousand

This is the difference you feel. Starting a program or loading a level is mostly small reads from scattered places, and for each one a hard drive sits through two mechanical waits. Rotational latency is the time for the sector you want to arrive under the head: half a revolution on average, so average rotational latency is 30000 / rpm milliseconds, and nothing a firmware engineer does changes it. Seek time - the actuator moving the head to the right cylinder - is the other wait, and makers have stopped publishing it: the 2.5-inch Toshiba manual gives 5.56 ms of latency and no seek time, and the 32 TB Exos M manual gives 4.16 ms and no seek time either. Add the two waits and divide into one second for a queue-depth-one random IOPS figure, remembering that the seek column is a range across vendors and generations rather than a specification:

Drive Rotational Seek (typical, not specified) Access Random IOPS at QD1
5,400 rpm laptop or desktop disk 5.56 ms unpublished ~17 ms ~60
7,200 rpm desktop disk 4.17 ms 8-9 ms ~12.5 ms ~80
SATA SSD none none 80-200 µs up to 13,000 (870 EVO)
NVMe SSD none none 20-70 µs ~14,000 at 70 µs

Server disks spin faster. 10,000 and 15,000 rpm SAS drives reach about 7 ms and 4.3 ms, or 140 and 230 IOPS, and need a controller an ordinary PC does not have (drive interfaces). The spindle stopped at 15,000 rpm because windage power rises with the cube of velocity: going from 7,200 to 15,000 rpm at the same radius costs about nine times the power, all of it heat inside a sealed enclosure. That is why 15K drives used smaller platters, and why nobody built a 20K. Flash has no mechanical wait, and one outstanding read through a real kernel and filesystem sees roughly the device figure in that table, less some software overhead.

Queue depth widens the gap

The gap widens dramatically once work is queued. A SATA link carries at most 32 outstanding native command queueing commands, and NVMe allows 65,536 queues of 65,536 commands each. That is not an incremental improvement, it is a different order of object. At SATA’s queue ceiling the numbers look like this:

SATA SSD at the interface ceiling
  perfect 32-way concurrency at 100 µs device latency
    32 / 100 µs                  = 320,000 IOPS   theoretical
  observed ceiling               ≈  90,000 IOPS

7,200 rpm drive
  80 IOPS × 4 KiB                =    320 KiB/s of random 4K reads
SATA SSD at QD32
  90,000 × 4 KiB                 =    352 MiB/s

Three orders of magnitude, from the same interface, on the same cable. That is why 15K SAS died as a product category: it delivered roughly three times a desktop drive’s IOPS for many times the price per terabyte, and flash arrived offering a thousand times. At the top of the range, Samsung rates the 2 TB 990 EVO Plus at up to 1,000,000 random read IOPS and high-end PCIe 5.0 parts go above two million. Those are ratings at high queue depth across many threads; one process doing one read at a time still sees 20 to 70 microseconds.

Sequential speed is much closer than people assume

Random access is where flash wins by a thousand. Sequential is where it wins by two to four over SATA and thirty to fifty over a current M.2 drive, which surprises people in both directions. The rated figures:

Drive Rated sequential transfer Source
2.5-inch 5,400 rpm laptop disk up to 140 MB/s Seagate BarraCuda 2.5-inch datasheet, May 2020
3.5-inch 8 TB desktop disk up to 190 MB/s Seagate BarraCuda manual, March 2023
3.5-inch 2 TB 7,200 rpm desktop disk up to 220 MB/s Seagate BarraCuda manual, March 2023
3.5-inch 32 TB nearline disk 285 MB/s Seagate Exos M product manual
2.5-inch SATA SSD up to 560 MB/s Samsung 870 EVO data sheet, January 2021
Portable USB SSD up to 1,050 MB/s Samsung T7 data sheet, 2020
M.2 NVMe SSD, PCIe 4.0 up to 7,250 MB/s Samsung 990 EVO Plus data sheet, October 2024

A film needs far less than the slowest of those, and a USB hard drive runs at its disk’s speed rather than its port’s. The ceiling on a SATA SSD is set by the interface, not the flash: 600 MB/s, about 550 in practice. PCIe x4 tops out at 3.938 GB/s on 3.0, 7.877 on 4.0 and 15.754 on 5.0. Drive interfaces works through which of these a given slot actually gives you, which is rarely the one printed on the box, along with the SAS links.

A hard drive’s sequential figure is also not one number. Platters spin at constant angular velocity while linear bit density is held roughly constant, so bytes per revolution scale with the circumference of the track, which scales with radius. The data band on a 3.5-inch platter runs from an outer radius to an inner one roughly half of it:

throughput ∝ 2πr × linear density × revolutions per second
           ∝ r

outer tracks : inner tracks  ≈  2 : 1

A desktop drive rated up to 190 MB/s will finish a benchmark sweep near 95. That is geometry, not a fault, and it is why manufacturers quote “up to”. Confirm it on a drive you own with a benchmark that sweeps the whole surface, such as HD Tune Pro on Windows, a paid utility with a trial. On Linux, take a buffered read at the start and near the end:

sudo hdparm -t --direct /dev/sda         # buffered read from the outer tracks
sudo hdparm -t --direct --offset 14000 /dev/sda   # 14,000 GiB in, near the end of a 16 TB drive

--offset counts in GiB, so pick a number that lands inside your drive, and --direct skips the page cache; it is -T that times your RAM.

Bigger hard drives are slower per terabyte

Areal density is the number the industry publishes. Access density - performance per terabyte stored - decides whether a drive suits a job, and it has been going the wrong way for fifteen years. A hard drive’s random IOPS is set by one actuator and one spindle speed, neither of which has changed since the 2000s, while capacity per drive has gone up by more than an order of magnitude:

2010   2 TB   at ~80 random IOPS  =  40.0 IOPS per TB
2026  32 TB   at ~80 random IOPS  =   2.5 IOPS per TB

capacity ×16,  IOPS unchanged,  IOPS per terabyte ÷16

Sequential bandwidth per terabyte falls too, because the newest capacity gains come from packing tracks closer together rather than from shorter bits. Going from the perpendicular Exos X24 to the HAMR Exos M, Seagate’s own figures show track density up 44.1 percent and linear density down 3.2 percent, so 24 TB at 285 MB/s is 11.9 MB/s per terabyte and 32 TB at 285 MB/s is 8.9. At the same 24 TB capacity, the HAMR part is rated slower, 240 MB/s, than the perpendicular part it replaces.

The consequence a buyer feels is the whole-drive pass: the first full backup, a restore, a move to a new disk. On a desktop disk that is a long night, and on the biggest drives it is most of two days:

best case, outer tracks only, which no real workload achieves
   8 TB ÷ 190 MB/s  =  42,105 s  = 11.7 hours
  32 TB ÷ 285 MB/s  = 112,281 s  = 31.2 hours

realistic, inner tracks run at roughly half the outer rate
  average ≈ (285 + 143) / 2 = 214 MB/s
  32 TB ÷ 214 MB/s  = 149,533 s  = 41.5 hours

A full-surface read of the largest single drive is now a forty-hour operation, before any contention from live traffic. That is the real constraint on RAID rebuild, and drives for a NAS works through what it does to array design.

Seagate’s MACH.2 dual-actuator drives reach up to 480 MB/s sustained against SATA’s practical 550 and roughly double per-drive IOPS, to about 11.4 per terabyte at 14 TB. But nearly every drive you are likely to buy, and nearly every drive on this site’s listings, has one actuator.

The flash tail: when an SSD read is as slow as a hard drive

Flash’s median latency is excellent and its tail is not. A read from NAND is not one operation with one cost: it is a ladder of error-correction attempts, and the controller climbs it until the data comes back clean. The rungs and their costs, from Cai, Ghose, Haratsch, Luo and Mutlu’s survey of flash errors and recovery:

Rung Cost
BCH hard decoding with read-retry 70 µs per iteration
LDPC hard decoding 80 µs
Neighbor-cell-assisted correction 140 µs for two neighbor reads, plus 70 µs per additional adjacent value
LDPC soft decoding 80 µs per level, so about 400 µs at five levels
Superpage-level parity recovery about 10 ms

Look at the bottom row. Ten milliseconds is a hard drive. A 7,200 rpm access is about 12.5 milliseconds, so a worn or hot SSD that falls all the way through to superpage parity recovery serves that read in roughly the time the platter it replaced would have taken: 125 times its fast path of 80 µs.

This is why enterprise SSD datasheets quote percentile latency rather than average: Solidigm’s D5-P5316 publishes a 99.999th-percentile read of 600 microseconds. Flash wins decisively in the median and the gap closes sharply in the tail, and a consumer drive that publishes no percentile figure at all is telling you it would rather you did not look.

The ladder gets climbed more often as a drive wears, heats up or sits unpowered between reads, because all three widen the voltage distributions the controller is trying to separate. A drive that used to be fast and is now intermittently slow is usually not broken. It is spending more time on rung three.

Why a small SSD is slower than a big one

A single NAND die is not fast: representative figures for the TLC generation, from generation summaries and conference papers rather than any current datasheet, are 77 to 100 µs to read a page, 1,300 to 1,630 µs to program one and about 15 ms to erase a block. The IOPS come from parallelism, not from speed. Take a drive with eight channels and four dies per channel, a plausible mid-range arrangement:

32 dies operating concurrently at tR = 90 µs
  32 / 90 µs  = 355,000 page reads per second

A small drive has fewer dies, so a 500 GB model of the same product is genuinely slower than the 2 TB model and not just artificially limited. And a workload that hits the same die repeatedly gets none of the parallelism, which is why a queue depth of one looks so much worse than the headline figure.

Both slow down when full, and only flash falls off a cliff

The hard drive’s decline is the radius argument. Files written late land on inner tracks, so a drive at 95 percent full is writing at roughly half its headline rate. Predictable, gradual, and visible in a sweep.

The SSD’s decline is about free blocks. NAND is written in pages and erased in much larger blocks - 256 to 1,024 pages per block - so the controller can never overwrite in place. It writes elsewhere, updates a mapping table and garbage-collects the stale pages later, and that collection needs spare blocks to work with. Every drive has a hidden reserve from the gap between binary and decimal units alone, 9.05 percent of the NAND fitted in a 1 TB drive, the same gap drive capacity explained covers for the number your operating system shows. Enterprise drives add far more on top, and SSD endurance reads the label that gives it away.

The cliff is the SLC cache. A consumer TLC drive writes incoming data one bit per cell into blocks that will eventually hold three, because single-level programming needs only one coarse threshold, while TLC and QLC have to be written in stages. Caching one gigabyte therefore occupies three gigabytes of eventual TLC capacity, and on a QLC drive one occupies four. Most consumer drives size this cache dynamically from whatever is free, so:

Empty 2 TB TLC drive:  large dynamic cache, writes at interface speed
Drive at 90% full:     ~200 GB free ÷ 3  ≈ 67 GB of cache left
Drive at 98% full:     essentially none - writes go straight to TLC

Past the cache, Samsung rates its 870 EVO, a TLC drive, at 300 to 530 MB/s and its 870 QVO, a QLC drive, at 80 to 160 MB/s, by size. A nearly full QLC drive, past its cache, can write more slowly than a hard drive. That is the product working as designed, not a defect, and it is the most common reason someone concludes their SSD has “gone bad”.

The practical rule is the same for both media: keep 15 to 20 percent free. On flash it preserves the cache and keeps write amplification down; on platters it keeps you off the inner tracks. On a laptop’s only drive that margin comes straight out of the size you buy.

Write amplification is the other cost of small writes: in JEDEC’s worked example the host writes 8 sectors into a 256-sector block and the drive rewrites all 256. Small random writes consume flash. Total volume does not. SSD endurance works through what to do about it and how to read the odometer on a used drive.

HDD vs SSD lifespan and reliability

The important difference is not which fails more often. It is what failure looks like from where you are standing.

Hard drive SSD
Wear mechanism Bearings, actuator, head-media spacing, lubricant NAND oxide degradation per program/erase cycle
Typical warning Reallocated and pending sectors, slow reads, noise Often none
Typical end state Degrades; parts of it still read Read-only, or gone entirely
Media defects Grow slowly, remapped from a spare pool Blocks retire into spare area, tracked as spare remaining
Controller failure Donor PCB swap is sometimes viable The mapping table is the data; recovery needs a lab
Sudden total loss Head crash, seized spindle Controller or firmware failure, power loss during write
Environmental sensitivity Shock while operating, which is a laptop’s daily life; rotational vibration Temperature, both for retention and for throttling
Slow-failure signature Rising pending sector count Rising time on the ECC ladder

A hard drive usually tells you it is dying, and an SSD usually does not. The platters carry data in a layout with a comprehensible relationship to logical block addresses, so a drive that is failing mechanically can often be imaged slowly with ddrescue and most of it recovered. When a growing defect list is the failure mode you get weeks of warning, and you can read it. On Windows, the open-source CrystalDiskInfo shows the counters and turns to Caution when the reallocated, pending or uncorrectable counts cross its threshold; its page says it reads NVMe on Windows 10 or later and only “a part of” USB drives. On Linux the same counters come from smartctl:

$ sudo smartctl -a /dev/sda
ID# ATTRIBUTE_NAME          FLAGS    VALUE WORST THRESH RAW_VALUE
  5 Reallocated_Sector_Ct   PO--CK   100   100   010    0
  9 Power_On_Hours          -O--CK   079   079   000    18942
194 Temperature_Celsius     -O---K   114   102   000    36
197 Current_Pending_Sector  -O--C-   100   100   000    0
198 Offline_Uncorrectable   ----CK   100   100   000    0

Attributes 197 and 198 moving off zero is the signal that matters: sectors the drive could not read and has not yet been able to reallocate. Attribute 5 climbing steadily means the spare pool is being consumed. Raw value encodings are vendor-specific, so watch the direction of travel on one drive rather than comparing two; buying used drives on eBay covers the full set of checks for a drive that has had a previous owner.

An SSD’s data has no such relationship to anything. The flash translation layer maps logical blocks to physical pages, that map is rebuilt and rewritten constantly, and it lives on the drive. Lose the controller and the NAND packages need chip-off desoldering and reverse engineering of a proprietary FTL to read. This is why SSD failure feels sudden and total: there is no partial-read path. So a backup matters more on the laptop you just upgraded, not less. The one warning flash does give is standardized on NVMe, and worth reading every few months. CrystalDiskInfo lists these same fields on Windows:

$ sudo smartctl -a /dev/nvme0
Percentage Used:                    7%
Available Spare:                    100%
Available Spare Threshold:          10%
Data Units Written:                 20,000,000 [10.2 TB]
Unsafe Shutdowns:                   14
Media and Data Integrity Errors:    0

Percentage Used is the controller’s own estimate of consumed endurance and can exceed 100. Available Spare falling toward its threshold is the closest thing to a pending sector count that flash offers; on SATA SSDs the equivalent attribute has no standard ID. SSD endurance reads the rest.

Two more asymmetries. Power loss during a write is a real risk on consumer SSDs and largely not one on enterprise drives, which carry capacitors to flush the mapping table; a hard drive losing power mid-write corrupts the sectors in flight and nothing else. And a hot hard drive wears its bearings and lubricant faster, while a hot SSD throttles to protect itself and, separately, loses retention margin.

Error rates: a client SSD is specified like a nearline hard drive

Uncorrectable error rates are the most directly comparable reliability figure, because both industries specify them the same way:

Device Specification Source
Laptop hard drive 1 unrecoverable sector per 10^14 bits read Seagate BarraCuda 2.5-inch datasheet
Client SSD UBER 10^-15 JESD218 application class table
Exos M nearline HDD 1 unrecoverable sector per 10^15 bits read Seagate Exos M product manual
Enterprise SSD UBER 10^-16 JESD218 application class table

A client SSD’s guaranteed error floor is ten times better than a consumer hard drive’s and the same as an enterprise hard drive’s, so against the disk a home buyer compares it with, the SSD holds the better number; only the enterprise SSD class is specified an order of magnitude beyond that. Across one whole-drive read:

1 per 10^14 bits, one full read of an 8 TB consumer disk
  8e12 bytes × 8 = 6.4e13 bits = 0.64 × 10^14
  expected bad sectors     0.64 per whole-drive read, at the spec limit

1 per 10^15 bits, one full pass of a 30 TB nearline disk
  30e12 bytes × 8 = 2.4e14 bits = 0.24 × 10^15
  expected bad sectors     0.24 per whole-drive read, at the spec limit

The specification is an upper bound, not an expectation - real drives are typically much better, or arrays would fail constantly - and it describes one sector, not one drive. Drives for a NAS works through what that does to a rebuild, and why an array is scrubbed.

What Backblaze’s fleet data shows

Backblaze publishes quarterly annualized failure rates for a fleet large enough to mean something, which makes it the only public dataset of its kind. The Q1 2026 report gives a quarterly AFR of 1.24 percent, down from 1.42 in Q1 2025, and a lifetime AFR of 1.39 percent across 529,968,464 drive-days. Full-year 2025 came in at 1.36 percent, down from 1.55 in 2024 and the best since 1.37 in 2022, across 344,196 drives and 30 models. Drives of 20 TB and above ran at 0.85 percent AFR across more than 86,000 units - better than the fleet average, not worse, which contradicts the standing folk belief that bigger drives are less reliable.

The average hides a spread that matters far more than the average does. In the same quarter:

Model Capacity AFR Basis
Toshiba MG11ACA24TE 24 TB 0.22% quarterly
WDC WUH722222ALE6L4 22 TB 0.38%, across 45,000+ units quarterly
Seagate ST10000NM0086 10 TB 3.13% lifetime
Seagate ST14000NM0138 14 TB 5.78% lifetime

That is a twenty-six-fold spread between models inside one fleet, with one operator - though two of those four rows are lifetime rates rather than quarterly ones, so it is a spread across unlike quantities rather than a clean ratio. Model and vintage dominate brand, and neither is knowable in advance, so any sentence of the form “brand X is more reliable than brand Y” is doing violence to that table.

On SSDs against hard drives, the unadjusted Q2 2021 figures (0.59 percent for SSDs at an average age of 14.2 months, 2.19 percent for hard drives at 52.4 months) compare a young population against an old one. Age-adjusted, the figures were 0.59 percent for SSDs against 1.27 for hard drives, with 95 percent confidence intervals of roughly plus or minus 0.5 on each, so the two intervals overlap and the difference is not significant. Backblaze’s own written conclusion is that the difference “was certainly not enough by itself to justify the extra cost.”

Three caveats come with this data, on top of the fact that none of it is a laptop: these drives sit in data center racks.

  • The SSD population is small and atypical. Lifetime SSD AFR of 0.90 percent over Q4 2018 to Q2 2023 covers about 3,144 drives, and they are boot drives - a light, sequential, mostly-read workload that resembles nothing a buyer does with a data SSD.
  • Day-one failures are missing. Backblaze writes that they have “probably been completely under-reporting day one failures”, and covers the gap with a pre-deployment qualification process you do not have.
  • A population statistic is not a prediction about your drive. Seagate says so itself in the Exos M manual: the annualized failure rate of 0.35 percent is “a population statistic not relevant to individual units”, and it assumes 8,760 power-on hours a year at a drive temperature of 10 to 30 °C. Against that 0.35 percent, Backblaze measures 1.24 percent across a real fleet: roughly three and a half times worse than specification is what observation looks like, in a professionally operated data center.

HDD vs SSD for backup and archiving

An SSD stores bits as trapped charge, and charge leaks. A hard drive stores bits as magnetic domain orientation, which does not. That is why the backup drive in the drawer should have platters. JEDEC specifies the floor for flash in JESD218, and the application class table is the whole of it:

Class Active use Powered-off retention FFR UBER
Client 40 °C, 8 hours a day 1 year at 30 °C 3% 10^-15
Enterprise 55 °C, 24 hours a day 3 months at 40 °C 3% 10^-16

Three things about that table are routinely misread.

It is a post-endurance requirement. JESD218 defines the endurance rating - the TBW number on the box - as the maximum host writes at which the drive simultaneously maintains its capacity, its class UBER and its functional failure requirement, and retains data with power off for the required time. The one year is not what a new drive does. It is what the drive must still do after being written to its full rated endurance. A lightly used drive has far more intact tunnel oxide and retains charge for very much longer. How much longer is not published by anybody, for any consumer drive.

Temperature dominates. The JEDEC presentation by Seagate’s Alvin Cox that spawned the “SSDs lose data in a week” scare charts retention against active-use and power-off temperature. The commonly quoted halving of retention per 5 °C is read off that chart; applied to the client specification it gives about six months at 35 °C and about three at 40 °C - a rule of thumb for a worn-out drive in a hot room, not a measurement of yours.

Enterprise is worse, not better. Enterprise drives are designed on the assumption that they are always powered, so more of the budget goes to performance and write cycles rather than retention margin. A decommissioned enterprise SSD is the worst thing you could choose for a shelf, and its low used price per terabyte makes that a trap rather than a bargain.

Platters have no equivalent decay mechanism on any timescale you care about. What actually kills a shelved hard drive is mechanical: lubricant migrating off the platter surface, bearing grease settling, and stiction on a drive that has not turned in years. The failure is “it does not spin up”, not “the data faded”.

Tape is the only medium in this comparison designed for the job. LTO-10 ships in 30 TB and 40 TB native cartridges, and the shelf-life figure that circulates for tape does not trace to the LTO consortium’s own published material; neither hard drive nor SSD vendors publish one at all.

For cold storage, the count matters more than the medium. Two copies on different drives in different places, checksummed, and verified on a schedule, beat any single medium’s specification. If you are choosing one medium anyway, choose hard drives, power them up once a year, and read the whole surface when you do - which, on an 8 TB drive, is a long night and on a 30 TB drive the forty-hour pass, so schedule it rather than starting it on a Friday.

HDD vs SSD in a NAS: power, noise and vibration

A NAS is the hard drive’s home ground, and drives for a NAS is this site’s full treatment. A few things matter in a box full of drives that never come up for one desktop SSD.

Many users means many random streams. Forty concurrent readers produce forty uncorrelated address streams, which pushes a hard drive toward its full seek-plus-rotation cost on every access while an SSD answers them across its flash dies in parallel. At 2.5 IOPS per terabyte, no capacity of hard drive supports many readers.

Spin-up sizes the power supply. The running figures are unremarkable: 6.763 W idle and 9.269 W under random read on the 32 TB Exos M. The number that sizes an enclosure is the startup current, a typical 2.6 A peak on the +12 V rail, which is 31.2 W a drive. Twelve spindles starting together want 374.4 W against 111.2 W under random read, so a twelve-bay enclosure needs three and a half times the 12 V capacity at the moment of power-on that it needs for the rest of its life. The fix is PUIS, Power-Up In Standby, an ATA feature that lets a drive come up without spinning until the host tells it to, or a backplane that staggers spin-up two drives at a time, turning a 31.2 A inrush into 5.2 A.

The workload rating is the hard drive’s DWPD. The Exos M manual specifies under 550 TB a year, reads and writes alike, and says exceeding it “may degrade drive MTBF and impact product reliability”. That is about eighteen full passes of a 30 TB drive, 8.3 times stricter than the 4,586 TB a year of Solidigm’s read-intensive 30.72 TB D5-P5316, and a monthly scrub of a 32 TB drive spends 70 percent of it alone. Drives for a NAS works that through.

Noise. Hard drive acoustics are published in bels of sound power, not the decibel figure a phone app measures: 2.8 bels at idle and 3.2 seeking, typical, on the nearline Exos drives. Idle noise can rise to operating levels during SMART offline activity, so a drive that gets audibly busy while nothing is using it is usually doing its background self-test, not failing.

Vibration. The Exos M manual promises more than 90 percent of its write throughput under rotary vibration, so losing up to a tenth of your throughput to vibration is the specified behavior of a good enterprise drive in a vibrating chassis (enterprise drives has the figures). Seagate’s own NAS drive guidance recommends rotational vibration sensors for systems with more than five drives, and flash has nothing to shake off-track at all.

Per dollar, the hard drive wins on capacity and nothing else

Hard drives do not win anywhere on performance economics. Take the median prices from the price section, a new 8 TB 3.5-inch disk and a new 2 TB M.2 SSD, and divide three ways. The speeds are the rated figures above, the IOPS the queue-depth-one ones:

8 TB 3.5-inch hard drive, median $61 a terabyte, about $485 a drive
  per terabyte                               = $61 / TB
  per MB/s            485 / 190              = $2.55 per MB/s
  per random IOPS     485 / 60               = $8.08 per IOPS

2 TB M.2 NVMe SSD, median $220 a terabyte, about $440 a drive
  per terabyte                               = $220 / TB
  per MB/s            440 / 7,250            = $0.06 per MB/s
  per random IOPS     440 / 14,000           = $0.03 per IOPS

ratios
  dollars per terabyte      HDD wins          3.6×
  dollars per MB/s          flash wins         42×
  dollars per random IOPS   flash wins        257×

The hard drive’s advantage is capacity and only capacity. A backup or a media library needs terabytes and tens of MB/s, so it stays on platters on economics rather than on tradition. A system drive needs IOPS, and no arrangement of hard drives fixes that, because buying IOPS from platters means buying spindles, and each spindle arrives with terabytes you did not want and watts you must pay for.

The data center comparison

Enterprise drives are the clean comparison at the other end of the market, because both are bought by the same buyer for the same rack. Price trackers and market reports put enterprise NVMe at $117 to $188 a terabyte and nearline hard drives at $12.50 to $16.63, roughly 10 to 1 on the midpoints. On this site on November 1, 2026, used, a 3.84 TB U.2 drive had a median of $253 a terabyte and an 8 TB 3.5-inch SAS drive $32, about 8 to

  1. At the extreme high end one index put a 30 TB enterprise SSD at a $22,600 reference price in August 2026 against $1,216 for an equivalent-capacity hard drive: $753 per terabyte against $41, a ratio of 18.6. On Solidigm’s and Seagate’s rated speeds that same pair comes out at 1.37 to 1 for flash on dollars per MB/s and 543 to 1 on dollars per random IOPS.

The two hard drive figures, $12.50 to $16.63 and $41, come from different sources and differ by a factor of two and a half. When two respectable sources for one category differ by that much, the category is being allocated rather than priced. Enterprise drives covers what the class buys. One data center argument does not transfer to a home: rack density favors flash on total cost of ownership - one analysis put a 10 PB deployment at 19 racks all-flash against 102 racks on disk - but for a buyer paying no rack rent and no cooling contract, that argument is worth exactly nothing.

Putting an SSD in front of hard drives

On a desktop the plain version needs no software: put what you wait on on the SSD and the library on the disk. Putting flash in front of platters as a cache works only under one condition: the hot working set has to be much smaller than the capacity, and it has to be reused. A cache does nothing for a linear backup pass, a media scan, or a first-run import, because every block is cold exactly once.

Two different things get the same name. A cache holds a copy; losing the device costs you performance. A tier holds the only copy of what is on it; losing the device costs you data. On ZFS, L2ARC and SLOG are the first kind and a special vdev is the second, so mirror it or lose the pool. It is also the arrangement that most reliably justifies itself on a hard drive pool, because directory traversal is the worst thing you can ask an actuator to do.

The hybrid drive (SSHD) carried a few gigabytes of onboard cache and has largely gone out of production. The arrangement that most often disappoints is a large L2ARC bought instead of RAM. Add RAM first; it is the same cache one level up - the memory site’s caching guide works out how much, and NVMe caching covers when a flash cache earns its place after that.

The one thing neither technology will do is make a single copy safe. Flash fails suddenly and platters fail slowly, the tail of a flash read is a hard drive’s median, and the plan that survives both is the same plan.

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.