External SSDs: Which Spec Sheet Numbers Predict Real Speed, and Which Are Noise

01

The big number measures one thing you rarely do

The largest number on the box is sequential read speed: the rate at which the drive streams long, contiguous blocks of data while it is cool and mostly empty. That describes one real task, which is copying a few enormous files. Most of what people actually do is random access, meaning thousands of small reads and writes scattered across the drive. Opening a project folder, running software from the drive, or scanning a photo library all live there. Random performance is quoted in IOPS, or in much smaller megabyte figures at four kilobyte block sizes, and manufacturers usually leave it off the packaging.

There is a second reason to distrust the headline. Every conventional external drive sits behind a bridge chip that translates USB into the language the flash controller speaks, and that translation adds latency to every single request. Latency is what random performance is made of. The practical result is that two drives separated by a huge margin on sequential numbers can land within a few percent of each other on small-file work over the same port. If your workload is small files, you are mostly buying the enclosure's protocol support and bridge quality rather than the flash inside it.

Queue depth is the missing qualifier on nearly every random figure you will see quoted. Benchmarks that produce impressive IOPS run dozens of requests in flight at once, which is what a database server does and not what happens when you double click a folder. The figure that tracks how a drive feels is random read at queue depth one, single threaded, and it sits roughly an order of magnitude below the headline. An internal NVMe drive might manage sixty to ninety megabytes per second there. Put the same flash behind a USB bridge and it typically lands between twenty five and forty. A mechanical hard disk does about one.

02

The SLC cache cliff, and why big transfers slow down halfway

Modern flash stores three or four bits in each memory cell, which is cheap but slow to write. Controllers hide this by running a slice of the flash in a fast one-bit mode and dumping incoming writes there, folding the data down to its dense final form later while the drive is idle. That fast slice is the SLC cache. It is usually dynamic, sized as a fraction of the free space available, so it shrinks as the drive fills. On an empty one terabyte three-bit drive the fast window is often somewhere between one hundred and two hundred gigabytes. On the same drive at ninety percent full it can be a tenth of that.

How far the rate falls when the cache runs out depends on the flash. Three-bit designs usually settle between four hundred and nine hundred megabytes per second, still comfortably ahead of any spinning disk. Four-bit designs commonly collapse to around one hundred megabytes per second and stay there for the remainder of the copy, which is slower than the mechanical drive you were replacing. The number that matters is therefore not the peak but the post-cache write rate, which almost nobody publishes and reviewers have to measure. Work out the largest single transfer you routinely make. Twenty gigabytes onto a half-empty drive never meets the cliff. Unloading camera cards meets it every time.

One wrinkle applies only to external drives. Folding cached data down into dense storage happens during idle time, and an external drive spends most of its idle time unplugged. Copy two hundred gigabytes, eject immediately, and the housekeeping never runs. Next time you connect it the fast slice is still occupied, so the window before the cliff is shorter than it was, sometimes dramatically. Leaving the drive attached and powered for ten or fifteen minutes after a heavy write lets the controller catch up. This is also why a drive can benchmark beautifully once and disappoint on the second real transfer of the same day.

03

USB naming is broken, and your port decides your speed

Four speeds exist in practice: five, ten, twenty and forty gigabits per second. The five and ten gigabit tiers have been renamed repeatedly and surface on packaging as 3.1, 3.2, Gen 1, Gen 2 and assorted marketing labels that all reduce to those same two things. Twenty gigabit, sold as USB 3.2 Gen 2x2, is a dead branch. It required a specific host controller that mainstream laptops never shipped, it was left out of the USB4 specification, and on an unsupported machine those drives quietly negotiate down to ten gigabits with no warning anywhere in the interface.

Forty gigabit ports, badged either USB4 or Thunderbolt, are the genuine step up. Enclosures from the Thunderbolt 3 era typically top out near two thousand six hundred to two thousand eight hundred megabytes per second, limited by the four lanes of PCIe behind the tunnel. Current USB4 bridge chips go further, past three and a half thousand on a capable host. They also cost more, run considerably hotter and pull more power from a laptop battery. Buy for the port you own today. Protocol overhead matters too: a ten gigabit link carries about one thousand two hundred and ten megabytes per second after its encoding scheme, and real transfers land nearer one thousand.

Then there is the cable. A USB-C connector says nothing about the wires behind it. Charging cables and many bundled cables carry only USB 2.0 data lines, which cap transfers near forty megabytes per second no matter what the drive and port can do. This is by far the most common reason a new SSD feels slow, and it is invisible until you check. Keep the cable that shipped with the drive, label it, and resist grabbing whatever is nearest when you travel. Length matters as well, since passive cables above about a metre often drop to a lower tier rather than failing outright.

• 5 Gbps: 500 MB/s on paper after 8b/10b encoding, realistically 400 to 450 MB/s.
• 10 Gbps: about 1,210 MB/s on paper after 128b/132b encoding, realistically 950 to 1,050 MB/s.
• 20 Gbps needs a host controller almost no laptop shipped, and unsupported machines silently fall back to 10 Gbps.
• 40 Gbps USB4 or Thunderbolt: around 2,800 MB/s from Thunderbolt 3 era enclosures, over 3,500 MB/s from current USB4 bridges.
• Hubs, docks and monitor pass-through ports routinely negotiate a slower tier than the port built into the laptop.

04

Heat, throttling, and when the metal shell actually earns its price

Sustained writing heats the controller and the bridge chip far more than the flash itself, and firmware responds by cutting speed to keep silicon inside its temperature limit. Some compact drives begin throttling within a minute of continuous writing. A metal body is not styling in this category. It is the heatsink, and it separates a drive that holds its rate through a large transfer from one that saws up and down. Vented plastic with a thermal pad bridging the board to the shell works about as well. Sealed plastic with an air gap around the board does not.

Here is the part that saves money: throttling is mostly a fast-interface problem. On a ten gigabit connection the link itself limits throughput to a level that generates modest heat, so many drives never reach their throttle point at all. On a forty gigabit connection the same flash is pushed three times harder and heat becomes the binding constraint rather than the flash or the link. If you are buying a ten gigabit drive, chassis material is a minor consideration. If you are buying a USB4 class drive, it is among the most important things about it.

Two habits help more than any purchase. Leave the drive on a hard surface while it works rather than on a laptop lid or inside a padded sleeve, because these enclosures shed heat through the shell and soft materials trap it. Then break enormous transfers into batches where you can, since throttling is fully recoverable and a few minutes of cooling restores the opening rate. Worth knowing too: the bridge chip throttles on its own schedule, independently of the flash controller, and the fastest bridges run hottest. A drive too warm to hold comfortably after ten minutes of writing is not faulty, but it is showing you where its ceiling comes from.

05

DRAM-less controllers cost you more externally than internally

Solid state drives keep a map of where data physically lives, and that map is consulted constantly. Drives with a DRAM chip hold it in fast memory. DRAM-less drives keep it in flash and cache fragments in a small amount of on-controller SRAM. Inside a computer, a DRAM-less drive can borrow system memory through a mechanism called host memory buffer. Across a conventional USB bridge that route is unavailable, because the bridge translates NVMe commands into SCSI ones and nothing in that translation exposes host memory. Going DRAM-less therefore costs you more in an external drive than it would in an internal one.

The penalty shows up as inconsistency rather than a lower top speed. Random reads suffer, the drive gets slower as it fills, and long mixed workloads turn uneven in a way benchmarks with a fresh empty drive never reveal. For a drive that receives large sequential backups once a week, none of that is worth paying to avoid. For a drive you edit from, run software from, or keep permanently three quarters full, it is worth the difference. Manufacturers rarely state which they have shipped, so read an unusually low price beside a high sequential figure as a strong hint.

One exception is worth knowing. A Thunderbolt or USB4 enclosure tunnels PCIe rather than translating storage commands, so the drive inside appears to the computer as a genuine NVMe device. Host memory buffer can work in that arrangement, though whether it does depends on the host operating system and the enclosure firmware rather than anything printed on the packaging. Everything connected through an ordinary USB bridge, which covers nearly every sealed portable drive sold, has no path to system memory at all. That is a real reason the cheap-controller penalty is harsher in this product category than reviews of internal drives suggest.

06

Match the drive to the job, then buy one size up

Decide which of three jobs the drive has before comparing anything. A backup target receives big sequential writes occasionally and sits idle otherwise, so sustained write rate and capacity matter while random performance barely registers. A working drive holds active projects and is read and written all day, which puts random performance and DRAM first. A drive used to run games or applications directly sits between the two and cares most about consistent random reads at low queue depth, the least advertised specification of the lot and the one most likely to separate two drives with identical box copy.

Capacity deserves a specific warning. Because the fast write cache is carved out of free space, a drive kept near full is slower than the same model kept half empty, sometimes by a factor of five. Buying one size larger than you need is a performance decision as much as a storage one, and it extends life because writes spread across more cells. The endurance rating, usually given in terabytes written, is almost always far beyond what a person reaches with an external drive, so read it as a warranty term rather than a lifespan estimate.

Free space is the one lever you still control after purchase. Keeping roughly a quarter of the drive empty preserves a usable write cache and gives the controller room to shuffle blocks during garbage collection. Deleting files does not automatically hand that space back, either: the TRIM command that tells the flash a block is dead only reaches the drive if the bridge chip implements it, and cheaper bridges do not. Where TRIM is missing, a drive that has been filled and emptied several times gradually loses write speed and never recovers it. Enclosures that support the faster USB storage protocol usually pass TRIM through correctly.

07

Verify what you actually bought in the first ten minutes

Test the drive while you can still return it, because every failure described here looks identical from the outside: it just feels slow. Open your operating system's system information panel and read what the port negotiated, which is reported as a link speed in gigabits. Seeing five where you expected ten usually means the cable, a hub, or a front-panel port wired to an older internal controller. Swap one variable at a time. Plugging directly into the machine with the supplied cable is the baseline that tells you whether the drive itself is the problem.

Then run a transfer that is deliberately larger than the cache, ideally forty or fifty gigabytes of real files rather than a synthetic benchmark, and watch the rate through the whole copy instead of glancing at the first few seconds. The shape of that curve is the specification nobody prints: where it steps down, how far, and whether it recovers. Repeat the same copy with the drive around three quarters full if you intend to live with it that way, since a test on an empty drive tells you about a condition you will occupy for one week of its life.

Formatting is worth a moment too. Portable drives ship formatted for maximum cross-platform compatibility, which means a filesystem with no journaling and a large allocation unit. That combination handles single big files well and thousands of small ones poorly, and it recovers badly from an unplanned disconnection. If the drive will spend its life with one operating system, reformatting to that system's native filesystem measurably improves small-file work and reduces the chance of a corrupted directory after a cable is yanked. If it genuinely travels between machines, keep the compatible format and eject properly every time.

08

When a cheap drive is genuinely the right answer

Cheap is fine more often than enthusiasts admit. If the drive exists to hold a backup, shuttle files between two machines, or store a media library you read from occasionally, the inexpensive option runs at the same speed as the expensive one for the first several gigabytes and finishes a large copy a few minutes later. Nobody notices a few minutes on an unattended weekly backup. Paying three times as much to shorten it is a poor trade, and the difference buys a second drive holding a second copy, which protects you from something that actually happens.

What separates good cheap from bad cheap is the assembly rather than the flash. External drives overwhelmingly fail at the connector or the bridge chip rather than in the memory, so a port anchored to the board with solder and a bracket instead of solder alone is worth more than any headline figure, as is a warranty process that replaces the unit without an argument. Rubberised armour and dust ratings protect the housing, not the data inside it. Whatever you buy, keep a second copy elsewhere. A single portable drive living in a bag is the least safe place your files can be.