A 7,000 MB/s rating does not mean an NVMe drive will open a photo project seven times faster than a SATA SSD. That figure describes a sequential transfer under favourable conditions; daily work mixes small files, waiting and transfers of different sizes, often with few concurrent requests.
What does that number measure?
Bandwidth is how much data passes per unit of time. Operations per second and latency describe something else: how many independent requests the drive can serve and how long each operation takes. Block size, queue depth and the read/write mix determine the result, which is why two tests with different numbers can both be correct.
Kingston makes a methodological point: the test must resemble the real environment and workload ( Kingston ). It holds for a data centre and for a home PC.
Why a fast NVMe sometimes feels slow
Possible causes include cache behaviour, memory management and temperature. Check the application, connection and background activity before blaming the drive.
The SLC cache. Many TLC or QLC SSDs use some cells in single-bit mode to accelerate short writes. Sustained speed may fall when this reserve is exhausted. Cache size, free space, firmware and workload matter: no fixed amount of data produces the same effect on every model.
Mapping data. Some drives without dedicated DRAM use HMB, when supported, to access a portion of system RAM. Not all do so, and DRAM alone does not predict performance. Compare the actual workload.
Thermal throttling. A controller may reduce performance at firmware-defined temperature limits. A heatsink or a position beneath a graphics card cannot predict the outcome by itself. Observe temperature and performance during work.
NVMe is a protocol, not a speed
NVMe defines protocols and interfaces for non-volatile storage; local SSDs commonly use it over PCIe ( NVM Express ). Parallel request support does not guarantee the same benefit in every application. The complete workload matters, not only advertised bandwidth.
| Situation | What matters | What to look at |
|---|---|---|
| Opening a project with many files | random access, latency | IOPS at low queue depth |
| Copying a large archive | bandwidth and SLC cache | sustained write speed after the cache fills |
| Editing high-bitrate video | sustained bandwidth and heat | throttling curves, heatsink |
| Databases or virtual machines | latency and endurance | TBW rating, DRAM |
Set up a test around your workload
Reproduce the task you want to speed up using non-critical data. Record configuration, free space, temperature and software. Repeat under comparable conditions and distinguish operating-system cache from drive cache: two runs alone do not isolate them.
Free space can affect writes and internal maintenance, but there is no universal 70% threshold. State the conditions compared and prepare a verified backup before testing.
Repeat the same workload after the drive warms up and with roughly the free space you expect in daily use. That is more informative than the advertised peak alone. Compare the time saved on your work with the replacement drive’s cost, capacity and compatibility before deciding to upgrade.
Sources
- How to test an enterprise SSD — metrics ↗ — Kingston · accessed 2026-11-01
- Use your exact environment ↗ — Kingston · accessed 2026-11-01
- NVM Express specification overview ↗ — NVM Express Inc. · accessed 2026-11-01
- Host Memory Buffer (HMB) technology ↗ — KIOXIA · accessed 2026-11-01
- Internal SSD product information: Dynamic Thermal Guard and TurboWrite ↗ — Samsung · accessed 2026-11-01


