Skip to main content
Image-generation workloads usually depend more on VRAM than raw GPU count. Start with enough GPU memory for the model and workflow, then add GPUs only when your software can use them or you want to run jobs in parallel. Compute currently uses NVIDIA RTX 5090 GPUs with 32 GB of VRAM per GPU.

​
A practical starting point

Start with one GPU when the workload fits. A second GPU does not automatically make a single Stable Diffusion or ComfyUI job twice as fast.

​
Check VRAM before adding GPUs

After starting the workload, monitor GPU memory with:
If the job approaches the 32 GB VRAM limit, first check whether you can reduce memory use through the application itself. Depending on the tool, useful options can include a smaller model, lower resolution, smaller batch size, tiled processing, or memory-efficient attention. If the workload genuinely needs more than one GPU, confirm that the application or framework supports distributing work across several cards.
Multiple GPUs have separate VRAM. Their memory is not automatically combined into one larger pool.

​
When multiple GPUs help

Consider a multi-GPU instance when you:
  • Run several image-generation jobs at the same time and can assign them to different GPUs.
  • Use a training or fine-tuning framework with explicit distributed or multi-GPU support.
  • Use rendering or compute software designed to scale across several GPUs.
  • Have tested one GPU and know the workload can benefit from additional cards.
For ordinary interactive image generation, one RTX 5090 is usually the simpler place to start.

​
Change configuration later

Compute does not resize an existing instance in place. To move to another GPU configuration:
  1. Save any models, outputs, configuration files, or other data you need.
  2. Stop or terminate the current instance as appropriate.
  3. Create a new instance with the GPU preset you need.
  4. Restore your files or reuse a custom template to rebuild the environment.
Terminating an instance permanently deletes its local data. Copy anything important somewhere safe before you terminate it.

​
Check current availability

GPU presets depend on location and current capacity. The Compute creation screen shows the GPU count, vCPU, RAM, disk, bandwidth, availability, and active price for the presets you can launch now. See GPU types and sizes for current fleet guidance.