GPU planning guides

From format and budget to your GPU choice: compare memory, rental terms and results on your workload.

• How to choose a GPU for LLM inference

  Match a GPU to your model, context and concurrent requests. An 8B memory example, quantization limits and a practical latency test plan.

• LLM VRAM: weights, KV cache and context

  Why weight size is not the full VRAM requirement. A worked KV cache example for 8,192 tokens and four requests, with limits of the formula.

• LoRA and QLoRA: plan your GPU configuration

  What uses memory during LoRA and QLoRA: base weights, adapters, optimiser state and activations. Configure and test before fine-tuning.

• Cloud GPU, dedicated server or API: which should you choose?

  Compare cloud GPUs, dedicated servers and managed model APIs: responsibilities, serverless, total costs and questions to ask a provider.

• How to estimate GPU rental cost: compute, storage and idle time

  Build a GPU rental budget from billable time, storage, data transfer and additional costs, with a diagram and a calculator using your own inputs.

• RTX 4090 or L40S: which GPU should you rent for AI?

  Compare the RTX 4090 24 GB and L40S 48 GB for LLMs, image generation and fine-tuning: memory, limitations, performance testing and cost per result.

UZ Dynamics is in early access. These guides help describe a workload; they do not confirm GPU availability, a ready environment or execution speed. Availability, the exact configuration, prices and access terms need confirmation when discussing your request.