This is a review methodology, not a certification program. It is still useful for anyone buying U.2, U.3, E1.S or E3.S drives for a workstation, home lab or server. It also shows how to read any vendor spec sheet or review of these drives.
What was announced
Tom's Hardware says it is returning to enterprise SSD testing after a hiatus of several years. It cites demand driven by AI workloads and the spread of enterprise form factors into high-end workstations and home labs.
A full run on a 7.68TB PCIe 5.0 drive takes most of a day and writes the drive end-to-end many times. The suite is automated and crash-resumable from the last stage. A shorter express mode handles quick triage, and variant runs repeat a test or change one setting to track down odd results before the site goes to the vendor.
The site also says what it will not do. It excludes dual-port failover, power-loss validation and SR-IOV, which it says belong in a qualification lab.
The test bench
- CPU: an Intel Xeon 658X, a 24-core, 48-thread Granite Rapids part with a 4.9GHz boost. The testers say clock speed affects how many IOPS a core can push. The core count lets them spread load across eight fio jobs.
- Motherboard: an Asus Pro WS W890E-Sage SE. Its MCIO x8 connector splits into x4/x4 for two U.2 or U.3 drives. It also has seven PCIe 5.0 slots.
- Memory: a 64GB G.Skill DDR5 R-DIMM kit running at the DDR5-5600 JEDEC speed. The authors say memory speed does not affect enterprise SSD results.
- Boot and power: a Corsair MP700 Pro XT 2TB as the OS drive and a Corsair HX1500i power supply.
- Drive carriers: Icy Dock adapters for E1.S drives, a JMT adapter with an industrial Noctua fan for E3.S, and an open-air E-ATX bench to limit heat build-up as a variable.
- Stated gap: no RTX Pro graphics card yet. The AI tests are therefore storage I/O simulations, not full GPU workloads.
Several vendors supported the build, and the article names Asus, Corsair, G.Skill, Icy Dock and Noctua. Readers should weigh that, though the supplied parts are described as test equipment, not endorsements.
Software choices
The suite runs on a custom Debian 13 image with a pinned kernel, Linux 7.2.6 from backports. According to the article, Debian's stock 6.12 kernel predates support for the board's Intel E610 10GbE controller. The testers prefer Debian to Ubuntu because it has fewer background services, no Snap and no unattended upgrades by default.
They remove AppArmor and disable scheduled jobs such as the weekly TRIM timer and the drive-health poller. They also turn off NVMe autonomous power-state transitions (APST) and confirm the drive reports it off.
They chose fio over Oracle's Vdbench for one set of latency accounting across all charts. They also note that Vdbench has had no release since 2018.
This is a Linux bench, so the numbers describe the drive's raw block-device behavior. They do not describe how Windows, NTFS or a given driver stack would perform. That is the right call for comparing drives. Windows admins should treat the results as a ceiling and a ranking, not a prediction for their own filesystem.
Steady state and preconditioning
Each run starts with a secure erase through nvme format. If the drive refuses, the suite falls back to a plain format, then a full-drive discard, and records which was used.
Preconditioning uses SPRandom. Independent sources confirm what it is. SanDisk describes it as a method for rapidly preconditioning large SSDs using knowledge of physical over-provisioning. SanDisk says it writes every logical address, distributes over-provisioning across the drive, and reaches steady-state performance in a single physical write. Sandisk contributed it to the fio project as an open source extension.
The vendor claims big savings. TechRadar reports that conventional preconditioning of a 128TB drive takes over 160 hours, against 6.5 hours with SPRandom. Those figures come from Sandisk and apply to very large drives. They are not Tom's Hardware measurements on a 7.68TB drive.
NVMe drives do not report their raw flash. For its first 7.68TB drive, Tom's Hardware assumes 8TiB of NAND, which works out to about 14.5% over-provisioning. This is an assumption, and the tester can override it.
The site does not trust SPRandom alone. A steady-state gate runs 4K random writes at QD256 and borrows SNIA's two steady-state tests with limits ten times tighter, applied to the last 30 minutes:
- The window is split into five six-minute rounds.
- The round averages must sit within 2% of each other.
- The trend line across the window must not move more than 1%.
A drive that has not settled after three hours is marked as not reaching steady state, but the run continues. A later repeat of the four-corner 4K write flags the run if the two results differ by more than 3%. The authors say an earlier, looser gate passed their first drive while its write speed was still climbing.
The suite also re-runs a test immediately if its slowest moment falls below 70% of its average. Both results are kept, and retests are noted in the review.
After the random tests the drive is erased again and filled once with 128K sequential writes at QD64. On the first test drive, sequential write results on a randomly written drive varied more than threefold. After a sequential fill they held flat from QD2 through QD1024.
Drives are formatted to 512-byte logical blocks when supported. Drives without that format are tested as shipped, and their sub-4K results are marked unsupported.
Keeping the bench out of the results
The authors found three ways the host could distort results:
- CPU C-states: waking cores from deep idle can take up to 170 microseconds. Holding cores awake cut the first drive's average 4K QD1 read latency by about a sixth.
- Single-job saturation: one fio job can max out a core before a PCIe 5.0 drive is fully loaded. Random tests now spread queue depth across up to eight jobs.
- Repeated random seeds: By default, fio seeds its random generator the same way every run, so it replays the same sequence of random offsets. On the first drive this was the main reason 4K random writes came in at more than twice the sustained figure. Every random test now uses a fresh seed.
The main charts keep interrupt-driven completions and the default IOMMU mode, because that is how most applications use a drive. That sets a ceiling of roughly 2.8 million 4K random read IOPS on this bench. The fastest drives may therefore score below their spec sheets, and the authors say that is the bench's limit, not the drive's. A separate set of polled-I/O charts shows the drive's limit at both ends. It covers peak 4K random read at QD1024 across eight jobs, plus 4K random reads and writes at QD1.
What gets measured
| Test | What it shows |
|---|---|
| Four corners | 4K random and 128K/1MB sequential headline numbers on a preconditioned drive |
| Random and mixed sweeps | 4K and 8K from QD1 to QD4096, across eight read/write mixes |
| Sequential sweep | How much parallelism a drive needs for full speed |
| Block-size sweep | 512 bytes to 1MB at QD32, exposing read-modify-write penalties |
| Tail latency | P50 to P99.999 at QD1, QD4 and QD16 |
| Saturation curve | 4K 70/30 across 13 queue depths, to find where latency climbs |
| Sustained write | One hour of 4K random writes at QD256, to show flat or sawtooth behavior |
| Write-to-read recovery | QD1 reads for four minutes after a write burst |
| Idle recovery | Write bursts after idle gaps of 0 to 300 seconds |
| Simulated database I/O | Log writes with sync, 8K 70/30 at QD16, 256K scans |
| AI checkpoints | 105GiB over eight writers and 912GiB over 64, modeled on 8B- and 70B-class models |
| Checkpoint interference | A fixed 4K read load alone and during a checkpoint save |
| Thermals | Temperature during the hour-long write test against the drive's own thresholds |
A few of these deserve explanation:
- Tail latency matters because averages hide stalls. At a million IOPS, the slowest 0.001% still amounts to ten requests a second.
- Checkpoint interference times each read from when it was due, not when it was sent. Reads stuck behind a stalled one therefore count against the drive.
- The saturation curve gives a practical operating ceiling. On the first drive, that point was set by write latency.
What this means for buyers
- Look for steady-state figures, not fresh-out-of-box numbers. A drive can look excellent for the first minutes and then fall well below that.
- Check tail latency and recovery behavior if the drive will share duty between serving reads and absorbing writes.
- Treat any review's top-end IOPS carefully. As this bench shows, the host can be the limit.
- Remember that results from different outlets use different hardware, preconditioning and profiles, so the numbers are not directly comparable. The authors make this point themselves.
The caveats are real. The suite is a single-host Linux evaluation, with no GPU in the loop yet. It also leaves out reliability, power-loss and multipath testing. The methodology is transparent, and the early findings, such as the seed problem and the C-state effect, show the authors checked their own bench as well as the drives. The first reviews will show how well it holds up.
References
- How we test enterprise SSDs Tom's Hardware · 2026-11-17T12:32:43+00:00
- SanDisk just cut SSD preparation times from ten days to only hours using an open source storage breakthrough techradar.com
- From Bottleneck to Breakthrough: Solving the SSD Preconditioning Challenge sandisk.com