Cloud & infrastructure

Native NVMe in Windows Server 2028: Check Drivers and Measure Gain

Windows Server 2028 preview turns native NVMe on by default, but vendor drivers may see no change. Check which driver each drive uses and benchmark before and after with DiskSpd.

By EpiFive Editorial TeamReviewed/updated 13 min readHow we check our guides
On this page

Windows Server 2028 turns on native NVMe by default in its public preview, so the first question is whether your drives will benefit. They benefit only if they use the in-box Windows NVMe driver rather than a vendor driver. Check the driver in Device Manager, then run the same DiskSpd test before and after and compare IOPS, latency and CPU together.

Who this is for

This guide is for server admins and storage owners who run NVMe SSDs in Windows Server and want to plan for Windows Server 2028 or for the Windows Server 2025 opt-in. By the end you will know what Microsoft has said about native NVMe, how to find out which driver a drive uses, how to run a repeatable benchmark, and how to read the results without overstating them.

Windows Server 2028 is a public preview. Microsoft's announcement says the security, performance and management features are available in preview through the Windows Server Insider Program. No general availability date or pricing was published on the pages we read, so none is given here. Do not use the preview on production servers. For the wider picture see Windows Server 2028 preview: what is new, and for a test VM see install the Windows Server 2028 preview in a Hyper-V lab.

What Microsoft has said about native NVMe

Two Microsoft sources matter, and they say different things.

The Windows Server 2028 announcement on the Windows Server blog says: "Windows Server 2028 adds native NVMe storage enabled by default and NVMe over TCP and RDMA transports", along with the ability to boot from ReFS. That post does not mention StorNVMe.sys, the in-box driver, vendor driver caveats or a build number.

The detail comes from the earlier post "Announcing Native NVMe in Windows Server 2025" (Tech Community, December 15, 2025, updated January 5, 2026). It says native NVMe in Windows Server 2025 is generally available but disabled by default, names the in-box driver as StorNVMe.sys, and states that "some NVMe device vendors provide their own drivers, so unless using the in-box Windows NVMe driver, you will not notice any differences."

Put together, the careful reading is this: Microsoft says native NVMe is on by default in the 2028 preview, and Microsoft's 2025 guidance says the gain depends on using the in-box driver. Whether the 2028 preview treats vendor drivers the same way is not stated on the 2028 post, so test it rather than assume.

What changed in the stack

A StorageReview analysis of the Windows Server 2025 feature describes the old behavior: Windows translated storage commands into SCSI regardless of the hardware's protocol, which added a translation layer. The new stack removes that layer so commands go more directly to NVMe hardware. StorageReview names NVMeDisk.sys as replacing Disk.sys in the new architecture and says Microsoft plans to enable native NVMe by default in the next Windows Server version. Treat those driver names as StorageReview's reporting, not a Microsoft statement for 2028.

Prerequisites

You need:

  • A lab server or VM running the Windows Server 2028 preview, or a non-critical Windows Server 2025 machine with the October 2025 cumulative update or later if you want to test the 2025 opt-in.
  • At least one NVMe drive you can test on. A virtual disk in a VM does not tell you how physical NVMe behaves, so use bare metal for real results.
  • Administrator rights and a free drive that holds only test data.
  • Microsoft's DiskSpd tool, downloaded from its GitHub project.
  • A written note of the drive model, firmware, driver version and BIOS settings before you change anything.

Back up anything on the machine first. StorageReview, which tested the Windows Server 2025 feature, advises trying it on a non-critical server first and reports user complaints about NVMe drives when deduplication is enabled, with an official fix said to be coming. Commenters on Microsoft's post also report problems with some backup agents; Microsoft asked for details, and those reports are unverified.

Step 1: Record the starting state

Capture what Windows reports for the physical disks. This is a documented cmdlet; the example output on Microsoft's page shows columns for FriendlyName, CanPool, OperationalStatus, HealthStatus, Usage and Size.

Get-PhysicalDisk

For the full property set of each disk, pipe the output to Format-List * and save it to a file so you can compare it afterward.

Get-PhysicalDisk | Format-List * | Out-File C:\Temp\physicaldisk-before.txt

Step 2: Check which driver each NVMe drive uses

Use Device Manager, which is the verification method Microsoft's post describes.

  1. Open Device Manager (right-click Start, then Device Manager).
  2. Expand Disk drives and Storage controllers and find your NVMe device and its controller.
  3. Open the device's Properties, select the Driver tab and choose Driver Details.
  4. Note the driver file names and the provider.

If the driver file is StorNVMe.sys, the device is using the in-box Windows NVMe driver that Microsoft's post names. If the provider and file name belong to the drive or controller vendor, the device is using a vendor driver, and Microsoft's 2025 post says you should not expect a difference from native NVMe.

After the native path is active, Microsoft says to confirm that all attached NVMe devices appear under the Storage disks section of Device Manager.

For a script-friendly record, Get-PnpDeviceProperty returns device properties as key and value pairs, and you can pass key names with -KeyName. Which key names exist for your storage devices is something to confirm on your build, so start with the unfiltered output.

Get-PnpDevice -FriendlyName '*NVMe*' | Get-PnpDeviceProperty

Do not change drivers during the test. If you want to compare a vendor driver against the in-box one, treat that as a separate experiment with its own before and after runs.

Decision table

What you findWhat it meansWhat to do
Driver file StorNVMe.sysIn-box Windows NVMe driverExpect native NVMe to be relevant; benchmark
Vendor-supplied driverMicrosoft says no difference unless the in-box driver is usedBenchmark anyway, but set expectations at zero
Drives behind a RAID card or HBAController driver sits in the pathCheck with the vendor; StorageReview says Microsoft was working with vendors on RAID and HBA drivers
Virtual disk in a VMNot physical NVMeUse for functional checks only

Step 3: Build a repeatable DiskSpd test

Microsoft's own test used this command, which we reproduce from the 2025 post:

diskspd.exe -b4k -r -Su -t8 -L -o32 -W10 -d30 testfile1.dat > output.dat

According to the DiskSpd wiki's parameter page, these switches mean:

SwitchMeaning
-b4k4 KiB I/O size (default is 64K)
-rRandom I/O, aligned to the block size by default
-SuDisable software caching (same as FILE_FLAG_NO_BUFFERING)
-t88 threads per target
-LMeasure latency statistics
-o3232 outstanding I/O requests per thread per target
-W1010 seconds warm-up before measurement
-d3030 second test duration

With no -w switch the workload is 100% reads, which is what this command measures. The wiki warns that write tests destroy existing data on the target without warning, so point them only at a test file or disk you can lose. The command above does not specify -c, so testfile1.dat must already exist; add -c with a size such as -c20G to create it, and keep the file larger than any drive cache you want to bypass.

To compare fairly, change only one thing between runs:

  1. Run the command three to five times and keep every result.
  2. Use the same file size, location and thread settings each time.
  3. Let the server idle for a few minutes between runs so thermal throttling does not skew later runs.
  4. Run the write variants (-w100 or a mix such as -w30) separately, on a disposable file.

Microsoft's own run used an Intel dual-socket server with 208 logical processors and 128 GB RAM and a Solidigm SB5PH27X038T 3.5 TB NVMe device, and reported "up to ~80% more IOPS" and "a ~45% savings in CPU cycles per I/O" on 4K random reads on NTFS volumes compared with Windows Server 2022, adding that results may vary. Treat those as best-case figures from one configuration.

Step 4: Change one variable and repeat

For Windows Server 2028 preview, native NVMe is on by default, so the comparison is between a baseline you already have and the new build. If you can, run your tests on the same hardware under Windows Server 2025 first, then install the preview on the same machine (or boot it from a separate disk) and repeat with identical parameters.

For the Windows Server 2025 opt-in, Microsoft's post gives a registry method. Run it in an elevated prompt on a non-critical server, after the October 2025 cumulative update or later:

reg add "HKLM\SYSTEM\CurrentControlSet\Policies\Microsoft\FeatureManagement\Overrides" /v 1176759950 /t REG_DWORD /d 1 /f

The post also describes a Group Policy method that uses an MSI it links to, with the policy under "KB5066835 251014_21251 Feature Preview" in the Windows 11, version 24H2, 25H2 section of the Local Group Policy Editor. This applies to Windows Server 2025 only; do not apply it to other versions. Check the post for the restart guidance before you change production-adjacent systems.

Step 5: Read the results

Compare three numbers per run: IOPS, average latency and CPU use. Microsoft suggests Performance Monitor with the Physical Disk, Disk Transfers/sec counter on the instance for your drive if you want a live view, and Windows Admin Center is also mentioned.

What an independent test showed

StorageReview tested Windows Server 2025 on a dual-socket AMD server with two 128-core EPYC 9754 CPUs, 768 GB DDR5 and fifteen 30.72 TB Solidigm P5316 NVMe SSDs in JBOD, using FIO rather than DiskSpd. Selected results, non-native to native:

WorkloadIOPSAvg latency (ms)CPU (%)
Random 4K read1,598,959 to 2,636,5160.169 to 0.10472.67 to 74.22
Random 64K read1,217,176 to 1,493,6370.239 to 0.20768.44 to 65.11
Sequential 128K read710,978 to 758,2520.613 to 0.60861.56 to 49.56
Random 4K write472,725 to 460,3830.992 to 1.02826.00 to 20.67

Redone from those figures: random 4K read IOPS rose 64.9% and latency fell 38.5%; random 64K read IOPS rose 22.7%; sequential 128K read IOPS rose 6.6% while CPU fell 19.5%; random 4K write IOPS fell 2.6% and latency rose 3.6%. StorageReview itself says most write results were within run-to-run variance, that this was one platform and one drive model, and that an earlier unrelated update may have narrowed the gap. It also notes that on a fast CPU throughput may stay flat while CPU use falls.

The lesson is to expect read-heavy, small-block workloads to benefit most, and to judge your own server with your own workload rather than quoting anyone's headline number.

Troubleshooting

No difference after enabling native NVMe

This is the expected result when the drive uses a vendor driver. Re-check the driver in Device Manager. If it is the in-box StorNVMe.sys and there is still no change, your test may be limited by the drive, the file size, thread count or a different bottleneck; try more threads and outstanding I/O.

Results vary wildly between runs

Add warm-up with -W, repeat each run, and watch for thermal throttling. The DiskSpd -D switch sets the IOPS sampling interval (default 1000 ms), and the text output reports the IOPS standard deviation, which helps you judge noise.

Backup or deduplication issues

Reports of problems with deduplication and with some backup agents exist for the Windows Server 2025 feature. Test your backup and dedup jobs on the non-critical server before any wider rollout, and send feedback to the address Microsoft gives in its post ([email protected]).

"Access denied" or a missing test file

Run the prompt as administrator and make sure the target path exists. Add -c to create the file, or DiskSpd cannot open it.

Drives behind RAID or HBA controllers

StorageReview says Microsoft was working with vendors on new drivers for devices such as RAID cards and HBAs. Ask the controller vendor whether a native-NVMe-aware driver exists for your model.

Checklist

  • Treat Windows Server 2028 as preview; lab and test servers only.
  • Record the drive model, firmware, driver version and BIOS settings.
  • Confirm in Device Manager whether each drive uses StorNVMe.sys or a vendor driver.
  • Save Get-PhysicalDisk output before and after.
  • Run the identical DiskSpd command three to five times per configuration.
  • Compare IOPS, latency and CPU together, for reads and for writes.
  • Test backup and deduplication jobs before trusting the change.
  • Re-check Microsoft's release notes whenever you move to a newer preview build.

See also

What we checked

  • Opened the Windows Server 2028 announcement on the Windows Server blog on 2026-11-22: it states native NVMe storage enabled by default and NVMe over TCP and RDMA, and does not mention StorNVMe.sys, vendor drivers or a build number.
  • Opened Microsoft's "Announcing Native NVMe in Windows Server 2025" post on 2026-11-22 and quoted its in-box driver name, vendor-driver caveat, registry key path, value name 1176759950, REG_DWORD value 1, and the DiskSpd command and hardware it lists.
  • Opened the DiskSpd wiki's "Command line and parameters" page on 2026-11-22 and matched each switch in the table to its description there, including the 100% read default and the warning that writes destroy existing data.
  • Opened the Microsoft Learn pages for Get-PhysicalDisk and Get-PnpDeviceProperty on 2026-11-22; Get-PhysicalDisk's documented example output shows the columns FriendlyName, CanPool, OperationalStatus, HealthStatus, Usage and Size.
  • Cross-checked the 2025 caveat that vendor drivers see no gain against StorageReview's report, which says the author tested only one drive model and that Microsoft plans to enable native NVMe by default in the next Windows Server version.
  • Recalculated percentage changes from StorageReview's table by hand: (2,636,516 / 1,598,959) - 1 = +64.9%; (0.104 / 0.169) - 1 = -38.5%; (1,493,637 / 1,217,176) - 1 = +22.7%; (758,252 / 710,978) - 1 = +6.6%; (49.56 / 61.56) - 1 = -19.5%; (460,383 / 472,725) - 1 = -2.6%; (1.028 / 0.992) - 1 = +3.6%.
  • Not checked: no NVMe hardware, preview build or DiskSpd run was available, so no benchmark was run here and the registry change was not applied. The Device Manager steps follow the verification Microsoft's post describes, and menu wording may differ on your build. Which PnP property keys expose the storage driver was not confirmed.

Last tested: 2026-11-22

References

Questions people ask

Is native NVMe enabled by default in Windows Server 2028?

Microsoft's Windows Server 2028 preview announcement says the release adds native NVMe storage enabled by default, along with NVMe over TCP and RDMA transports. It is a public preview through the Windows Server Insider Program with no announced general availability date, so use it for lab testing only.

Will my NVMe drive get faster on Windows Server 2028?

Only if the drive uses the in-box Windows NVMe driver. Microsoft's Windows Server 2025 native NVMe post says that unless you use the in-box driver you will not notice any differences, because some vendors ship their own drivers. Results also vary by hardware and workload.

How do I check which driver an NVMe drive uses?

Open Device Manager, find the NVMe device and check its driver details, and compare against the in-box StorNVMe.sys that Microsoft names. Microsoft's post also says attached NVMe devices appear under Storage disks once the native path is active.

How should I benchmark native NVMe before and after?

Run the same DiskSpd command, on the same test file and the same hardware, before and after the change, and repeat each run several times. Compare IOPS, average latency and CPU use together, because CPU can fall even when throughput stays flat.

Windows Server 2028NVMeStorNVMe.sysDiskSpdWindows Server 2025