One work session
$2.72
1 GPU × 8h · compute only
Model adaptation
Rent GPUs for LLM fine-tuning with the method defined first.
Direct answer
LoRA, QLoRA and full fine-tuning have very different GPU memory and runtime profiles. A 24–48 GB RTX or L40S can handle many parameter-efficient jobs, while A100 or H100 fits larger models and full-state training. Define the method, sequence length and checkpoint plan before choosing a cloud GPU.
Reviewed by the GPURento infrastructure team · September 1, 2026
Transparent planning windows
Example rental cost on one RTX 4090.
These are compute rental windows, not predictions of job duration or performance. Replace the hours with a measured pilot and add storage in the calculator.
Forty-hour window
$13.60
1 GPU × 40h · compute only
One-hundred-twenty hours
$40.80
1 GPU × 120h · compute only
| Lower-memory path | LoRA / QLoRA | Adapter and quantization choices change quality and speed. |
|---|---|---|
| Higher-memory path | Full fine-tuning | Weights, gradients and optimizer states all matter. |
| Key variable | Sequence length | Activation memory can grow substantially. |
| Success metric | Cost to accepted checkpoint | Include evaluation and failed runs. |
Best for
- Domain adaptation with a fixed evaluation set
- LoRA and QLoRA experiments
- Full fine-tuning with a measured memory plan
- Teams saving reproducible checkpoints
Not the best fit when
- Training without a baseline or quality gate
- Uploading sensitive data without access controls
- Choosing hardware before defining the tuning method
Decision method
What to verify before you deploy capacity.
The content below separates published specifications, GPURento catalog references and decisions that still require a workload benchmark.
01
Separate parameter-efficient and full tuning
QLoRA can bring some jobs onto a 24 GB device, but the exact fit depends on base model, sequence length, batch, quantization and library. Full fine-tuning usually needs significantly more memory and may require multiple data-center GPUs.
02
Pilot with the same data and evaluation
Run enough steps to observe peak memory, tokens per second, checkpoint time and evaluation quality. Extrapolate only after the data pipeline and accumulation settings match the planned run.
- Pin the base model revision
- Record adapter rank and precision
- Store configuration beside each checkpoint
03
Do not pay to preserve an idle GPU
Separate reusable model and dataset storage from temporary compute. Stop compute after the checkpoint is safe, but include persistent storage and later reload time in the economic model.
Catalog shortlist
Relevant GPU options.
AdaAvailable
NVIDIA RTX 4090
24 GB GDDR6X · $0.34/hr
Image, video and mid-size inference
EU West · ParisEU Central · Frankfurt
AdaAvailable
NVIDIA L40S
48 GB GDDR6 ECC · $0.79/hr
Production inference, fine-tuning and video
EU West · ParisEU Central · Frankfurt
AmpereAvailable
NVIDIA A100 80GB
80 GB HBM2e · $1.19/hr
Training, fine-tuning and HPC
EU West · ParisEU Central · Frankfurt
HopperAvailable
NVIDIA H100 SXM
80 GB HBM3 · $2.69/hr
Intensive training and FP8 inference
EU West · ParisEU Central · Frankfurt
Questions
Clear answers, including the limits.
Can I fine-tune an LLM on an RTX 4090?+
Many LoRA or QLoRA jobs fit within 24 GB, but the result depends on the model, sequence length and settings. Full fine-tuning usually requires more memory.
A100 or H100 for fine-tuning?+
A100 has a lower GPURento reference rate; H100 may finish supported transformer workloads sooner. Measure cost to the same evaluation result.
What should persist after the GPU stops?+
Keep the accepted checkpoint, tokenizer, configuration, evaluation outputs and provenance. Temporary caches can be recreated if storage cost outweighs reload time.
Sources and scope
- NVIDIA GeForce RTX 4090 specificationsOfficial RTX 4090 architecture and memory specification.
- NVIDIA A100 product pageA100 architecture and 80 GB configuration from the manufacturer.
- NVIDIA H100 product pageArchitecture and memory specifications from the manufacturer.
Last reviewed September 1, 2026. GPURento rates are current catalog references; provisioning state remains visible in the workspace, and manufacturer specifications do not substitute for workload benchmarks.
Continue the research
Deployment plan
Define the tuning recipe before the GPU.
Save the model, method, sequence length, image and storage plan in a funded workspace request.