It’s likely mostly a security patch release. Those drivers tend to have short or vague release notes.
List of issues that still affect the 620.73.04 driver:
- Random black screens when playing intensive games, sometimes resulting in a partial GPU driver crash. It is sometimes possible to switch to a TTY, but attempting to switch back to a graphical session will crash the driver for good.
- Desktop memory usage is excessively high(over 2GB sometimes).
- A large amount of GPU Memory is unaccounted for in
nvidia-smi/ NVML. - Applications crash when VRAM is full.
- Shared RAM under Linux seems broken. It’s almost like there is an incredibly small amount(256MB) and once that’s gone things start crashing.
- DLSS3 / DLSS4 / Smooth Motion is broken. It causes massive input lag and micro stuttering and frame pacing / slowness. This is better in 615.78.08 but it’s still unusable with 3x or 4x FG.
- Gnome-shell / GDM can freeze on login from a fresh boot. Switching to a TTY and relogging in works around this.
- Ray Tracing is broken on a lot of games, resulting in an XID 109 or some other error.
- Some games show DLSS upscaling options but no frame generation(Portal 2: RTX). Maybe a Proton issue?
- DX12 games have a massive performance penalty.
- VRR causes backlight strobing/flickering.
- Shader cache doesn’t save properly, forcing it to be computing again.
- Video decoding from a GTX 960 using Sunshine/Moonlight causes blocky artifacts.
- HDR is broken.
- Using a secondary dedicated GPU often fails with
[drm:__nv_drm_gem_nvkms_prime_dup [nvidia_drm]] *ERROR* [nvidia-drm] [GPU ID 0x00000c00] Failed to import NVKMS memory to GEM object. This error can happen regardless of whether you usedxvk.deviceFilteror in-game display adapter settings. Other times games start but run at 5 FPS. - Display sometimes fails to wakeup from sleep unless switched to TTY.
- Applications crash due to failing to allocate VA mapping.
- Using cgroups can cause a full system crash.
NVML bugs, just because:
- NVML is advertised as thread safe, when it is not.
- Performance limiters are broken on Blackwell.
- GPM shows hardware functionality as supported when it is not.
- GPM functions return a success return value on GPUs it should not.
nvmlDeviceGetPcieSpeedappears to return info reported bynvmlDeviceGetGpuMaxPcieLinkGeneration, not the PCIe speed.- GPU performance state min/max thresholds via
nvmlDeviceGetDynamicPstatesInfowill always return 0 on newer GPU archs. - The memory clocks and graphics clocks returned via
nvmlDeviceGetSupportedMemoryClocksandnvmlDeviceGetSupportedGraphicsClocksdo not take into account overclocking on newer GPU archs. - PowerMizer mode resets on application exit.
nvmlDeviceGetMaxPcieLinkWidthalways returns 16x even on GPUs that only support 8x width.
Regression: Hard system lockup when entering the open 3D world (RTX 4080 / NVIDIA 620.70.06 / Wayland)
System Information
- Game: Forza Horizon 6 (Steam AppID:
2483190) - OS: Fedora Linux 45 (64-bit)
- Desktop Environment: KDE Plasma 6.8.3 (Wayland)
- CPU: AMD Ryzen 7 7800X3D
- GPU: NVIDIA GeForce RTX 4080
- NVIDIA Driver:
620.70.06 - Proton Version:
Proton Experimental(1794497090 experimental-11.0-20261112-x86_64)
Regression Timeline
The game ran completely fine on this exact hardware and OS setup last month (October 2026) with DLSS and Raytracing enabled, without any freezes or stability issues. Since recent updates (NVIDIA 620 driver series and November Proton Experimental builds), the system locks up hard every single time when transitioning into the open world.
Symptom Description
- The game launches normally and reaches the main menu without issues.
- As soon as I click to enter/load the open 3D game world, the entire system suffers a complete hard freeze.
- No recovery is possible: display output freezes, audio stops, and even switching to a TTY via
Ctrl + Alt + F3is completely unresponsive, requiring a hard power reset. journalctl -b -1cuts off abruptly at the moment of the freeze without flushing anyNVRM: Xidor kernel panic message to disk.
Already Tested (No Effect)
- Cleared local Steam/VKD3D shader cache (
shadercache/2483190) and verified game file integrity via Steam. - Disabled Steam Overlay (
gameoverlayui). - Disabled Raytracing, DLSS Upscaling, DLSS Frame Generation, and NVIDIA Reflex in the main menu (switched to TAA) before entering the world.
- Tested launch options
VKD3D_CONFIG=no_upload_hvvas well asPROTON_VKD3D_HEAP=1 VKD3D_CONFIG=no_upload_hvv,descriptor_heap,enable_experimental_features.
I fixed this issue. Not rebased against newest driver tho. If interest, let me know.
Any idea whether the 100% GPU utilization fix, which is in 620.70.06 is also in 620.73.04? I get bit by these symptoms pretty often… I’m referring to this one:
* GPU utilization reported 100% following SIGINT-terminated CUDA workloads. We now perform a context reset at the end of the save sequence in FECS ucode.
Actually, there’s a hint that there’s a security related update. But it’s for confidential computing, not relevant for gamers. There are various other changes: Color pipeline fixes, HDMI/HDCP fixes, locking fixes, multi GPU fixes, … A lot of minor changes, as the changelog said. And memory hardening… ;-) Some more details would be appreciated anyways.
The binary GSP firmware surely got some improvements and changes, too. Probably also the closed-source user-space driver. But we cannot see what changed there.
Interested… I could add it to my collection of patches if it is not already there.
OTOH, so far there has been no corresponding DC driver release, which suggests this release contains mostly desktop-related fixes.
Intuitively guessing, the real fix may be in the closed-source core or firmware, so the open kernel module needs an adjustment and is not the real fix itself. As far as I understand, NVIDIA uses the same driver core and firmware across different platforms - Windows, Linux, DC driver, gaming driver.
Encountering an issue with 620 series drivers when running Space Engineers, which is not reproduceable in 615 or 610 series drivers. Game core-dumps on world load:
Nov 19 14:58:17 Pegasus kernel: NVRM: Xid (PCI:0000:01:00): 109, pid=21055, name=Interim.bin, channel 0x00000012, errorString CTX SWITCH TIMEOUT, Info 0x1a801a
Nov 19 14:58:37 Pegasus kernel: NVRM: Xid (PCI:0000:01:00): 31, pid=21055, name=Interim.bin, channel 0x00000013, intr 00000000. MMU Fault: ENGINE GRAPHICS GPC2 GPCCLIENT_GCC faulted @ 0x0_00122000. Fault is of type FAULT_PDE ACCESS_TYPE_VIRT_READ
Nov 19 15:17:21 Pegasus kernel: NVRM: Xid (PCI:0000:01:00): 109, pid=26023, name=SpaceEngineers., channel 0x00000010, errorString CTX SWITCH TIMEOUT, Info 0x168016
Nov 19 15:17:34 Pegasus kernel: NVRM: Xid (PCI:0000:01:00): 31, pid=26023, name=SpaceEngineers., channel 0x00000011, intr 00000000. MMU Fault: ENGINE GRAPHICS GPC2 GPCCLIENT_GCC faulted @ 0x0_0011b000. Fault is of type FAULT_PDE ACCESS_TYPE_VIRT_READ
Nov 19 16:12:23 Pegasus kernel: NVRM: Xid (PCI:0000:01:00): 109, pid=30329, name=SpaceEngineers., channel 0x00000010, errorString CTX SWITCH TIMEOUT, Info 0x168016
Nov 19 16:12:42 Pegasus kernel: NVRM: Xid (PCI:0000:01:00): 31, pid=30329, name=SpaceEngineers., channel 0x00000011, intr 00000000. MMU Fault: ENGINE GRAPHICS GPC2 GPCCLIENT_GCC faulted @ 0x0_00169000. Fault is of type FAULT_PDE ACCESS_TYPE_VIRT_READ
Failure is totally independent of Proton Version, and fails when being run through:
- Proton GE 11-12
- Proton Hotfix
- Proton Stable
- Community Maintained Linux Native Client (Pulsar).
Experiment concludes that the nvidia-620 series driver is responsible.
Environment:
GPU: RTX 5070 mobile
Driver Version: 620.70.06
OS: Fedora 45
Kernel Version: Linux 7.3.4-300.fc45.x86_64
I upstreamed my fix here, with help of this nvidia-driver maintainer. Take are look, if you wanna understand how it work.
I will rebase against “clean” driver later.
Thanks. Yes, this patch is already in my patch collection, too. During review and testing, I found a few more fixes which I pushed to my branch.
Additional evidence for the NVIDIA DRM event-handling regression in 620.73.04
I have been investigating intermittent NVIDIA DRM failures on an RTX 4090 running Arch Linux, kernel 7.3.5-arch1-1, NVIDIA Open 620.73.04, and KDE Plasma 6.8.3 on Wayland.
I noticed the recently published patch addressing the NVIDIA DRM event-handling deadlock:
My logs show a potentially relevant timing pattern during SDDM Wayland greeter initialization:
[14.789234] Flip event timeout on head 0
[17.797261] Flip event timeout on head 1
[17.797439] nv_flip == NULL
[17.800393] nv_flip == NULL
The two timeouts occur approximately three seconds apart. The same warning signature was observed in the NVIDIA DRM event-handling path on an earlier boot.
Across six consecutive boots, the problem occurred intermittently with both default and custom GPU power/clock settings. On affected boots, SDDM initialization was delayed, while on unaffected boots both displays initialized normally.
This appears consistent with a timing-sensitive synchronization issue, although I have not tested the proposed patch.
I have posted the detailed six-boot comparison and system configuration in the related KWin/Wayland regression thread:
Can NVIDIA confirm whether the proposed event-handling change is being evaluated for inclusion in an official driver release? I would also be interested to know whether it is expected to address the SDDM startup timeouts.
nvidia-test-enabled-boot2.txt (1.1 KB)
Thank you.
Thanks everyone for the reports on “Flip event timeout” / nv_flip == NULL / page-flip timeout issue on driver 620.70.06 with multiple displays
We’ve filed internal bug 7126445 , and we’re working on reproducing it internally.
As a diagnostic experiment, could someone try setting the nvidia_modeset.opportunistic_flip_sync=0 kernel module parameter and report whether it changes the behavior?