What SSD speed numbers actually tell you

The headline number on an SSD box is sequential read: how fast the drive moves one large file when many requests are in flight at once. That is the condition your machine produces least often. Booting, launching an application, and most of what you actually wait on depend on a different figure entirely.

That does not make it useless. It makes it a narrow claim about one kind of work.

What sequential and random speed actually measure

Sequential and random are not two speeds of the same thing. They are quoted in different units, because different things limit them.

Sequential throughput is stated in megabytes per second, and it describes one long transfer moved end to end. Bytes per second is the right unit because the work is a single object of a known size, and how long it takes is that size divided by the rate.

Random performance is stated in IOPS — input/output operations per second — usually at a request size of 4 KB. Move the same total in 4 KB pieces and it stops being one object and becomes millions of them. Operations are the right unit because the bytes stop mattering. Each request carries a fixed cost: working out where the data physically sits, issuing the command, returning the answer. At 4 KB that cost is most of the time spent, so the drive is bounded by how many requests it can retire, not by how wide its path to the flash is.

There is a third figure, and almost nothing states it. Latency is the time one request takes when nothing else is waiting. Throughput and IOPS describe a drive under load; latency describes it when you are the only thing asking, which is most of a desktop's life.

Which kind of work your machine actually does

Datasheet figures are produced under load a desktop rarely creates: several workers issuing requests continuously, keeping the drive's queues full. Ordinary desktop use sits a few requests deep, often at one.

Queue depth is how many requests the drive has been handed and not yet answered. It is where the interfaces genuinely differ, though not in the way it is usually put. SATA is not one request at a time: native command queueing gives it a single queue holding up to 32 commands. What NVMe adds is many queues rather than one, each far deeper, with less host work per command. That is worth a great deal to a server. It is worth much less to one desktop session.

The reason a desktop cannot fill those queues is that its small reads depend on each other. Opening an application reads a file that names the next file to read. Booting walks a long chain of them. Each step has to finish before the next one is known, so they cannot be issued in parallel however many queues are available, and the total is a sum of latencies rather than a division by bandwidth. That is why a drive quoting several times another drive's headline figure does not open your mail client several times faster.

Why the number on the box is sequential read

Manufacturers print sequential read because it is the figure that behaves best under measurement. It is stable run to run, much the same on a full drive as an empty one, and bounded by the interface — so it rises cleanly with each generation and gives this year's drive something visibly larger to print than last year's.

It is also the cheapest claim to stand behind. Random performance depends on how the drive is being used, how full it is, and what it was doing a minute ago. Sequential read on a fresh drive is close to a property of the hardware.

Why write speed falls partway through a large copy

Write speed is not one number. Nearly every drive takes incoming data into a fast cache region and folds it into its slower permanent form afterwards, so there is a cached rate and a native rate underneath, and the box states the first. That region is the SLC cache, flash run at one bit per cell. How to choose an SSD covers what sets the rate underneath, and TLC, QLC, and DRAM-less SSDs covers where the cache comes from and why writing into it costs the drive a second write later.

What matters here is the shape of the curve. A copy starts at the cached rate, holds it while the cache lasts, then falls — sometimes by a little, sometimes to a fraction of it — and stays down for the rest of the transfer. Two things decide whether you meet that drop: how much you are writing in one go, and how much cache the drive has at that moment.

This is why a benchmark that writes only a gigabyte or two tells you little. That amount lands entirely in the cache on most drives, so the result is the cached rate however slow the drive is underneath. A test that ran until the cache was exhausted would report two rates and the point where the first became the second. Almost none do, and no datasheet does.

The cache also needs idle time to empty. Start a second large copy soon after the first and it can begin slower, because the drive is still clearing the first. Sizing an SSD covers the free-space side of it, the other lever on how much cache you get.

How to read a speed rating without being misled

Treat the rating as a tier marker rather than a score. It sorts drives into rough bands, and inside a band it says very little. The ratio between two figures does not survive contact with your workload: a drive rated twice as fast is not twice as fast at anything you do outside a large file copy.

A rating is a ceiling under favourable conditions, not a floor. It was measured on an empty drive, cool, with the queues kept full. Your drive is none of those things for most of its life.

Read the write figure next to the read figure rather than on its own. And when two drives are separated by a multiple rather than a margin, that is usually the interface generation talking, which is a fit question before it is a speed one. NVMe or SATA covers which one your machine takes.

What we compare on, and what we cannot see

We compare SSDs on three figures: capacity, sequential read, and sequential write. Higher is better on each of them, and that is the whole of what our comparison knows about speed.

We carry sequential read knowing it is the weakest number on the box. We carry it because it is the one figure virtually every datasheet states, at every price tier. Random figures and latency would tell you more, but they are stated by some makers on some drives, and a comparison that worked only on generously documented drives would drop the cheaper half of the market. That is not a comparison. It is a shortlist of drives with good paperwork.

That is the same trade we make when a deal score falls back to a manufacturer's list price and marks itself low confidence: a weaker figure that covers everything beats a better one that covers a third, as long as we say which one we used. How we score deals sets out that side of it.

Random read and random write IOPS are fields we hold on an SSD, but they are optional, and most listings we ingest never state them. Neither is something we compare drives on, and we do not filter on them either. That absence is the market's, not a judgement that the numbers do not matter.

Every SSD we track is on one page.