Burst speed and sustained speed are two different products

This works through peak and sustained throughput in the order the parts actually depend on each other.

Why a burst exists at all

Almost every device that transfers or processes data keeps a small fast area in front of a larger slow one. That buffer accepts work at high speed and releases it to the slower stage in the background, which is invisible while the buffer has room.

Short tasks therefore complete at the speed of the buffer, and that speed is genuine for exactly as long as the buffer lasts. Because most benchmark runs and most marketing tests are short, they measure the buffered rate rather than the underlying one. The buffer is not a trick; it is a sensible design that makes common small operations fast, and it stops helping at a definable point.

What happens when the buffer fills

Once incoming work exceeds what the slow stage can absorb, the device can only run at the slower stage's rate. The transition is usually abrupt rather than gradual, which is why sustained transfers show a sharp step down partway through.

The size of the buffer therefore decides how long the fast phase lasts, and buffer size is frequently unpublished. Some designs vary the buffer dynamically with free space, so a nearly full device behaves quite differently from an empty one. This is why the same product can be described accurately as both very fast and rather slow depending on the task used to measure it.

Heat as the other limit

Sustained work generates heat, and every semiconductor has a temperature above which it must slow down to protect itself. Throttling is a designed behaviour rather than a fault, and its onset depends on the cooling arrangement rather than on the chip alone.

Judged against the category, the same component in a well-ventilated housing and in a sealed one produces very different sustained figures. Ambient temperature shifts the whole curve, so results measured in a cool laboratory do not transfer to a warm room. A product that cannot dissipate heat will always converge on the rate its cooling permits, whatever its peak specification says.

Access pattern changes everything

Sequential work reads or writes in long continuous runs, which suits almost every storage and transfer mechanism. Random access jumps between locations, and the overhead of each jump can dominate the time taken regardless of raw speed.

The measurable part is this: real workloads are usually a mixture, and the mixture matters more than either extreme figure taken alone. Headline figures are almost always the sequential ones, because they are the largest numbers a device can honestly produce.

Where a specification quotes both sequential and random figures with the test conditions, it is describing the product rather than advertising it.

The rest of the system decides the result

A chain of components runs at the speed of its slowest link, and the advertised part is rarely the one limiting you. Connections, protocols, controllers and software layers each add overhead that no component specification includes.

Judged against the category, negotiated link speeds fall back automatically when cables or ports do not support the maximum, usually without telling anyone. Power delivery can limit sustained performance in portable equipment, since the device may be unable to draw what it needs continuously. Upgrading the fastest component in a chain changes nothing, which is why identifying the bottleneck precedes any purchase decision.

Where a claim here has not been tested directly it is inference, and it should be read as inference.

Reading a speed claim

Look for whether the figure is described as peak, maximum, up to, burst or sustained, since those words are the specification. Check for the duration and the file or task size used, because a figure without a duration is a statement about the buffer.

Prefer measurements that show a curve over time rather than a single number, as the shape reveals both the buffer and the throttle point. Consider how you will actually use the device, since a pattern of short tasks genuinely does run at the buffered speed. Where sustained performance matters, evidence from long-duration testing is the only relevant evidence, and it is not on the box.

The takeaway

Ask how long the advertised speed lasts, because the answer is usually the specification you actually needed.

Buy for the failure you can live with, not the feature you will use twice.