Why the same PCIe Gen4 SSD shows four different speeds
A single PCIe Gen4 SSD datasheet can quote a sequential read figure near 7,000 MB/s, a sequential write figure a few hundred MB/s lower, and random read IOPS with figures in the hundreds of thousands or millions. Those numbers describe four different things: the link, the flash, the controller's queue handling, and the workload the vendor chose to publish. Reading them as one "speed" is the most common way buyers mis-compare drives.
This guide walks through what each number actually measures on a PCIe Gen4 class M.2 or U.2 drive, which conditions must be identical before two figures can be set side by side, and which checks catch the misleading comparisons.
The ceiling first: what PCIe Gen4 x4 can physically deliver
Before judging any drive number, check the bus it sits on. PCI Express 4.0 transfers at 16 GT/s per lane and uses 128b/130b encoding. One lane therefore carries roughly 1,969 MB/s of encoded payload, and a four-lane link (x4, the standard for NVMe M.2 and most U.2 drives) carries about 7,877 MB/s in one direction.
- Gen4 x4 one-way theoretical ceiling: ~7,877 MB/s
- Gen4 x4 combined both directions: ~15,754 MB/s (not a read or write limit on its own)
- Gen4 x1: ~1,969 MB/s
That 7,877 MB/s figure is a link-level number. Transaction-layer packet headers, controller overhead, flash programming time and host software all consume part of it, which is why a well-designed drive typically publishes sequential reads in the low-to-mid 7,000 MB/s range rather than at the wire rate. Any datasheet claiming a single-drive sequential read above the x4 ceiling without explaining a cache burst is worth questioning.
Sequential throughput: where the number comes from
Sequential figures measure large, contiguous transfers — usually 128 KiB or larger block sizes — and they are dominated by how many flash dies the controller can keep busy in parallel. Practical consequences for interpretation:
- Capacity changes the number. A drive with more NAND packages has more parallelism. Vendors often publish the peak sequential figure for the largest capacity only; smaller SKUs of the same model can be materially slower. Always match the capacity the spec was stated for.
- Read differs from write, and write differs by state. Write figures frequently assume a fresh drive with SLC cache available. Once the cache is exhausted, sustained write can fall to a fraction of the published figure. Look for a separate "sustained write" or "post-cache write" value.
- Host platform matters. A CPU-attached M.2 slot and a chipset-attached slot, or a drive behind a riser or adapter, will not produce the same result. Datasheets usually note the platform used.
Random IOPS: the queue depth is the specification
Random figures are quoted in IOPS at a stated block size and queue depth (QD). A random read IOPS figure without those two parameters is uninterpretable. The common pairs you will see:
- QD1 (or QD1, single thread): latency-bound; this is the number that predicts everyday responsiveness in light desktop or single-stream application use.
- QD32 / QD256: parallelism-bound; this is what predicts behaviour under database, virtualisation or multi-stream workloads.
- Mixed read/write ratios (e.g. 70/30): often quoted for enterprise-class drives, and not comparable to a 100% read figure.
A drive can look excellent at QD256 and mediocre at QD1. If your workload is a single-user workstation, the QD1 figure — often published in microseconds of latency rather than IOPS — is the one that matters.
Do not confuse JEDEC endurance workloads with performance tests
Datasheets often cite JEDEC standards near the performance table, which invites the wrong inference. JESD219A (SSD Endurance Workloads), used together with JESD218, defines the endurance workload used to rate and verify drive endurance for Client and Enterprise categories. It does not prescribe a universal "4 KiB QD32" performance test. JEDEC's published category conditions differ sharply: Client is specified around 40 °C with 8 hours per day of use and a 1-year verification, while Enterprise is specified around 55 °C with 24-hour operation and a shorter verification period, with stricter uncorrectable bit error rate targets for enterprise.
So a drive that says "endurance rated per JESD219A Enterprise workload" and separately quotes "4 KiB, QD256" IOPS is telling you two different things. Neither statement lets you infer the other, and neither makes the numbers comparable to a drive quoting 4 KiB QD32 without stating whether it was pretested to steady state.
A comparison worksheet that prevents bad matches
Before placing two PCIe Gen4 drives side by side, confirm every row below is the same. If any row is unknown, treat the comparison as unverified.
| Field | What to record |
|---|---|
| Capacity / SKU | Exact model and capacity the figure was published for |
| Bus and form factor | PCIe Gen4 x4, M.2 2280 or U.2, host slot type |
| Block size | e.g. 128 KiB sequential, 4 KiB random (state whether 4,096 bytes) |
| Queue depth and threads | QD1 / QD32 / QD256, single or multi-thread |
| Read/write mix | 100% read, 100% write, or stated ratio |
| Drive state | Fresh, preconditioned, or steady state |
| Cache condition | SLC cache available, or post-cache sustained |
| Thermal condition | Heatsink fitted, ambient temperature, throttling noted or not |
| Test tool and platform | Tool name, OS, driver, CPU-attached or chipset-attached |
| Standard cited | Which JEDEC or vendor method, and whether it covers endurance or performance |
Thermal throttling: the hidden variable in every sustained test
Gen4 controllers run hot, and most datasheets publish peak figures from short bursts before the drive reaches its thermal limit. When a spec sheet mentions a heatsink, an airflow requirement or a throttling threshold, treat the headline number as a burst figure. For sustained workloads — video ingest, large dataset copies, NAS caching — ask for the sustained figure or measure it after the drive has been under load long enough to stabilise. This is also the reason two reviews of the same drive can disagree: one ran a short benchmark on an open bench, the other a long test inside a chassis.
Reading a datasheet in the right order
- Check the bus: Gen4 x4 gives you a ~7,877 MB/s one-way ceiling, so any sequential claim should sit below it.
- Check the capacity the figure belongs to.
- Check whether the number is burst (cache) or sustained.
- Check block size and queue depth for every IOPS figure.
- Check the drive state and thermal conditions.
- Check whether a cited JEDEC standard covers endurance, performance, or both — and do not transfer one to the other.
Only after those six checks are two numbers genuinely comparable. A lower figure measured under a heavier, better-documented condition is often the more trustworthy specification.