How to use Dedicated GPU Mode?
If your Run button shows an airplane icon, you can directly click the Run button and you'll see a popup window.
You'll need to select your desired GPU model, runtime duration, and any plugins you don't want to use. Some plugins take longer to start up, so you can choose to disable plugins that you won't be using.
After clicking start, the GPU will initialize. You'll see the bottom toolbar change to look like this. This means that the GPU has been initialized.
Clicking Run button will run the workflow.
An A10G GPU takes about 10 seconds to generate a 512x512 image using SD1.5 with 20 steps.
The overall experience is similar to using ComfyUI locally - after nodes are run, they won't be re-executed if the parameters haven't changed. When debugging workflows, this is faster compared to Serverless mode.
You can stop the GPU at any time by clicking the Stop button. The unused time's Credit will be returned to you.
If your Run button shows a car icon, that means you're currently in Serverless mode. You'll need to click the dropdown button next to run and select Dedicated GPU Mode.
What's the difference between Dedicated GPU mode and Serverless mode?
First, the cost difference is as follows:
In Serverless mode, you only pay for GPU runtime, not idle time. For example, you don't pay when adjusting your workflow. However, the GPU needs to restart ComfyUI each time it's used, and you pay for these startup costs:
In Dedicated GPU mode, you pay for both GPU runtime and idle time. For example, you pay even when adjusting your workflow. However, ComfyUI only needs to start once, which means the cost is lower. But since adjusting the workflow still consumes GPU runtime, the cost is higher. If you need to spend a long time aligning your workflow, the cost in Dedicated GPU mode may be higher.
Second, in terms of running Workflow mode, there are some subtle differences:
In Serverless mode, no matter how you adjust your workflow (e.g., only changing the parameters of 2 nodes), clicking Run will cause all nodes to be re-executed.
In Dedicated GPU mode, the opposite is true. If you only adjust the parameters of 2 nodes, only these 2 nodes will be re-executed. This means that the running speed is faster, so it's more suitable for debugging workflows.
Why or when to use Dedicated GPU Mode?
If you're debugging a workflow, Dedicated GPU Mode is a good choice since it won't re-run nodes that have already been executed, making debugging faster.
Additionally, if you want to generate multiple images consecutively, Dedicated GPU Mode is recommended. This is because in Serverless mode, the GPU needs to be restarted between image generation tasks, which makes it slower.