SSD Over-Provisioning Explained: Why 3.84TB and Not 4TB
Buy a 3.84TB enterprise SSD and a 4TB one from the same range and you may find they contain identical flash. The difference is how much of it the drive keeps for itself.
That reserved capacity is over-provisioning, and it is the specification that explains why enterprise SSD capacities look odd, why two drives with the same flash perform differently under sustained load, and why leaving a drive nearly full is a performance decision rather than just a capacity one.
Why an SSD needs spare capacity at all
The reason comes down to a mismatch in how flash is written and erased.
Flash is written in pages, typically a few kilobytes, but it can only be erased in much larger blocks made up of many pages. A drive cannot overwrite a page in place. To change data, it writes the new version to a fresh page and marks the old one invalid.
Over time the drive accumulates blocks full of a mixture of valid and invalid pages. To reclaim that space it has to copy the still-valid pages elsewhere and erase the whole block. That process is garbage collection, and it needs somewhere to put the valid pages while it works.
Over-provisioning is that somewhere. It is flash the drive owns but never reports to the host, held back specifically so the controller always has free blocks to work with.
How to spot it in the capacity number
The pattern is visible on any enterprise SSD range once you know to look for it.
Flash is manufactured in binary quantities. A drive built from 4TiB of raw flash has roughly 4.39TB in decimal terms. What the vendor reports as usable is where the over-provisioning shows.
Sold as 3.84TB, that is around 7 percent held back — the standard level for read-intensive enterprise drives. Sold as 3.2TB, it is around 28 percent held back, which is typical of mixed-use. Sold as 1.6TB from the same raw capacity, the reserve is very large and the drive is a write-intensive part.
This is why the capacity ladder in an enterprise range runs 1.6, 3.2, 3.84, 7.68 rather than in round numbers. The odd figures are the endurance and performance tiers, expressed as capacity.
It also explains a price comparison that otherwise makes no sense. A 3.2TB drive costing more than a 3.84TB drive in the same family is not mispriced. It contains the same flash with more of it reserved, and it is sold for heavier write workloads.
What the reserve actually buys
Three things, and they matter to different buyers.
Sustained write performance. A drive with generous free blocks can absorb writes without stopping to reclaim space. A drive running short has to garbage collect in the write path, and throughput falls sharply once that starts. This is the difference between benchmark figures and behaviour an hour into a real workload.
Endurance. More spare area means the controller copies fewer valid pages per block erased, which lowers write amplification — the ratio between what the host writes and what the flash actually writes. Lower write amplification means the endurance budget covered in our DWPD and TBW guide lasts longer for the same host workload.
Bad block replacement. Flash blocks fail over the drive's life. The reserve is where replacements come from, which is part of why a drive can report wear steadily rather than failing abruptly.
The part most buyers can act on
Free space on the filesystem behaves as additional over-provisioning, and this is the lever available without buying a different drive.
When the drive knows a region is unused, it can treat those blocks as spare. A drive running at 60 percent full has considerably more working room than one at 95 percent, and behaves accordingly under sustained writes.
The mechanism that tells the drive a region is unused is TRIM on SATA, or UNMAP on SAS and NVMe. Without it, deleting a file frees space in the filesystem but the drive still believes those blocks hold valid data and keeps relocating them during garbage collection.
Two situations where this quietly fails.
Behind a RAID controller. Support for passing TRIM or UNMAP through to member drives varies by controller and firmware. Where it is not passed through, the drives gradually reach a state where every block looks used, and sustained write performance settles at the drive's worst case.
Inside thin-provisioned or virtualised storage. Space reclaimed at the guest level has to be passed down through each layer to reach the drive. Each layer that does not forward it breaks the chain.
If a drive that performed well when new has become slow under load, this is worth checking before concluding the drive is worn.
Setting over-provisioning yourself
You can increase the reserve on a drive you already own, and there are two ways.
The simple method is to leave capacity unpartitioned. Partition 80 percent of the drive, leave the rest untouched, and provided the drive has never had data written to that region the controller treats it as spare. On a drive that has been in use, secure erase it first, otherwise the controller still considers those blocks occupied.
The formal method is a host-protected area or namespace resize, supported by most enterprise drive toolkits. This tells the drive explicitly to report less capacity, which is more reliable than relying on unpartitioned space.
The trade is direct and worth stating plainly. You are exchanging capacity you paid for against write performance and endurance. For a read-heavy workload it is wasted money. For a sustained-write workload on a read-intensive drive it can rescue a configuration that would otherwise need replacing.
How much is enough
Driven by workload rather than by a general rule.
Read-heavy work — boot volumes, web content, reference data, analytics over data written once — is well served by the standard 7 percent. Buy usable capacity and do not pay for reserve you will not use.
Mixed workloads — virtualisation hosts, general databases, application servers — suit the 20 to 28 percent tier. This is where most enterprise buying lands.
Sustained random writes — transaction logs, caching tiers, high-frequency ingest — justify the write-intensive tier, where reserve can reach half the raw flash. Expensive per usable terabyte, and the correct purchase where the workload demands it.
The measurement that settles it is the same one in the endurance guide: actual write volume over a representative week, divided by the usable capacity of the drive you are considering.
Where this leaves consumer drives
Consumer SSDs typically hold back very little, which is how a 4TB consumer drive and a 3.84TB enterprise drive can be built from the same flash at very different prices.
That thin reserve is the mechanism behind the behaviour people report when consumer drives are put into servers: acceptable performance initially, then a collapse under sustained load as the drive fills and garbage collection moves into the write path. Combined with the absence of power loss protection, it is the substance of the consumer versus enterprise distinction.
Sourcing
We supply enterprise SSDs across read-intensive, mixed-use and write-intensive tiers in SAS, SATA and NVMe, with the exact manufacturer part number, endurance rating and condition stated on every line.
Tell us the workload and the write volume if you have it, and we will recommend a tier rather than quoting the largest usable capacity for the money. Send a bill of materials through the bulk quote page or email sarah.jane@techsellerusa.com.
Common questions
What is SSD over-provisioning?
Flash capacity the drive holds back and never reports to the host, reserved so the controller always has free blocks for garbage collection and bad block replacement. It is the main reason enterprise capacities read 3.84TB and 3.2TB rather than 4TB.
Why is a 3.2TB SSD more expensive than a 3.84TB one?
Because they often contain the same raw flash, with more of it reserved on the 3.2TB part. The larger reserve delivers higher sustained write performance and endurance, so it is sold as a mixed-use or write-intensive drive at a higher price per usable terabyte.
Can I add over-provisioning to a drive I already own?
Yes. Either leave a portion unpartitioned after a secure erase, or use the vendor toolkit to resize the reported capacity through a host-protected area or namespace change. The second method is more reliable. You trade usable capacity for write performance and endurance.
Why has my SSD got slower under load over time?
Often TRIM or UNMAP is not reaching the drive, so it still believes deleted blocks hold valid data and keeps relocating them. This is common behind RAID controllers that do not pass the command through, and in thin-provisioned or virtualised layers. Check that before assuming wear.
Does keeping free space on the filesystem help?
Yes, provided TRIM or UNMAP reaches the drive. Free space the drive knows about behaves as additional reserve, so a drive at 60 percent full has considerably more working room under sustained writes than one at 95 percent.
How much over-provisioning should I buy?
Read-heavy work is fine on the standard 7 percent tier. Virtualisation and general databases suit 20 to 28 percent. Sustained random writes justify the write-intensive tier. Measure write volume over a representative week and divide by usable capacity to settle it.
Send us the workload and we will size the reserve as well as the capacity.
