Reference
GPUs
The four GPU types, their memory, and how auto-selection and multi-GPU work.
GPU types
| Value | GPU | Memory per GPU | Typically used for |
|---|---|---|---|
| l40s | NVIDIA L40S | 48 GB | Distilled and smaller models such as Z-Image Turbo and Wan 2.2 5B. |
| h100 | NVIDIA H100 | 80 GB | Full-size models such as Qwen-Image, Qwen-Image-Edit and Wan 2.2 14B. |
| h200 | NVIDIA H200 | 141 GB | Large video models, long clips, high resolution. |
| b200 | NVIDIA B200 | 192 GB | The largest models and the highest throughput. |
Automatic selection
gpu: auto picks the smallest GPU type that fits the model and the requested output size, which is usually the cheapest per output. clockingcloud plan shows the choice before you commit. If your resolution or clip length needs more memory than the default, name a larger GPU or raise gpus.
Multiple GPUs
Set gpus to 2, 4 or 8 to place one replica on several GPUs of the same type. Video models use sequence parallelism to split a long clip across them, which reduces latency. Billing is the per-second rate multiplied by the number of GPUs.
Availability
Capacity for a given GPU type can be temporarily constrained. If a replica cannot be allocated within the --timeout, the instance moves to failed, the GPU is released, and you are not billed for time before attachment. Retry, or choose auto to allow a different compatible GPU.
Something unclear or missing? Email hello@clockingcloud.com.