¿Qué GPU debo alquilar?

Lista de configuración

• Use a serving engine with continuous batching (vLLM, SGLang or TGI) — it's what makes concurrent requests cheap.
• Download weights in the quantized format you sized for (AWQ/GPTQ for 4-bit, FP8 on Hopper/Ada) rather than quantizing on the box.
• Put model files on a persistent volume so a restart doesn't re-download tens of GB.
• Load-test with your real prompt lengths before sending traffic; long contexts shrink how many requests fit in memory.

Cómo se calcula

The wizard starts from memory, because a GPU that can't hold the job is useless at any price. For running a model we size the weights at the precision your priority implies — 4-bit for cheapest, 8-bit for balanced, 16-bit for fastest — plus KV cache for an 8K context. Fine-tuning uses QLoRA (or LoRA when you pick fastest), training uses full mixed-precision state, and image models use typical VRAM needs for each model family.

Among the GPUs that fit and have a live price, cheapest ranks by hourly rate, fastest by memory bandwidth (which sets generation speed), and balanced by price per unit of bandwidth — the most speed per dollar. If no single card fits, it shows the smallest multi-GPU shapes instead. Every result links to the GPU's page with all providers, because the cheapest provider isn't always the most reliable one.