Sequential read speed sells drives. Almost nothing you do all day is a sequential read. Here is what actually determines whether storage feels fast — and the two workloads where a spinning hard disk is still the correct purchase in 2026.
Two drives sit on our bench. One advertises 14,000 MB/s and costs three times as much as the other, which advertises 5,000 MB/s. In every test that resembles something a person actually does — booting, launching applications, loading a game level, opening a large project — the difference between them is between zero and 1.1 seconds. In one test, the expensive drive was slower, because it thermally throttled.
This is not a criticism of fast drives. It is a criticism of the single number used to sell them, which measures an operation that barely occurs outside benchmarks.
Why the number on the box is nearly useless
The headline figure is a sequential read at high queue depth: one enormous contiguous file, dozens of requests in flight simultaneously, ideal conditions. That describes copying a large video file. It does not describe booting an operating system, which reads thousands of small files scattered across the drive, mostly one or two requests at a time.
We ran both kinds of test on the same three drives:
| Drive | Advertised sequential | Measured sequential | Mixed low-QD | Cold app launch |
|---|---|---|---|---|
| PCIe 5.0 flagship | 14,000 MB/s | 13,410 MB/s | 241 MB/s | 2.8 s |
| PCIe 4.0 mid-range, DRAM | 5,000 MB/s | 4,880 MB/s | 228 MB/s | 2.9 s |
| PCIe 4.0 budget, DRAM-less | 3,500 MB/s | 3,390 MB/s | 84 MB/s | 6.1 s |
| SATA SSD | 560 MB/s | 545 MB/s | 176 MB/s | 3.4 s |
| 7,200 rpm HDD | 240 MB/s | 218 MB/s | 1.9 MB/s | 41.7 s |
Read that table twice. The flagship is 2.8× faster than the mid-range drive on the sequential test and 5% faster on the test that resembles real use. Meanwhile the budget DRAM-less drive, which looks respectable on paper, is 2.7× slower than its mid-range sibling on the mixed workload — and slower than a SATA drive costing less.
Sustained write speed after the SLC buffer is exhausted, and random read performance at queue depth 1–4. These two numbers determine how a drive feels, and neither appears in marketing material. Look for reviews that measure them — if a review only reports the headline sequential figure, it has told you almost nothing.
Queue depth: what your workload actually looks like
Queue depth is how many I/O requests are outstanding at once. NVMe was designed for enormous parallelism — 65,535 queues of 65,535 commands each — and consumer workloads use almost none of it.
We instrumented real sessions on three machines for a week:
| Activity | Typical queue depth | Dominant pattern |
|---|---|---|
| Booting Windows | 1–4 | Random read, small blocks |
| Launching an application | 1–8 | Random read, mixed sizes |
| Loading a game level | 4–16 | Mixed read, medium blocks |
| Compiling a large project | 2–12 | Random read + write |
| 4K video editing, timeline scrub | 8–32 | Sequential-ish read |
| Copying one huge file | 32–128 | Pure sequential |
| Benchmark headline figure | 32–256 | Pure sequential |
Everything a person does day-to-day lives in the top half of that table. The marketing number lives at the bottom. That single mismatch explains why upgrading from a good PCIe 4.0 drive to a flagship PCIe 5.0 drive feels like nothing at all.
The DRAM cache, and why cheap drives stutter
An SSD maintains a mapping table translating logical addresses to physical flash locations. On a 2 TB drive that table is roughly 2 GB. Where it lives determines how the drive behaves under random access.
- DRAM-equipped drives keep the table in dedicated on-board memory. A lookup is immediate.
- DRAM-less drives keep it in flash and borrow a slice of system memory through Host Memory Buffer. Lookups cross the PCIe bus, and on a cache miss they hit flash.
For sequential reads this is irrelevant — addresses are predictable. For random reads it is decisive, and random reads are what application launches and game asset streaming consist of. It is why the budget drive in our table posts a respectable sequential figure and a dismal mixed one.
They are acceptable as a media archive or a secondary game library for titles you play rarely. As a boot drive, a development drive or a primary game drive they produce exactly the stuttering that people later blame on their processor or graphics card. The price difference is around £20 on a 1 TB drive.
The SLC write buffer and the cliff behind it
Consumer drives write incoming data to a fast pseudo-SLC region first, then reorganise it into denser storage in the background. The advertised write speed is the speed of that buffer. When the buffer fills, you fall off a cliff.
Sustained-write test on a 2 TB drive, writing continuously until the buffer was exhausted:
| Data written | Write speed | State |
|---|---|---|
| 0–220 GB | 4,880 MB/s | SLC buffer |
| 220–280 GB | 1,640 MB/s | Buffer folding, transition |
| 280 GB+ | 760 MB/s | Native TLC write |
| Same drive at 90% full | 310 MB/s | Small buffer, heavy folding |
Two things follow. First, the buffer is dynamic — it shrinks as the drive fills, because the fast region is borrowed from free space. A drive that sustains 4,880 MB/s when half empty may manage 310 MB/s when nearly full. Second, if you routinely write more than a couple of hundred gigabytes at a stretch — video ingest, large dataset work, disk imaging — the post-buffer figure is your real write speed, and almost nobody publishes it.
Copy a 300 GB folder to it and watch the transfer rate. You will see the moment the buffer runs out — the speed drops sharply and stays there. That second number is your drive's real sustained write speed, and on many budget drives it is genuinely shocking.
TLC, QLC and what endurance ratings mean
Flash cells store multiple bits by holding distinguishable voltage levels. More bits per cell means more capacity per wafer and less margin for error.
| Type | Bits/cell | Voltage states | Rated P/E cycles | Typical role |
|---|---|---|---|---|
| SLC | 1 | 2 | ~100,000 | Caches and industrial use |
| MLC | 2 | 4 | ~3,000 | Largely gone from consumer drives |
| TLC | 3 | 8 | ~1,000 | The consumer default |
| QLC | 4 | 16 | ~300 | High-capacity, read-heavy |
Is QLC actually a problem?
Less than the numbers suggest, for most people. Endurance is expressed as TBW — terabytes written over the warranty. A typical 2 TB QLC drive is rated around 450 TBW. Our instrumented desktop machines averaged 34 GB written per day, which works out at about 36 years to reach that rating.
The real QLC issue is not wear, it is the sustained-write cliff: QLC drives have a larger SLC buffer and a much slower native write speed behind it, sometimes below 150 MB/s. For a game library or a media archive — written once, read many times — that never matters. For a scratch disk or a video ingest drive it matters a great deal.
A large QLC drive as a game library or media archive is a sensible purchase: cheap per terabyte, written rarely, read constantly. The mistake is using one as a scratch disk, a video ingest target or a database volume, where the post-buffer write speed becomes the number that governs your day.
PCIe 5.0, 4.0 and SATA in practice
| Interface | Practical ceiling | Real-world benefit | Caveats |
|---|---|---|---|
| SATA III | ~550 MB/s | Transformative over HDD; fine for most tasks | Cheap, universal, ageing |
| PCIe 3.0 ×4 | ~3,500 MB/s | Indistinguishable from 4.0 in daily use | Still excellent value |
| PCIe 4.0 ×4 | ~7,400 MB/s | The current sweet spot | Widely supported, cool-running |
| PCIe 5.0 ×4 | ~14,000 MB/s | Only for genuinely sequential work | Often needs active cooling; premium price |
The thermal caveat deserves emphasis. Several PCIe 5.0 drives we have tested throttle within ninety seconds of sustained load without a substantial heatsink, and some need a small fan. That means another source of noise in your case, and a drive whose sustained performance is worse than a well-cooled PCIe 4.0 unit. Under a graphics card, in a cramped case, this is a real and common outcome.
Motherboard M.2 slots frequently share PCIe lanes with the primary graphics slot or with SATA ports. Populating the wrong slot can drop your graphics card to ×8 or disable two SATA ports. The manual has a table; read it before ordering, not after.
When a hard disk is still the right answer
Two cases, both about cost per terabyte for data you rarely touch.
Bulk archival storage
At the time of writing, a high-capacity hard disk costs roughly $0.012 per gigabyte against $0.055 for a mid-range SSD. For 20 TB of photographs, video masters or backups, that is a difference measured in hundreds of pounds. The data is written once and read occasionally; sequential throughput of 200 MB/s is entirely adequate.
Backup targets
Backups are sequential writes and rare reads — the one workload where mechanical drives are genuinely competitive rather than merely cheaper. And an SSD left unpowered for long periods can lose charge over time, which makes a cold-storage hard disk arguably the safer medium for a drive that sits in a drawer.
Where hard disks are now indefensible: as a boot drive, for applications, for game libraries, or for any interactive work. Look again at that 41.7-second cold application launch. Life is short.
The most cost-effective configuration for most people with large media libraries: a 2 TB NVMe drive for the operating system, applications and current projects, plus a high-capacity hard disk for everything finished. Fast where speed is felt, cheap where it is not.
Matching a drive to five real workloads
| You mostly do this | Buy this | Why |
|---|---|---|
| Browsing, office work, media | 1 TB PCIe 4.0 with DRAM, or a good SATA SSD | Latency dominates; interface does not matter |
| Gaming | 2 TB PCIe 4.0 with DRAM | Random reads at low QD; capacity prevents the full-drive cliff |
| Photo and video editing | 2 TB PCIe 4.0 scratch + 8 TB+ HDD archive | Fast scratch, cheap masters |
| Software development | 2 TB PCIe 4.0 with DRAM, high TBW | Constant small reads and writes; endurance matters |
| Home server / NAS | HDD array + small SSD for cache and metadata | Capacity economics with a fast index |
Notice that PCIe 5.0 appears nowhere. There is one workload where we would recommend it — sustained ingest of very large sequential files, such as multi-camera high-bitrate video — and even then only with a proper heatsink and airflow.
Reading drive health before it fails
SSDs usually fail gradually and announce it clearly, if you look. Four SMART attributes are worth checking twice a year:
- Percentage Used — the controller's own wear estimate. Under 10% after several years is normal for desktop use.
- Available Spare — reserve blocks remaining. A decline below 100% means blocks are being retired. Below the threshold, replace the drive.
- Media and Data Integrity Errors — uncorrectable errors. Should be zero. Any non-zero value that grows is a replacement, today.
- Unsafe Shutdowns — power loss without flush. A high count points at a power supply problem rather than a drive problem.
# Linux / macOS with smartmontools
sudo smartctl -a /dev/nvme0
# Windows, PowerShell
Get-PhysicalDisk | Get-StorageReliabilityCounter |
Select DeviceId, Wear, ReadErrorsTotal, WriteErrorsTotal, PowerOnHoursSet a recurring reminder. Five minutes with a SMART tool twice a year will usually give you weeks of warning before a drive fails outright, which is the difference between a planned migration and an unplanned data recovery.
A buying checklist you can use in a shop
- Does it have a DRAM cache? If the specification does not say, assume not. This is the single biggest divider between drives that feel fast and drives that do not.
- What is the TBW rating? Under 300 TBW on a 1 TB drive signals low-endurance flash.
- What is the warranty? Five years is standard. Three suggests the manufacturer knows something.
- What is the sustained write speed after the buffer? Rarely published. If a review does not test it, assume it is unimpressive.
- Does your motherboard slot run at full width? Some M.2 slots share lanes with the graphics card or drop to ×2. Check the manual.
- Is there a heatsink, and room for it? Especially under a graphics card.
- Buy one size larger than you need. Performance collapses above about 80% utilisation, so a 2 TB drive holding 1.4 TB is comfortable while a 1 TB drive holding the same data is not.
The two questions that separate a drive you will be happy with from one you will quietly resent: does it have DRAM, and is it large enough to stay below eighty per cent full. Everything else is detail. Daniel Okonkwo, Hardware Lead
No amount of endurance rating or SMART monitoring changes this. Drives fail suddenly, get stolen, get encrypted by ransomware and get deleted by accident. Keep three copies on two kinds of media with one off-site — the approach we set out in the home network security guide.
Storage is the component where specification sheets diverge most sharply from experience, and where the cheapest possible choice does the most damage to how a machine feels. Get DRAM, get capacity, keep it below eighty per cent, and put the money you saved on PCIe 5.0 into a better graphics card or more memory.
Questions readers keep asking
For almost no one in 2026. Our measured difference against a good PCIe 4.0 drive was 5% on mixed workloads and 0.1 seconds on application launches. The exception is sustained ingest of very large sequential files — multi-camera high-bitrate video, large dataset transfers — where the sequential advantage is real. Even then, budget for a substantial heatsink and airflow, because several PCIe 5.0 drives throttle within ninety seconds without one.
Far longer than the TBW rating implies for normal use. Our instrumented desktops averaged 34 GB written per day; against a typical 600 TBW rating on a 2 TB drive that is roughly 48 years of writes. In practice drives are replaced because they are too small, not because they wear out. Heavy write workloads — video scratch disks, busy databases, virtual machine hosts — are the exception and deserve a high-endurance drive.
Not for performance — there is no mechanical seek time to optimise, so partitioning gains you nothing there. It can be useful for organisation, for dual-boot setups, or to make imaging simpler. Just remember that free space is shared across the whole drive for wear levelling and SLC caching, so a nearly-full partition on a mostly-empty drive is less of a problem than it looks, while a nearly-full drive is a real one.
Three common causes, in order of likelihood. It is more than 80% full, so the SLC buffer has shrunk and garbage collection is working constantly. It is thermally throttling, especially if it sits under a graphics card. Or TRIM has been disabled — occasionally by a third-party “optimiser” utility. Check free space first; that is the answer roughly three times in four.
Only for bulk data you rarely touch. At roughly a fifth of the cost per terabyte, hard disks remain the sensible choice for photo and video archives, media libraries and backup targets — all sequential, all rarely read. For anything interactive, they are indefensible: our cold application launch took 41.7 seconds on a 7,200 rpm drive against 2.9 seconds on a mid-range NVMe.
