Free guide
GPU Sizing Guide
Which NVIDIA card for which model. A method to work out GPU memory and card count from your workload, for inference, fine-tuning, training, image and video generation or rendering, then prove the choice on a trial node before you commit.
PDF · 16 pages · Updated October 2026
For: CTOs, ML engineers and founders choosing GPU capacity for inference, fine-tuning, training or rendering.
Get the guide
What’s inside.
- Start from the workload, not the card
- Memory for inference
- Memory for fine-tuning and training
- More than one card
- Concurrency and context drive cost
- Images, video and rendering
- Matching the workload to our cards
- The host matters too
- Prove it on a trial node
- One-page sizing worksheet
What you’ll be able to do.
- Estimate GPU memory from the model: weights, KV cache and headroom.
- Size fine-tuning and training, including LoRA and QLoRA.
- Know when you need more than one card, and when NVLink matters.
- Match the workload to L40S, RTX PRO 6000, H200 and the rest of the NVIDIA range.
- Prove the estimate on a trial node with a written test plan.
Rather have us do it?
GPU cloud, run by our 24/7 team. We reply within one business day.