ComfyUI Multi GPU: How to Use Multiple GPUs Without Breaking Workflows

Learn what ComfyUI can and cannot do with multiple GPUs, how to choose a CUDA device, and how to run separate ComfyUI instances on separate ports.

ComfyUI can run on a specific GPU, and you can run multiple ComfyUI instances on different GPUs.

But a normal ComfyUI workflow usually does not become twice as fast just because your machine has two GPUs. Most users get the best result by assigning one ComfyUI instance to one GPU and another instance to another GPU.

Quick Answer

Goal	Best Setup	Example
Use GPU 1 instead of GPU 0	Start ComfyUI with `--cuda-device 1`	`python main.py --cuda-device 1`
Run two queues at the same time	Start two ComfyUI instances on different ports	`--cuda-device 0 --port 8188`, `--cuda-device 1 --port 8189`
Keep all GPUs visible but prefer one	Use `--default-device` only if you know why	`python main.py --default-device 1`
Make one KSampler split across GPUs	Usually not supported by default	Use a model or node that explicitly supports it
Fix VRAM pressure	Use VRAM flags or smaller workflows	`--lowvram`, smaller resolution, quantized models

What Multi GPU Means in ComfyUI

There are three different ideas people mix together:

Choosing one GPU: ComfyUI runs on a selected device.
Running multiple instances: two browser sessions, two ports, two queues, two GPUs.
Splitting one workflow across GPUs: a single model or workflow uses more than one GPU at the same time.

The first two are practical. The third is not something most standard ComfyUI workflows do automatically.

Step 1: Check Your GPUs

On Windows or Linux with NVIDIA GPUs, run:

nvidia-smi

Look for the GPU index in the left column. The first card is usually 0, the second is usually 1.

Then confirm PyTorch can see CUDA:

python -c "import torch; print(torch.cuda.device_count()); print(torch.cuda.get_device_name(0))"

If torch.cuda.device_count() returns 0, you do not have a multi-GPU problem yet. You have a PyTorch/CUDA install problem. Start with GPU Compatibility.

Step 2: Choose One GPU With `--cuda-device`

For a manual install:

python main.py --cuda-device 1

For the Windows portable package, edit or copy your launch .bat file and add the flag after main.py:

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --cuda-device 1

The official ComfyUI argument description says --cuda-device sets the CUDA device this instance will use and hides the others from that process.

That means ComfyUI may still print cuda:0 inside the selected process. This can be normal because the chosen GPU becomes the only visible GPU to that process.

Step 3: Run Two Instances on Two GPUs

Use different ports so the two servers do not collide.

Instance A:

python main.py --cuda-device 0 --port 8188

Instance B:

python main.py --cuda-device 1 --port 8189

Then open:

http://127.0.0.1:8188
http://127.0.0.1:8189

This is the most reliable multi-GPU pattern for production work: one queue per GPU.

Use separate user directories for serious parallel work

If both instances are active every day, consider separate --user-directory paths so browser state, workflow tabs, and user settings do not fight each other.

Example:

python main.py --cuda-device 0 --port 8188 --user-directory user-gpu0
python main.py --cuda-device 1 --port 8189 --user-directory user-gpu1

Step 4: Understand `--default-device`

ComfyUI also has a --default-device option. It sets the default device while keeping other devices visible.

That sounds attractive, but it is not the same as automatic multi-GPU execution. Use it only when a workflow, custom node, or advanced setup needs other devices to remain visible.

For most users, --cuda-device is easier to reason about.

Step 5: Do Not Use Multi GPU to Hide a VRAM Problem

If a workflow fails because one GPU runs out of VRAM, adding a second GPU usually will not make the same model magically fit. Try these first:

reduce image resolution
reduce batch size
close other GPU-heavy apps
use a quantized model when appropriate
start ComfyUI with --lowvram
use a smaller model or workflow

See Python Out of Memory in ComfyUI for a deeper diagnosis path.

Common Mistakes

Mistake	Why It Fails	Better Approach
Opening two browser tabs on the same server	Both tabs still use the same ComfyUI process	Start a second process on another port
Adding `--cuda-device 1` in the wrong place	The flag is not passed to `main.py`	Put it after `main.py` in the launch command
Expecting one workflow to split across two GPUs	Standard nodes usually run on one selected device	Use separate queues or specialized nodes
Running two instances on port `8188`	Only one process can own the port	Use `8188` and `8189`
Installing CPU-only PyTorch	ComfyUI cannot use any NVIDIA GPU	Reinstall PyTorch with CUDA

How to Verify the Setup

Start ComfyUI and read the startup log. You want to see your GPU name near the device line:

Device: cuda:0 NVIDIA GeForce ...

Then watch nvidia-smi while generating. If you run two instances, queue a small workflow in each one and confirm each GPU shows activity.

If only one GPU is busy, check:

which port your browser is using
whether both processes are still running
whether each command has a different --cuda-device
whether the second process failed to start because the port was already occupied

How Wonderful Launcher Helps

Wonderful Launcher is a good fit when you want multiple stable ComfyUI environments instead of one fragile folder:

keep a clean GPU 0 environment for normal images
keep a separate GPU 1 environment for experiments or video nodes
avoid custom node dependency conflicts between queues
preserve working workflows before changing launch flags

If you are still choosing hardware, start with GPU Compatibility. If ComfyUI launches but the browser keeps disconnecting, see ComfyUI Reconnecting Error.