The transfer starts beautifully. Five gigabytes a second, the progress bar moving faster than you can read the file names, exactly the number on the box. Then, somewhere in the middle, it falls off a cliff — to 400 MB/s, sometimes lower — and stays there for the rest of the copy.
The drive has not throttled and nothing has overheated. You have simply run out of the fast way to write, and the drive has fallen back to the way its memory actually works.
Why writing is the slow direction
A flash cell stores charge, and how much charge it holds is how the drive knows what bit it contains. With one bit per cell there are two states to tell apart, and the gap between them is enormous. With three bits — the TLC that almost every consumer drive uses — there are eight levels crammed into the same range. With four bits, QLC, there are sixteen.
Reading is quick regardless: sweep a comparison across the cell and see where it lands. Writing is the hard direction, because the controller cannot simply set a level. It nudges the charge up in small pulses and checks after each one, repeating until the cell sits inside its target band without overshooting — because there is no way to nudge back down without erasing the whole block.
The narrower the bands, the more pulse-and-verify cycles it takes. That is the entire reason a modern drive is slow to write and fast to read, and it is a property of the memory, not of the interface it is plugged into.
SLC mode is the same silicon used deliberately badly
Here is the trick that makes the first part of your copy fast. The controller takes ordinary TLC cells and writes just one bit into each of them — two levels instead of eight, with a huge margin between them.
Suddenly the cell needs very few pulses to land safely inside a band, and it can be programmed roughly an order of magnitude faster. Same chips, same drive, no special hardware: the memory is simply being used at a third of its capacity in exchange for speed.
That is the cache. It is not a separate fast chip bolted on — it is a region of the same flash, running in a fast, wasteful mode.
The cache is built out of your free space
Which sets up the fact that governs everything you will ever experience about this.
Because SLC mode uses one bit where TLC uses three, three blocks' worth of capacity in SLC mode hold one block's worth of data. So the arithmetic is simple and unforgiving:
Every 3 GB of free space buys about 1 GB of cache.
The controller borrows empty blocks, runs them fast, and hands them back later. A drive with plenty of room has an enormous cache. A drive that is nearly full has almost none — there is nothing left to borrow. Measured behaviour on real drives follows exactly this shape: a 2 TB drive that offers around 200 GB of cache when empty can be down to 10–15 GB by the time it is 85 per cent full, and drives commonly lose half their cache by 80 per cent.
This is why the same drive, doing the same copy, is fast in January and slow in November. Nothing degraded. You filled it.
The cliff, with the numbers worked through
Take a 1 TB TLC drive that is 80 per cent full, and copy a 120 GB game onto it.
- Free space: 200 GB → cache of roughly 66 GB.
- First 66 GB at 5 GB/s — 13 seconds.
- Remaining 54 GB at 400 MB/s — 135 seconds.
- Total: about 2 minutes 28 seconds.
Now the same drive and the same file, with the drive nearly empty. The whole 120 GB lands in cache and the copy takes about 24 seconds.
Same hardware, same file, same cable — six times the wall-clock, decided entirely by how full the drive was. No specification on the box distinguishes these two cases, because the box quotes the first one.
Why the review's figure is a ceiling you will not see again
Reviewers test on empty drives, and they are right to — it is the only condition that is reproducible between products. But it means the published sequential write number describes the drive on the day it arrived, and the interesting reviews are the ones that keep writing until the cliff appears and publish the whole curve: the height of the plateau, the length of it, and the floor underneath.
Read those three numbers rather than the headline. The plateau is marketing. The floor is what your drive does for the second half of every large copy, and across TLC drives it typically lands somewhere between 400 and 1500 MB/s depending on the controller and the flash. On QLC drives it can fall to around 100 MB/s — slower, for that stretch, than the mechanical drive you replaced.
The part that happens after the copy finishes
The data written in SLC mode is sitting in three times the space it needs. Before the drive can offer that capacity again, the controller has to read it back and rewrite it as TLC — compacting three blocks into one — and it does this quietly in the background while the drive looks idle.
Two consequences follow, and both are things people notice without explaining.
The second big copy is often slower than the first. If you start it before the folding has finished, the drive is writing your new data and rewriting the old at the same time, with less free space to build a cache from than it had an hour ago.
And every byte you wrote has now been written twice. That is write amplification, and it is spent out of the drive's endurance budget. For ordinary use this is irrelevant — consumer endurance ratings are far beyond what a desktop writes in its life — but it is the honest cost of the trick, and it is why keeping a drive stuffed to 95 per cent is bad for it as well as slow.
Why the drive is fast again tomorrow
Leave the machine alone for an hour and the copy that crawled yesterday runs at full speed today. Two mechanisms are behind that, and both can fail in ways worth knowing.
The first is the folding described above finishing its work, which hands the borrowed blocks back and lets the controller rebuild a cache.
The second is TRIM. When you delete a file, the operating system marks it gone in its own records — the drive is not told anything by that act alone. Until a TRIM command informs the controller that those blocks no longer hold anything wanted, the drive must assume the data is live, keep it during garbage collection, and refuse to count it as free space for cache purposes.
On a normal internal drive with a modern operating system this happens automatically and you never think about it. It is worth knowing where it does not: some external enclosures do not pass the command through, and some storage layers in between swallow it. A drive in that situation behaves exactly like a full drive no matter what the file manager says — the cache never comes back, and every large copy runs on the floor.
Not every cache is dynamic
Two design choices decide how badly the cliff bites, and neither is on the front of the box.
Some drives set aside a fixed region as SLC in addition to, or instead of, borrowing free blocks. A static cache is smaller than a dynamic one on an empty drive — and it is still there when the drive is 90 per cent full, which is the case you actually live in. A drive with a modest static cache can be the better buy over one with a huge dynamic cache that evaporates.
The other is whether the drive has its own DRAM. The controller keeps a map from logical addresses to physical blocks, and that map has to be consulted for every access. A drive with DRAM keeps it to hand; a DRAM-less drive borrows a little system memory over the interface instead. Under light use the difference is hard to see. Under exactly the conditions of this article — a long write, with folding happening underneath, on a fairly full drive — the map is being consulted and updated constantly, and the borrowed-memory design is doing that over a bus rather than on-board.
If you want to know which you are dealing with, the transfer dialog will tell you. Copy something large and watch the shape: the height of the plateau, how long it lasts before the fall, and where it settles. Those three numbers are your drive’s real specification, measured on your drive at its real fill level, and it takes one copy to find them.
Where you will meet this and where you will not
It would be wrong to leave you thinking your drive is slow. For almost everything a desktop does, the cliff does not exist.
- Gaming does not touch it. Playing is reading, and reads are unaffected. Loading a level never fills a cache.
- Ordinary desktop use does not touch it. Documents, browsers, updates — the writes are small and the cache is refilled between them.
- Installing and updating games does. A large title lands as tens of gigabytes of writes in one run.
- Video ingest, backups and library moves do, and this is where a slow floor genuinely costs you time.
So the shape of the advice is not "buy a better drive". It is: know which of those two groups you are in, because the specification that matters is different for each.
What to do
- Keep 15–20 per cent free. This is the single most effective thing on the list. It is not superstition about wear levelling — it is the cache, and it is the difference between the 24-second copy and the two-and-a-half-minute one.
- Buy the larger drive rather than the faster one if you write a lot. More capacity is more free space is more cache, and it moves the floor as well as the plateau.
- Read the sustained-write curve, not the headline. Plateau height, plateau length, floor. The floor is your drive's real write speed.
- Give it a moment between large copies if you can. The folding wants idle time, and it will take it out of your next transfer otherwise.
Two things we would not buy
A QLC drive as the working drive for anything write-heavy. It is excellent value per gigabyte and perfectly good as a game library that is mostly read — but its post-cache floor is in hard-drive territory, and a large ingest on a full QLC drive is a genuinely unpleasant experience.
And a drive chosen on its peak sequential number when the plan is to keep it 90 per cent full. At that fill level the peak is a number the drive can sustain for a few seconds. You will have bought the plateau and be living on the floor.
How this was put together
The cache-size behaviour comes from whole-drive fill testing — the methodology AnandTech established and TechPowerUp and Tom's Hardware continue, where the drive is written end to end and the speed is plotted the whole way, so the plateau, the cliff and the floor are all visible rather than averaged into one figure. The post-cache floors quoted here are the ranges those reviews find across current TLC and QLC controllers, and the shrinking-cache numbers are the measured fill-level behaviour published for current drives.
The derived figures are ours: the three-to-one rule that turns free space into cache, the 66 GB of cache on a 1 TB drive at 80 per cent full, and the six-fold difference between the same 120 GB copy onto that drive when it is nearly full and when it is nearly empty.
