ComfyUI reservers an unreasonable amount of RAM sometimes freezing up my system

[panteraleo555-lab]

Hello,

I use the follow start up commands in terminal in Pop!os with native nvidia drivers.

conda activate comfyui_env
cd /home/user/ComfyUI
python main.py --normalvram --enable-dynamic-vram --enable-manager

my whole system slows down while using comfyui. It loads and holds 26.2 GB ram of the 31.3 available GB RAM.
Couple updates ago, it wasnt like that. It actually freezes up my computer.

the more generations I make, the more RAM it takes. Eventually freezing my computer.

[gershu-ar]

Hey, amigo. It's unclear by what you're sharing what is taking up all the RAM.
Check the "Processes" tab. ComfyUI runs on Python, look for "python.exe" and check RAM usage.

Also, with a 3090 I wouldn't even dare to use --normalvram: use --highvram or leave it empty, ComfyUI is good at managing VRAM.
There's no need to forcefully enable dynamic VRAM neither, it gets enabled by default on Nvidia GPUs. Avoid that switch.

I recommend adding --fast, to enable fp16 accumulation and use Sage Attention (--use-sage-attention). I too have a 3090 and I make the girl fly.

These are my switches:
--fast --use-sage-attention --highvram

Cheers.

[panteraleo555-lab]

Hello, thanks for your reply. I will try these commands. I am also not sure by what you mea by this "It's unclear by what you're sharing what is taking up all the RAM." My workflow with Qwen-image-2512 GGUF Q6 with 2 lora models and Qwen control net + img2img workflow takes about 22.6 of the 24 gb of the VRAM of the GPU. I now use CachyOS, which uses less than 1 GB VRAM. I might have been that Pop!OS was not as effective at using RAM than CachyOS. Since the problem seems to be alleviated.

thanks for your time.

[gershu-ar]

Hey. Screenshots of Task Manager don’t provide much actionable info — you need console logs or profiler data to see which processes are consuming resources, or at least a screenshot from the Processes tab.

Running Qwen with 2 LoRAs plus ControlNet is pushing your 3090 close to its limits. It’s not only about VRAM or RAM usage, but also raw compute throughput: Tensor Core scheduling, memory bandwidth, and kernel efficiency all come into play. Ampere is strong, but every workflow has its sweet spot — there are no universal settings.

Also keep in mind your system RAM: DDR4 at 3600 MT/s is decent, but slower compared to DDR5 system RAM setups — and nowhere near the bandwidth of the GDDR6X your 3090 uses. When workloads spill over into system memory, bandwidth and latency become another bottleneck. Avoid configs that spill into RAM as much as possible, keep memory usage tight and efficient. I have 64GB DDR5 6000 MT/s and even so I NEVER allow Comfy to spill models/LoRAs into RAM: keep your stuff locked into VRAM whenever possible.

Remember:
System's RAM DDR4/DDR5 -> Made for low latency and quick responsiveness.
GPU's RAM GDDR6X -> Made for massive bandwidth over immediate response speed.

Try --fast (fp16 accumulation), Sage Attention (about 15–20% it/s speed increase on a 3090 in my own tests), and experiment with fp8/int8 weights or encoders if your models tolerate quantization. fp8/int8 can make a huge difference, with minimal artifacts and massive VRAM savings — ideal for Qwen and WAN on big image renders and especially video (t2v, i2v, v2v).

Squeezing (but not choking) the most out of Ampere while keeping stability and resources in check → jackpot.