Which GPU for which workload
For each workload, the memory it needs - worked out on its page from parameters and bytes per parameter - and the cheapest card that has that much, at its cheapest live on-demand rate as of .
| Workload | Needs | Cheapest card that fits | From, per GPU-hour |
|---|---|---|---|
| GPUs for LLM inference An 8B model at 16-bit | 24 GB or more | NVIDIA V100 32GB  Geodd | $0.16 captured |
| A 70B model at 8-bit | 80 GB or more | NVIDIA RTX PRO 6000 Blackwell  Hostinger | $0.65 captured |
| A 70B model at 16-bit on one GPU | 141 GB or more | AMD Instinct MI300X  TensorWave | $1.71 captured |
| GPUs for fine-tuning QLoRA on an 8B model | 16 GB or more | NVIDIA RTX A4000  GPU.ai | $0.12 captured |
| LoRA on an 8B model, 16-bit base | 24 GB or more | NVIDIA V100 32GB Geodd | $0.16 captured |
| Full fine-tuning of an 8B model on one GPU | 141 GB or more | AMD Instinct MI300X TensorWave | $1.71 captured |
| GPUs for training Cards with 80 GB or more | 80 GB or more | NVIDIA RTX PRO 6000 Blackwell Hostinger | $0.65 captured |
| GPUs for image generation SDXL at 16-bit | 12 GB or more | NVIDIA GeForce RTX 3060  Edgevana | $0.07 captured |
| FLUX.1 [dev] at 16-bit | 32 GB or more | NVIDIA V100 32GB Geodd | $0.16 captured |
| Cheapest GPU memory per hour Every card, cheapest GB first | any memory | NVIDIA V100 32GB Geodd | $0.16 captured |
Data updated: