A Full SSD Costs You Writes, Not Frames

The advice to keep a fifth of your SSD empty has been repeated for so long that it has stopped being a claim and become furniture. Four drives filled to within 10 GB of capacity and run through games and creative apps say the framerate half of it is simply wrong — no meaningful loss in average FPS, none in 1% lows. What a full drive does take from you is write throughput, spare blocks and endurance budget, and that bill never shows up in a frame counter.

Framerate Is Not What a Full Drive Takes From You

The lineup covered the shapes people actually buy: a DRAM-less WD Blue SN580 1 TB leaning on host memory buffer, a Corsair MP600 in both 1 TB and 2 TB capacities with real DRAM, and a Gen5 Crucial T705 2 TB. Baseline was roughly 450 GB of game installs and assets — under half of the 1 TB drives, under a quarter of the 2 TB ones — and from there the fill was pushed to 60, 80, 90 and 95 percent with large video files plus a scatter of small office documents. Horizon Zero Dawn Remastered, F1 24 and Cyberpunk 2077 ran at every step.

Nothing moved. Across all four drives and all three titles, average frame rate and 1% lows sat flat from baseline to full-minus-10 GB. Whatever variance existed was run-to-run noise, with no trend against fill percentage. If a full drive were quietly costing frames, the effect would have to be smaller than a single benchmark run's spread — which is another way of saying it doesn't matter to anyone actually playing.

Load times were the one place with a real signal. The FFXIV Dawntrail scene-load benchmark had more spread than the frame data, and its shape was consistent: the 1 TB drives loaded almost 10% faster below half capacity than packed to the brim, and the T705 picked up a little extra latency past 90%. Visible in a benchmark, invisible in practice, and nowhere near enough to make an NVMe feel like a hard drive.

The Cliff Is in Writes, Not Frames

TLC stores three bits per cell and pays for that density with slower, fussier writes. To keep sequential writes quick, controllers carve a slice out of the empty NAND and run it as pseudo-SLC — one bit per cell, fast to program. The arithmetic is the whole trick and also the trap: a 1 TB drive with 150 GB set aside as pSLC holds 450 GB at that density, and the cache is built out of space you haven't used yet. Push past halfway and the controller converts pSLC back to TLC to keep the capacity honest, after which any write larger than what's left in the cache falls back to dense-cell speed.

That's the mechanism behind the advice, and it is not a framerate mechanism — games mostly read. A level load pulls textures and geometry off the drive and writes almost nothing, which is why the frame data is boring. Start writing and the picture changes. In the transfer test, a 41 GB video file copied onto each drive at each fill level, the SN580 1 TB averaged around 2,990 MB/s at 60% full and about 1,354 MB/s at 90%. The MP600 1 TB opened near 3,000 MB/s at baseline and fell to roughly 900 to 1,000 MB/s past 60%. Nobody should build a rule on the exact figures; the direction is unmistakable and it tracks the cache math.

Under the cache sits garbage collection. NAND writes in 4–16 KB pages but erases in 4–96 MB blocks, so deleting data strands valid pages among invalid ones, and the controller copies survivors elsewhere before the block can be reclaimed. A drive with almost no spare blocks does more copying to free less space. That's write amplification — one instruction from the operating system becoming several physical writes. A full drive doesn't slow your GPU. It makes storage work harder for every megabyte you hand it.

The Real Cost Is the Endurance Budget

NAND wears out. Each drive carries a terabytes-written rating — 600 TBW on the SN580 1 TB, 1,800 and 3,600 on the MP600 pair, 1,200 on the T705 — and the controller spreads writes evenly so no block dies early. Wear levelling needs room to shuffle. Fill the drive and that freedom shrinks, amplification climbs, and rated lifetime drains faster than the calendar suggests. This never shows up as a bad frame time; it shows up years later as a drive due for replacement, with a SMART wear figure almost nobody opens as the only warning.

The quieter problem is everything the machine writes when you aren't asking: page file, search index, logs, update staging, restore points. Squeeze free space toward zero and those routine writes land on a drive with no cushion, which is where stutter and strange behaviour come from. Creative work is worse because it writes constantly — autosaves, scratch files, recovery states. At full, Photoshop and its neighbours threw low-space warnings; past that, programs refused to launch outright. Ten gigabytes of headroom kept everything running.

What I Think

The 20% rule survives because it's directionally right and completely mis-explained. People repeat it as a performance tip, which makes it sound like superstition when their frames don't change, and then they ignore it and run the drive down to two gigabytes free. The honest version is that free space is a write buffer and an endurance reserve, and how much you need depends on what you ask the drive to do. A machine that installs games onto a 2 TB library and writes almost nothing else can sit at 95% full indefinitely. A workstation cutting ProRes proxies onto the same drive at 90% full is going to have a bad time, and it will blame the SSD.

Over 90% with a workload full of large writes is a real reason to add capacity, and the reason isn't FPS. Dollars per gigabyte have climbed hard and Warzone alone wants 116 GB, so "buy a bigger drive" isn't the answer it was two years ago. The better move for most people is a second NVMe as a scratch and library drive, leaving the boot drive its cushion and keeping installs off it. That solves the actual problem — writes, spare blocks, endurance — without replacing hardware that was never the bottleneck.