Fine-tuning cost estimator in taka
Enter the model, your data and the number of epochs: get the time and cost on each AI Computer — and print a one-page estimate for your supervisor, lab or client.
Fine-tuning Qwen2.5 7B with LoRA on 2,000,000 tokens × 3 epoch(s): about 38 minutes on NVIDIA L40S, costing about ৳217.
Fastest: NVIDIA B200 — 14 minutes, ৳376.
| AI Computer | Memory | Time | Estimated cost |
|---|---|---|---|
| NVIDIA L4৳147/hr | 18 GB / 24 GB | 1.6 hours | ৳234 |
| NVIDIA A10৳196/hr | 18 GB / 24 GB | 1.6 hours | ৳303 |
| NVIDIA L40S৳343/hr | 18 GB / 48 GB | 38 minutes | ৳217 |
| NVIDIA A100৳638/hr | 18 GB / 80 GB | 43 minutes | ৳453 |
| NVIDIA H100৳1,047/hr | 18 GB / 80 GB | 20 minutes | ৳342 |
| NVIDIA H200৳1,390/hr | 18 GB / 141 GB | 20 minutes | ৳454 |
| NVIDIA B200৳1,652/hr | 18 GB / 180 GB | 14 minutes | ৳376 |
| NVIDIA B300৳1,880/hr | 18 GB / 288 GB | 14 minutes | ৳428 |
Ways to spend less
- Start with QLoRA or LoRA on a 7–8B model — it is often good enough, and 10× cheaper than a bigger model.
- Try 1 epoch on 10 % of the data first, check the results, then run the full job.
- Remove duplicate and very long examples; shorter sequences train much faster.
- Set an auto-stop budget when you launch — the session stops itself at that amount.
Frequently asked questions
How accurate is this estimate?
It is a planning estimate based on the standard rule that training takes about 6 × parameters × tokens calculations (4× for LoRA/QLoRA), at 30–35 % of the GPU’s peak speed. Real runs are often within ±50 % — add a margin to your budget.
How many epochs should I use?
For LoRA fine-tuning, 1–3 epochs is usual. Watch the validation loss and stop when it starts rising.
Am I billed while the model downloads?
You are billed by the minute while your AI Computer runs, including downloading the model. The estimate adds about 10 minutes for that.
Try it on a real AI Computer
JupyterLab opens in about a minute. Pay in taka with bKash, billed by the minute — stop whenever you like.