Skip to content
senn-techsenn-tech
Back to blog
Hardware
Local AI and GPU hardware2026-11-19· By Franz Senn

GeForce 621.14 and 621.24 write up to 30 GB an hour to the SSD

Two Nvidia Game Ready drivers keep writing to the system SSD while idle: 621.14 from 4 November 2026 and 621.24 from 12 November 2026. Users in the Nvidia forum measure 515 MB in five minutes on 621.14 and up to 30 GB an hour on 621.24, with no game running and no visible activity. The file behind it is nvlog.nvlgstg, under C:\ProgramData\NVIDIA Corporation\nvtopps\nct\. By the reading of Tom's Hardware and PCGH the release notes of 621.24 list no open defect for it, and Nvidia has not said anything since. We measured our own fleet; we did not reproduce the forum rates.

The nvtopps log defect is a write with no order behind it. The NVIDIA Display Container LS service keeps writing log data into nvlog.nvlgstg on the 621.14 and 621.24 branches, with no game, no recording and no overlay, and the rate grows with the size of the file. Those affected have switched it off by rolling back so far.

Nvidia logo
The writes happen inside the NVIDIA Display Container LS service, with no application asking for them. (Quelle: Simple Icons, file nvidia.svg (CC0))

What users report

The thread on 621.24 is titled "Continuous writing of gigabytes to SSD to nvtopps file". Its author counts about 30 GB in an hour and 70 GB or more in longer sessions, with telemetry switched off and with restarts of the service and the machine achieving nothing. A second user reports more than 3 GB in ten minutes. The best measurement comes from the 621.14 thread, dated 6 November: 515 MB from the process in five minutes of idle, and under 2 MB on the same machine with 621.01. Tom's Hardware measured 270 MB on the drive, about 3 GB an hour, with the SMART counter rising at the same pace. Wccftech measured 1.43 MB per second on its own RTX 4090 machine with capture and overlay disabled, close to 5.2 GB an hour.

Where the writes come from

The NVIDIA Display Container LS service (NVDisplay.Container.exe) writes into nvlog.nvlgstg, reported as rewriting the whole file on every update instead of appending to it. One user documented it in January 2025 with Process Monitor: a single write at offset 4,096 with a length of 10,272,768 bytes, over 10 MB at once, into a file that had grown to 15 MB across eight hours; that machine was a Dell XPS 15 9570 with a GeForce 1050 Ti on driver 566.36. Three DDU reinstalls changed nothing, per Tom's Hardware. The mechanism is older than 621.14, what is new is the volume: PCGH names 471.41 as the first driver with it, Tom's Hardware 471.96 as the build whose release notes carried the entry in 2021.

The squeeze: 621.01 or the writes

A rollback is not a free choice. Build 621.01 from 21 October, the subject of our October post, causes black screens and video stalls in Chrome and Edge according to user reports, and 621.14 fixed exactly those according to PCGH. Going back from 621.24 trades write volume for dropped frames.

Checking a workstation with a GeForce cardIdentify the versionnvidia-smi, 621.14 or621.24?Look at the writerateResource Monitor,NVDisplay.Container.exeRead SMART twicetwo readings thirty minutesapartJudgeabove 1 GB an hour idle,derived by usActroll back only into the621.01 stalls
The order we would follow on an affected workstation. The threshold is ours, the findings are from the forums. (Quelle: Nvidia forum thread on 621.24, check extended by Senn-IT)

What this costs the SSD

Endurance comes from the drive vendor's datasheet, so we calculate on one reference case: a consumer drive with 2 TB and 640 TBW, the Kingston NV3. That figure is in Kingston's own datasheet, read on 19 November 2026, derived from the JESD219A client workload, with the added note that the drive is meant for desktops and laptops and not for server environments. We calculate over 24 hours and 365 days, without write amplification.

Write rateper dayper yearat 640 TBW
0.63 · CAD guest, avg 35.5 h15 GB5.5 TB116 years
0.92 · CNC guest, avg 33.4 h22 GB8.1 TB79 years
1.95 · host .180.3, morning window47 GB17.1 TB37 years
2.69 · terminal server, avg 33.5 h65 GB23.6 TB27 years
4.44 · host .202, late morning window107 GB38.9 TB16 years
5.2 · Wccftech reproduction125 GB45.6 TB14 years
3.0 · Tom's Hardware, drive level72 GB26.3 TB24 years
6.2 · user report on 621.14149 GB54.3 TB12 years
30 · user report on 621.24720 GB262.8 TB2.4 years

Unit GB per hour, 1 GB as 1,000,000,000 bytes, our own values measured on 19 November 2026 by Senn-IT; the last four rows are third-party reports and columns 2 to 4 are our extrapolation.

Two limits remain. SMART counts what the operating system hands to the drive, not the bytes landing in the NAND cells, yet the SMART cross-check from Tom's Hardware shows the question was fair. And TBW is a vendor warranty statement. At 30 GB an hour that is 41 percent of the specified write volume per year, calculated from 262.8 TB against 640 TBW. Tom's Hardware calculates smaller: their 270 MB per five minutes are 28 TB a year on a 1 TB drive with 600 TBW, which is 5 percent of the warranty in continuous operation and 15 percent at eight hours a day. Our Micron 7400 reports 73.10 TB written at 0 percent wear used.

What our own fleet writes

Our measurement ran on 19 November 2026 in two windows, 04:55 to 05:13 UTC and 08:38 to 08:53 UTC. The Windows values come from the block statistics of the virtualisation layer and are averages across the uptime of each guest, working phases included. The Linux values are windows over /proc/diskstats and they vary a lot: .180.3 reached 1.95 GB an hour in the morning window and 0.04 in the late morning one, .202 went from 2.90 to 4.44. The windows show utilisation at that moment, idle is a different number.

Write volume per hour, our own measurementWindows CAD guest, RTX A20000.63 · 35.5 h uptimeWindows CNC guest, RTX A20000.92 · 33.4 h uptimeLinux host .180.3, 4x RTX 50901.95 · morning windowWindows terminal server2.69 · 33.5 h uptimeLinux host .202, 2x R97004.44 · late morning window05
Our own measurement on 19 November 2026. Windows guests via the Proxmox block statistics across their uptime, Linux hosts via /proc/diskstats across windows of 10 and 15 minutes, in both cases the whole system rather than one process. The host values are the higher of the two windows. (Quelle: Senn-IT, own measurement on 19 November 2026)

None of our values reaches the highest forum rates, and there is a reason we had not seen before: our Windows fleet has no GeForce card in it. The two running guests with a passed-through GPU each have an RTX A2000, a third guest with a Quadro P400 is switched off. Which driver our Windows guests carry stays in the asset list, because the guest agents are blocked from executing commands. And the Linux values are idle figures for nobody: the 4.44 GB an hour from .202 sit in the range of the Wccftech reproduction at 5.2, while .202 did compute work in that window and the forum machines sat idle. On .180.3 we found no .nvlgstg file anywhere in the root filesystem.

Check your own machine in ten minutes

  1. Run nvidia-smi in a command prompt: if it reports 621.14 or 621.24 on a GeForce card, step 2 is worth doing.
  2. Resource Monitor, the Disk tab, sorted by write rate: if NVDisplay.Container.exe writes several MB per second while the machine idles, you have the case from the forums.
  3. Get-Counter '\PhysicalDisk(0 C:)\Disk Write Bytes/sec' -SampleInterval 2 -MaxSamples 30 samples for 60 seconds, and the mean times 3,600 is your hour.
  4. Read SMART twice thirty minutes apart, the field is called Data Units Written or Total Host Writes. The difference across one night is the most solid number.

We derive a check threshold of 1 GB an hour at idle from that: our Windows guests sit at 0.63 to 0.92 across their whole uptime, and several GB an hour on an otherwise idle machine deserves a second look. The threshold is ours, derived.

What this means for operations

A graphics driver is a change to half the fleet in a small business, even when it formally only concerns a game. We would change three things: driver updates for workstations through a pilot group of two or three GeForce machines, write volume as a line in the monthly review, and the TBW figure in the purchasing decision. What a higher TBW class costs, we have not researched. The AMD comparison closed faster: there atieclxx.exe did the writing, and after roughly three weeks Adrenalin 26.11.1 shipped the fix. Nvidia has neither a statement nor a fix. As of 19 November 621.24 is the newest Game Ready driver, so anyone running a GeForce workstation measures for themselves.

The earlier post on 621.01 is GeForce 621.01: black screens and video stalls, more on graphics cards and drives in our posts on local AI and GPU hardware. What we offer around Windows and server environments is under services.

Further reading

  • Nvidia forum, Continuous writing of gigabytes to SSD to nvtopps file: first report on 621.24 with 30 GB in an hour, 70 GB in longer sessions, telemetry off, restarts of service and machine without effect, plus a second user with more than 3 GB in ten minutes. Retrieved 19 November 2026.
  • Nvidia forum, GeForce GRD 621.14 feedback thread: the 515 MB in five minutes of idle on 621.14 and the under 2 MB on 621.01 on the same machine, dated 6 November 2026, with a desktop comparison at 14 KB file size. Retrieved 19 November 2026.
  • Nvidia forum, Driver excessively writing to nvlog(nvlgstg): the January 2025 documentation with Process Explorer and Process Monitor, a write at offset 4,096 with length 10,272,768 bytes, a Dell XPS 15 9570 with GeForce 1050 Ti on driver 566.36, 14 KB on a desktop machine with GeForce 4060. Retrieved 19 November 2026.
  • Nvidia forum, GeForce GRD 621.24 feedback thread: the thread Nvidia pinned, with 185 replies on 19 November 2026 and the open question whether a fix is coming. Retrieved 19 November 2026.
  • Tom's Hardware, Nvidia GeForce driver bug reportedly hammers SSDs: SMART field and write volume at drive level, 270 MB in five minutes, three DDU reinstalls without effect, jumps on Alt+Tab, the calculation with 600 TBW, driver 471.96 as the 2021 fix. Retrieved 19 November 2026.
  • PCGH, Geforce: driver loads the SSD with up to 30 GiB an hour: the user measurements per driver version, 621.14 as the fix for the black screens, the worked example with the Kingston NV3 and its 640 TBW, 471.41 as the first driver showing the behaviour, and the note that the 30 GiB rest on a single report. Retrieved 19 November 2026.
  • Wccftech, Nvidia GPU drivers reportedly writing up to 30GB per hour to SSDs: their own reproduction on an RTX 4090 machine at 1.43 MB per second, the note that process IO is not the same as NAND writes, and the AMD parallel. Dated 18 November 2026, retrieved 19 November 2026.
  • Borncity, Windows: Nvidia driver writes wildly to SSDs: German summary of 19 November 2026 with 21 GB an hour, more than 1.2 TB in five days, a rollback to 621.01 at 23 MB per second, and the note that AMD closed its pattern with Adrenalin 26.11.1. Retrieved 19 November 2026.
  • Kingston, NV3 PCIe 4.0 NVMe SSD datasheet: vendor figure of 640 TBW for the 2 TB model, derived from the JESD219A client workload, five year warranty, explicitly not intended for server environments. Document dated 2025, retrieved 19 November 2026.
  • Own measurement: Proxmox block statistics of the Windows guests and /proc/diskstats of hosts .180.3 and .202, 19 November 2026, two windows from 04:55 to 05:13 UTC and from 08:38 to 08:53 UTC, plus the SMART reading of the Micron 7400 on .180.3. PCI configuration of the passed-through cards from the guest definitions on nodes pve1, pve2 and pve3.
Questions?
Which GeForce drivers are affected?+

The writes to the SSD are reported for the Game Ready builds 621.14 from 4 November 2026 and 621.24 from 12 November 2026. The build before them, 621.01 from 21 October, is reported free of this writing behaviour, and it carries the black screens we described on 26 October. PCGH reads 621.14 as the driver that fixed those black screens.

How do I know whether my machine is affected?+

Open Resource Monitor, go to the Disk tab and sort by Write Rate, then look for NVDisplay.Container.exe. If the machine is idling and that process writes several MB per second to C:, the number is suspicious. As a second reading, note the SMART counter for data written twice, thirty minutes apart, and extrapolate the difference to one hour.

Can I just delete or lock the log file?+

The threads discuss deleting nvlog.nvlgstg and blocking it with file permissions for the system account. We have not tested that and we would not do it on a production machine while it is unclear which driver functions depend on the file. Forum reports say the file comes back after deletion anyway. Rolling back to an earlier driver is the documented route.