You are not logged in.
Pages: 1
Hello everyone. I use a laptop to work with ComfyUI, an RTX 5070 card with 8GB of VRAM. No other processes use the GPU, the card operates in hybrid mode. nvidia-smi shows 445 MB Reserved??!! Total - 8151, Free - 7569. Comfy shows "Total VRAM 7706 MB" when loading and uses only that.. But in Windows, Comfy can use all 8Gb. Can you tell me please how I can free up reserved VRAM? Or at least reduce it? Driver Version: 580.105.08 CUDA Version: 13.0
thnx.
Offline
No other processes use the GPU, the card operates in hybrid mode.
Only one of those claims can be true.
What's the actual and complete output of nvidia-smi?
Offline
Only one of those claims can be true.
Why so? If all other processes use only the integrated card, and the discrete one is used only for GPU Comfy
What's the actual and complete output of nvidia-smi?
nvidia-smi
Sun Nov 30 18:14:48 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.105.08 Driver Version: 580.105.08 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 5070 ... Off | 00000000:01:00.0 Off | N/A |
| N/A 32C P4 9W / 47W | 2MiB / 8151MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| No running processes found |
+-----------------------------------------------------------------------------------------+nvidia-smi -q -d MEMORY
==============NVSMI LOG==============
Timestamp : Sun Nov 30 18:25:55 2025
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
FB Memory Usage
Total : 8151 MiB
Reserved : 445 MiB
Used : 2 MiB
Free : 7705 MiB
BAR1 Memory Usage
Total : 8192 MiB
Used : 2 MiB
Free : 8190 MiB
Conf Compute Protected Memory Usage
Total : 0 MiB
Used : 0 MiB
Free : 0 MiBCan the integrated card reserve this memory for itself? ))
CPU Intel(R) Core(TM) Ultra 7 255H (16) @ 5.10 GHz
GPU NVIDIA GeForce RTX 5070 Max-Q / Mobile [Discrete]
GPU Intel Arc Pro 140T @ 2.25 GHz [Integrated]
Vulkan 1.4.318 - Intel open-source Mesa driver [Mesa 25.2.7-ar]
Display(s) 2560x1600 @ 1.25x in 16", 240 Hz [Built-in]Last edited by pernostro (2025-11-30 15:33:02)
Offline
Why so? If all other processes use only the integrated card, and the discrete one is used only for GPU Comfy
Because of the very definition of a "hybrid mode" - you're however not doing that according the output of nvidia-smi.
Seems reserved for the framebuffer
Blacklist nvidia-drm, resp. first try to just add "nvidia_drm.modeset=0 nvidia_drm.fbdev=0" to the https://wiki.archlinux.org/title/Kernel_parameters
Offline
Thanks for your participation, man.
sudo cat /sys/module/nvidia_drm/parameters/modeset
N
sudo cat /sys/module/nvidia_drm/parameters/fbdev
N
nvidia-smi -q -d MEMORY
==============NVSMI LOG==============
Timestamp : Sun Nov 30 20:15:44 2025
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
FB Memory Usage
Total : 8151 MiB
Reserved : 445 MiB
Used : 0 MiB
Free : 7706 MiB
BAR1 Memory Usage
Total : 8192 MiB
Used : 1 MiB
Free : 8191 MiBI even tried some livesessions with Cachy and Endeavour:
1 [liveuser@CachyOS ~]$ nvidia-smi
2 Sun Nov 30 18:55:56 2025
3 +-----------------------------------------------------------------------------------------+
4 | NVIDIA-SMI 580.105.08 Driver Version: 580.105.08 CUDA Version: 13.0 |
5 +-----------------------------------------+------------------------+----------------------+
6 | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
7 | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
8 | | | MIG M. |
9 |=========================================+========================+======================|
10 | 0 NVIDIA GeForce RTX 5070 ... Off | 00000000:01:00.0 Off | N/A |
11 | N/A 39C P4 8W / 42W | 15MiB / 8151MiB | 0% Default |
12 | | | N/A |
13 +-----------------------------------------+------------------------+----------------------+
14
15 +-----------------------------------------------------------------------------------------+
16 | Processes: |
17 | GPU GI CI PID Type Process name GPU Memory |
18 | ID ID Usage |
19 |=========================================================================================|
20 | 0 N/A N/A 1416 G /usr/lib/Xorg 4MiB |
21 +-----------------------------------------------------------------------------------------+
22 [liveuser@CachyOS ~]$ nvidia-smi -q -d memory
23
24 ==============NVSMI LOG==============
25
26 Timestamp : Sun Nov 30 18:56:13 2025
27 Driver Version : 580.105.08
28 CUDA Version : 13.0
29
30 Attached GPUs : 1
31 GPU 00000000:01:00.0
32 FB Memory Usage
33 Total : 8151 MiB
34 Reserved : 445 MiB
35 Used : 15 MiB
36 Free : 7692 MiB [liveuser@eos-2025.11.24 ~]$ nvidia-smi
Sun Nov 30 19:08:51 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.105.08 Driver Version: 580.105.08 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 5070 ... Off | 00000000:01:00.0 Off | N/A |
| N/A 40C P4 9W / 42W | 15MiB / 8151MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 1433 G /usr/lib/Xorg 4MiB |
+-----------------------------------------------------------------------------------------+
[liveuser@eos-2025.11.24 ~]$ nvidia-smi -q -d memory
==============NVSMI LOG==============
Timestamp : Sun Nov 30 19:09:10 2025
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
FB Memory Usage
Total : 8151 MiB
Reserved : 445 MiB
Used : 15 MiB
Free : 7692 MiB
BAR1 Memory Usage
Total : 8192 MiB
Used : 3 MiB
Free : 8189 MiB445 MB reserved everywhere )
Offline
Seems unconditional.
Does it not surrender it when requesting the memory through CUDA?
Does it still reserve it when blacklisting nvidia_drm ?
Offline
Yes, it still reserve:
lsmod | grep nvidia
nvidia_wmi_ec_backlight 12288 0
nvidia_uvm 4096000 0
nvidia_modeset 2277376 0
nvidia 15912960 13 nvidia_uvm,nvidia_modeset
video 81920 6 nvidia_wmi_ec_backlight,asus_wmi,asus_nb_wmi,xe,i915,nvidia_modeset
wmi 36864 4 video,nvidia_wmi_ec_backlight,asus_wmi,wmi_bmof
nvidia-smi -q -d memory
==============NVSMI LOG==============
Timestamp : Mon Dec 1 01:33:01 2025
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
FB Memory Usage
Total : 8151 MiB
Reserved : 445 MiB
Used : 0 MiB
Free : 7706 MiB
BAR1 Memory Usage
Total : 8192 MiB
Used : 1 MiB
Free : 8191 MiBNow used in FB - zero
)
And then I returned nvidia_drm back
lsmod | grep nvidia
nvidia_wmi_ec_backlight 12288 0
nvidia_uvm 4096000 0
nvidia_drm 147456 0
nvidia_modeset 2277376 1 nvidia_drm
nvidia 15912960 13 nvidia_uvm,nvidia_modeset
drm_ttm_helper 16384 2 nvidia_drm,xe
video 81920 6 nvidia_wmi_ec_backlight,asus_wmi,asus_nb_wmi,xe,i915,nvidia_modeset
wmi 36864 4 video,nvidia_wmi_ec_backlight,asus_wmi,wmi_bmof
nvidia-smi -q -d memory
==============NVSMI LOG==============
Timestamp : Mon Dec 1 01:35:28 2025
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
FB Memory Usage
Total : 8151 MiB
Reserved : 445 MiB
Used : 2 MiB
Free : 7705 MiB
BAR1 Memory Usage
Total : 8192 MiB
Used : 2 MiB
Free : 8190 MiBComfy doesn't see these 445 MB and immediately announces upon loading that it will work with what's left -)
Who pocketed them remains a secret -)
Offline
You should probably also unload nvidia_modeset (which is probably only loaded by nvidia_drm)
Who pocketed them remains a secret -)
https://docs.nvidia.com/deploy/nvidia-smi/index.html - nvidia admits to it being their driver ![]()
Offline
lsmod | grep nvidia
nvidia_wmi_ec_backlight 12288 0
nvidia_uvm 4096000 0
nvidia 15912960 12 nvidia_uvm
video 81920 5 nvidia_wmi_ec_backlight,asus_wmi,asus_nb_wmi,xe,i915
wmi 36864 4 video,nvidia_wmi_ec_backlight,asus_wmi,wmi_bmof
nvidia-smi -q -d memory
==============NVSMI LOG==============
Timestamp : Mon Dec 1 02:12:52 2025
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
FB Memory Usage
Total : 8151 MiB
Reserved : 445 MiB
Used : 0 MiB
Free : 7706 MiB
BAR1 Memory Usage
Total : 8192 MiB
Used : 1 MiB
Free : 8191 MiB![]()
They say "The driver may also reserve a small amount of memory for internal use"
smaaaal ![]()
Offline
You could try to disable the https://wiki.archlinux.org/title/NVIDIA … P_firmware but I'm not sure whether the GPU will even respond to nvidia-smi then and it will probably not do anything useful (so it's just to see whether that's what requires the memory, your GPU isn't supported by the proprietary driver)
Offline
I gave up
after looking at user bug reports on the Nvidia forum, which showed they also had quite a bit of reserved memory. Although their bug reports were not about that at all
But it's one thing if they have 32GB of VRAM and 500MB reserved, while mine is just a tiny bit. I had to squeeze every last bit of money to buy a laptop with 8GB VRAM. And if I had 32GB, I wouldn't worry about the reserved 500MB
So I guess it's just a driver issue and nothing can be done.
And thanks again for your participation, man. No one else responded to my problem either here or on the Nvidia forum.
So, "let it be", I said to myself ![]()
Look, from their bug reports, where they complain about some other error ![]()
Driver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:01:00.0
Product Name : NVIDIA GeForce RTX 5070 Ti
-----
FB Memory Usage
Total : 16303 MiB
Reserved : 466 MiB
Used : 15802 MiB
Free : 36 MiBDriver Version : 580.105.08
CUDA Version : 13.0
Attached GPUs : 1
GPU 00000000:2B:00.0
Product Name : NVIDIA GeForce RTX 4090
------
FB Memory Usage
Total : 24564 MiB
Reserved : 517 MiB
Used : 1847 MiB
Free : 22202 MiBSo I'm just a fool who thought his 500 MB was stolen ![]()
And no one on the Nvidia forum said, "Yeah, it's normal, don't worry about it, It's the same for everyone" ![]()
Because they apparently understand that this isn't normal ![]()
Last edited by pernostro (2025-12-01 20:58:04)
Offline
Check
modinfo nvidiafor whether there's any parameter that's suspicious to allowing you to control the amount of reserved FB VRAM
Then see https://github.com/NVIDIA/open-gpu-kern … les/issues
It seems inconceivable that this memory is pre-allocated if the device isn't even used for a framebuffer (please make extra sure to blacklist nvidia_drm and nvidia_modeset from the get-go resp. try to boot w/ "nvidia_drm.modeset=0 nvidia_drm.fbdev=0" to see whether it's only a problem to release the memory but it's not allocated if the framebuffer is never initialized) so you could raise an issue about that
(My gut feeling from the sheer size of allocated memory is however that it's indeed the GSP)
Offline
I checked and even tried to change some of them, but without success. Still reserved.
sort /proc/driver/nvidia/params
CoherentGPUMemoryMode: ""
CreateImexChannel0: 0
DeviceFileGID: 0
DeviceFileMode: 438
DeviceFileUID: 0
DmaRemapPeerMmio: 1
DynamicPowerManagement: 2
DynamicPowerManagementVideoMemoryThreshold: 200
EnableDbgBreakpoint: 0
EnableGpuFirmware: 18
EnableGpuFirmwareLogs: 2
EnableMSI: 1
EnablePCIeGen3: 0
EnablePCIERelaxedOrderingMode: 0
EnableResizableBar: 0
EnableS0ixPowerManagement: 0
EnableStreamMemOPs: 0
EnableUserNUMAManagement: 1
ExcludedGpus: ""
GpuBlacklist: ""
GrdmaPciTopoCheckOverride: 0
IgnoreMMIOCheck: 0
ImexChannelCount: 2048
InitializeSystemMemoryAllocations: 1
KMallocHeapMaxSize: 0
MemoryPoolSize: 0
ModifyDeviceFiles: 1
NvLinkDisable: 0
OpenRmEnableUnsupportedGpus: 1
PreserveVideoMemoryAllocations: 1
RegisterPCIDriver: 1
RegistryDwords: ""
RegistryDwordsPerDevice: ""
ResmanDebugLevel: 4294967295
RmLogonRC: 1
RmMsg: ""
RmNvlinkBandwidthLinkCount: 0
RmProfilingAdminOnly: 1
S0ixPowerManagementVideoMemoryThreshold: 256
TemporaryFilePath: "/var/tmp"
UsePageAttributeTable: 4294967295
VMallocHeapMaxSize: 0Of course, this is just a waste of resources
But after searching google down to the last page, I found out that it's normal practice for the driver to reserve up to 500MB of VRAM for an 8GB card.
There is no switch ![]()
Offline
Pages: 1