Stable Diffusion Inference: Memory Requirements, Speed and GPU Selection (opens in new tab)

When teams plan infrastructure for Stable Diffusion, the conversation usually starts with GPU speed. Should we use an L40S? Would an H100 generate images faster? How many images can it produce per minute? Those are useful questions. But they often skip the constraint that decides whether the workload can run properly in the first place: GPU memory. A model may generate one image quickly during testing and still struggle in production. Why? Because production adds larger resolutions, concurren...